[go: up one dir, main page]

WO2025017101A1 - Method and system for prediction of epilepsy - Google Patents

Method and system for prediction of epilepsy Download PDF

Info

Publication number
WO2025017101A1
WO2025017101A1 PCT/EP2024/070335 EP2024070335W WO2025017101A1 WO 2025017101 A1 WO2025017101 A1 WO 2025017101A1 EP 2024070335 W EP2024070335 W EP 2024070335W WO 2025017101 A1 WO2025017101 A1 WO 2025017101A1
Authority
WO
WIPO (PCT)
Prior art keywords
parameter
training
subject
epilepsy
eeg
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/EP2024/070335
Other languages
French (fr)
Inventor
Jolan HEYSE
Pieter VAN MIERLO
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Universiteit Gent
Original Assignee
Universiteit Gent
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Universiteit Gent filed Critical Universiteit Gent
Publication of WO2025017101A1 publication Critical patent/WO2025017101A1/en
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/40—Detecting, measuring or recording for evaluating the nervous system
    • A61B5/4076—Diagnosing or monitoring particular conditions of the nervous system
    • A61B5/4094—Diagnosing or monitoring seizure diseases, e.g. epilepsy
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/369—Electroencephalography [EEG]
    • A61B5/372—Analysis of electroencephalograms
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems

Definitions

  • the present invention is in a field of diagnosing epilepsy in a subject after a suspected first seizure using an electroencephalogram (EEG).
  • EEG electroencephalogram
  • Epilepsy is a chronic disorder of the brain, characterized by the unprovoked recurrence of seizures. It affects approximately 1% of the global population, making it the most common neurological disorder after migraine. Diagnosis and management of patients who experienced a first seizure with transient neurological deficit or loss of consciousness remains a challenging task for the consulting neurologist, evidenced by the fact that not more than 50% of patients are diagnosed appropriately (Krumholz, Neurology, 2007).
  • Errors committed in epilepsy diagnosis and seizure classification are known to lead to inappropriate decisions on the use or choice of anti-epileptic drugs and to other serious patient management errors.
  • Differential diagnosis of epilepsy encompasses psychogenic events, transient ischemic attacks, migraine, cardiac events, etc.
  • Clinical history, electroencephalography (EEG), and neuroimaging are the fundamental modalities to achieve a correct diagnosis.
  • EEG electroencephalography
  • neuroimaging more specifically MRI
  • 50%-99% of the MRIs and up to 80% of scalp EEGs may be negative, despite an underlying epileptic disorder.
  • EEG plays a central role in the diagnosis and management of epilepsy and the presence of transient events such as interictal epileptiform discharges (IEDs) is one of the diagnostic hallmarks of epilepsy. Photic stimulation, hyperventilation and a period of attempted sleep are often included in 20-30min routine EEG recordings to increase the occurrence probability of these IEDs. Despite these extra tests, the sensitivity of recording an IED with routine EEG remains low (29-55%). Longer recordings either with prolonged recording times or multiple EEG recordings can increase the yield of IEDs but are impractical for clinical application (Burkholder, Neurology, 2016). This invention uses advanced EEG analysis techniques to extract parameters that objectively describe the characteristics of the EEG signals and combines them into an epilepsy risk index.
  • IEDs interictal epileptiform discharges
  • a method for determining a likelihood of a test subject having epilepsy after having suffered a possible epileptic seizure comprising: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state wherein the measurement parameter set comprises Parameter A of Table 1, - determining from the measurement data the likelihood of the test subject having epilepsy.
  • the method is a computer-implemented method.
  • Parameter A is relative band power, RBP, of a power spectral density, PSD, curve limited to a delta frequency band and limited to a front right region of the head of the subject.
  • an increase in Parameter A of the measurement parameter set compared with a healthy reference Parameter A is indicative of an increased likelihood of the test subject having epilepsy.
  • the measurement parameter set further comprises one or more of Parameter B, Parameter C, Parameter D, wherein each parameter is derived from the EEG recording of the test subject in the resting state.
  • Parameter B is a global efficiency functional connectivity parameter that is a statistical combination of (SPLs) -1 limited to a delta frequency band, in all regions of the head of the subject, wherein the SPL is a shortest path length between EEG electrode pairs in all regions of the head of the subject determined according to the path which yields a smallest summation of inverse connectivity strengths, CS-1, values along that path.
  • Parameter C is a connectivity strength functional connectivity parameter that is a statistical combination of connectivity strengths, CS, limited to a delta frequency band between EEG electrodes pairs limited to a back right, BR, and/or front left, FL, region of the head of the subject.
  • Parameter D is a 1/f exponent parameter that is a statistical combination of slope of a decay of a power spectral density, PSD, curve limited to a broadband frequency band for each EEG electrode limited to a front left (FL), region of the head of the subject.
  • an increase in Parameter B of the measurement parameter set compared with a healthy reference Parameter B is indicative of an increased likelihood of the test subject having epilepsy
  • an increase in Parameter C of the measurement parameter set compared with a healthy reference Parameter C is indicative of an increased likelihood of the test subject having epilepsy
  • an increase in Parameter D of the measurement parameter set compared with a healthy reference Parameter D is indicative of an increased likelihood of the test subject having epilepsy.
  • the measurement parameter set further comprises one or more of Parameter E, Parameter F, Parameter G, Parameter H, wherein each parameter is derived from the EEG recording of the test subject in the resting state.
  • Parameter E is a global efficiency functional connectivity parameter that is a statistical combination of (SPLs)-1 limited to a broadband frequency band, and limited to a front right (FR), of the head of the subject, wherein the SPL is a shortest path length between EEG electrode pairs in all regions of the head of the subject determined according to the path which yields a smallest summation of inverse connectivity strengths, CS -1 , values along that path.
  • Parameter F is a peak frequency parameter that is a statistical combination of frequency of a peak (maximal) signal of a power spectral density, PSD, curve limited to a broadband frequency band for each EEG electrode limited to a back left, BL, region of the head of the subject.
  • Parameter G is relative band power, RBP, of a power spectral density, PSD, curve limited to a gamma frequency band limited to a front right region of the head of the subject.
  • Parameter H is a peak frequency parameter that is a statistical combination of frequency of a peak (maximal) signal of a power spectral density, PSD, curve limited to a theta frequency band for each EEG electrode limited to a front right, FR, region of the head of the subject.
  • a decrease in Parameter E of the measurement parameter set compared with a healthy reference Parameter E is indicative of an increased likelihood of the test subject having epilepsy
  • a decrease in Parameter F of the measurement parameter set compared with a healthy reference Parameter F is indicative of an increased likelihood of the test subject having epilepsy
  • a decrease in Parameter G of the measurement parameter set compared with a healthy reference Parameter G is indicative of an increased likelihood of the test subject having epilepsy
  • an increase in Parameter H of the measurement parameter set compared with a healthy reference Parameter H is indicative of an increased likelihood of the test subject having epilepsy.
  • the measurement parameter set comprises Parameter A, Parameter B, Parameter C, Parameter D, wherein each parameter is derived from the EEG recording of the test subject in a resting state.
  • a training method for training a probabilistic machine learning model, PMLM for determining a likelihood of a test subject having epilepsy comprising, the training method comprising: - receiving a plurality of training data sets, wherein: -each training data set has been acquired from a training subject and contains: - a training parameter set that has been measured in the training subject, and - a training tag that is indication of presence or absence of epilepsy in the training subject; - training the PMLM using the plurality of training data sets, wherein: - an input to the PMLM is the training parameter set of the data training set, and an output of PMLM is a pending tag; - the pending tag is compared with the training tag of the training data set, and - the PMLM is adjusted so that the pending tag approaches the training tag of
  • the determining is performed using the measurement data as an input to a trained PMLM, trained according to the training method described herein. According to a further aspect, the determining is performed by comparing the measurement parameter with a healthy parameter set.
  • a computing device or system configured for performing the method as described herein.
  • a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform the method as described herein.
  • a computer readable medium having stored thereon a computer program having instructions which when executed by a computing device or system cause the computing device or system to perform the method as described herein.
  • FIG.1 shows an exemplary arrangement of EEG electrodes on a head of a subject, showing standard electrode locations.
  • FIG. 2 shows a node map containing locations of different EEG electrodes (nodes) and connection strength between the different nodes.
  • FIG. 3 shows a node map containing locations of different EEG electrodes (nodes) and connectivity integration between the different nodes.
  • FIG.4-5 Graphs #1 to #16 show experimental data that are calibration curves for multiple probabilistic machine learning models (#1 to #15), each trained on only one but different parameter. The Brier Score (BrS) of each model is shown in the bottom right corner of the curve.
  • FIG.6-7 Graphs #1 to #16 show experimental data that are calibration curves for multiple probabilistic machine learning models (#1 to #15), each trained with an increasing number of parameters.
  • FIG.8-9 Graphs #1 to #16 show experimental data that are calibration curves for multiple probabilistic machine learning models (#1 to #15), each trained with an increasing number of parameters, but wherein the regional specificity of the parameter has been removed.
  • FIG.10 shows a calibration curve for a finally trained model.
  • FIG.11 shows a confusion matrix and evaluation data for a deterministic machine learning model.
  • FIG. 12 show box plots of predicted probably of epilepsy-, cardiac- and other-diagnosed subjects using the finally trained model of Example 1.
  • FIG.13 shows a calibration curves for data used in Examples 1 and 3.
  • the present invention provides a (computer-implemented) method for determining a likelihood of a test subject having epilepsy comprising: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state, wherein the measurement parameter set comprises at least parameter A from Table 1, - determining from the measurement data the likelihood of the test subject having epilepsy.
  • the test subject is a subject for whom a likelihood of epilepsy is sought, based on the measurement data comprising the measurement parameter set.
  • the test subject is typically human.
  • the test subject has suffered a first attack.
  • the measurement data is typically acquired within 1 to 4 weeks of the first attack.
  • the healthy subject is one that has suffered a first attack, after which a healthy dataset comprising a healthy parameter set was acquired.
  • the healthy dataset was typically acquired within 1 to 4 weeks of the first attack.
  • Subsequent to the first attack during a follow-up period it could be established that the healthy subject was not suffering from epilepsy.
  • the follow-up period was typically at least 2 years.
  • a healthy reference parameter may be measured that is a statistical indicator (e.g.
  • the measurement parameter set of the test subject may be compared with a healthy reference dataset, and depending on the deviation from a corresponding parameter values, a likelihood of the test subject having epilepsy can be determined.
  • the training subject is a subject from whom a training dataset has been obtained comprising a training parameter set and a training tag.
  • the training subject is one that has suffered a first attack, after which the training dataset was acquired.
  • the training dataset was typically acquired within 1 to 4 weeks of the first attack.
  • the first attack is a possible epileptic seizure, characterized by a transient neurological deficit. This neurological deficit can present itself in different ways, including but not limited to loss or alteration of consciousness, uncontrolled movements, and/or sensory hallucinations.
  • the first attack is the initial event based on which the medical doctor considers the possible diagnosis of epilepsy.
  • Epilepsy is a disease of the brain defined according the International League against Epilepsy (ILAE) (Fischer, 2014) by any of the following conditions: - at least two unprovoked (or reflex) seizures occurring >24 h apart, - one unprovoked (or reflex) seizure and a probability of further seizures similar to the general recurrence risk (at least 60%) after two unprovoked seizures, occurring over the next 10 years, - diagnosis of an epilepsy syndrome
  • ILAE International League against Epilepsy
  • Epilepsy is considered to be resolved for individuals who had an age-dependent epilepsy syndrome but are now past the applicable age or those who have remained seizure-free for the last 10 years, with no seizure medicines for the last 5 years.
  • the EEG recording is preferably a conventional EEG recording obtained using a conventional EEG system in which a plurality of electrodes are disposed at different locations on the head of the (test or training) subject so as to record the electrical activity of the brain. Electrode locations and names may be specified according to the International 10–20 system. Each electrode is connected to one input of a differential amplifier (one amplifier per pair of electrodes); a common system reference electrode is connected to the other input of each differential amplifier. The amplifiers amplify the voltage between the active electrode and the reference, and the signal from a channel is digitized via an analog- to-digital converter. Each active electrode provides an EEG signal to a different channel.
  • the plurality of electrodes may be incorporated into a wearable headpiece such as a cap.
  • Exemplary EEG systems of the art include NicoleteOne EEG (Natus), eego (ANT Neuro), RMS Super Spec 32 Channel EEG Machine (RMS), NS-EEG-D1 (Neurostyle), Brain Quick (Micromed), Easy EEG (BrainCapture), Neurofax / Cerebair / airEEG (Nihon Kohden), Arc EEG (Cadwell).
  • An exemplary arrangement of EEG electrodes on a head of a subject is shown in FIG.1.
  • the labelling convention shown in FIG.1 is standard and known in the art.
  • the EEG recording of the (test or healthy or training) subject is recorded during an EEG session.
  • the EEG may last around 10-30 minutes of continuous recording.
  • EEG recording of the (test or healthy or training) subject the subject is in a resting state.
  • resting state it is meant that the subject is under restful conditions such as sitting or lying, and the subject is not exposed to stimuli (i.e., there is no attempt to induce epileptic activity).
  • the eyes may be open or closed.
  • the EEG recording of the (test or training or healthy) subject, comprising multiple EEG signal channels, is preferably stored.
  • the EEG signal from each channel is preferably processed using one or more steps in the following order: - filtering with a high-pass filter (cut-off frequency 1Hz); - filtering with a 50Hz or 60Hz notch filter; - detecting and removing eye movement and eye blink artifacts. They may be detected and removed automatically using independent component analysis (ICA) and matching the components to pre-selected templates. The ICA components were removed if the correlation between the component and one of the templates was larger than 0.95. - average referencing. In average referencing an average of the EEG signals is subtracted from each channel. After average referencing, the overall amplitude across all channels sums up essentially to zero at each time point. - dividing into epochs (3s to 5s long).
  • ICA independent component analysis
  • One or more parameters is measured in a test subject and the measurement parameter set comprises the one or more parameters.
  • One or more parameters may be measured in a population of healthy subjects, and the healthy parameter set (where used) comprises the one or more healthy reference parameters, each healthy reference parameter being a statistical indication of that parameter measured in the population.
  • One or more parameters may be used to populate the training parameter set (where used).
  • Each parameter is determined from one or more EEG signal channels.
  • a parameter may be: - a spectral parameter, - a functional connectivity parameter.
  • a parameter may be particular (limited) to a frequency band.
  • the frequency bands are typically delta ( ⁇ , 1-4 Hz), theta ( ⁇ , 4-8 Hz), alpha ( ⁇ , 8-12 Hz), beta1 ( ⁇ 1, 12-20 Hz), beta2 ( ⁇ 2, 20-30 Hz), gamma ( ⁇ , 30-100 Hz), and broadband (bd, 1-100 Hz).
  • a frequency band may encompass a range of two or more of delta, theta, alpha, beta1, beta2, gamma. The range is typically continuous.
  • a frequency band delta-beta2 is in the range 1-30 Hz.
  • a parameter may be limited to a frequency band.
  • a parameter may or may not be particular to a region of the head of the (test or training) subject.
  • the region corresponds to the location of the EEG electrode(s) on the subject (excluding reference).
  • Typical regions are front left (FL), front right (FR), back left (BL), back right (BR), all regions (AR), one or more of AR (XR).
  • Each region (FL, FR, BL, BR, XR) contains at least one EEG electrode, preferably 2 to 10 EEG electrodes.
  • All region (AR) contains the total of all electrodes (excluding reference).
  • FL contains Fp1, F7, F3, Fz, T3, C3, Cz;
  • FR contains Fp2, F8, F4, Fz, T4, C4, Cz;
  • BL contains T3, C3, Cz, T5, P3, Pz, O1;
  • BR contains T4, C4, Cz, T6, P4, Pz, O2.
  • an epoch parameter value is preferably determined from each epoch (of multiple epochs) of the EEG signal(s).
  • An epoch is a continuous subperiod (e.g.3 to 5 seconds long) of the EEG signal(s).
  • some epochs may be discarded, for instance: - those epochs at the beginning (e.g. first minute) and/or end portion (e.g. final minute) from the EEG recording; - those epochs containing a stimulation event; - those epochs containing on or more artefacts.
  • the multiple epoch parameter values (of the same parameter type) obtained from the multiple epochs are combined to arrive at the parameter value that is a single scalar value for the session.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.
  • a spectral parameter may be one of the following (1.1 to 1.6): (1.1) Relative band power (RBP).
  • the parameter relative band power is a measure of signal power (signal area, power spectral density (PSD) curve) in a frequency band and particular (limited) to a region.
  • the relative band power is determined by measuring signal power (signal area) in the frequency band for each EEG electrode in the region, and combining the measured signal powers into a single scalar value.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g.
  • relative band power is preferably determined from an epoch relative band power of each epoch of the EEG signal of the EEG electrode(s) of the region.
  • the multiple values of epoch relative band powers across the session are combined into a single scalar value.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all an epoch relative band powers in the session.
  • the RBP parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of RBP parameters at multiple different time points, and therefore does not include a time- based component.
  • PSD power spectral density
  • Relative band power and methods for measurement of relative band power are known in the art, for instance, from Laman et al 2005, Wang et al 2016 and Krishnan et al 2020.
  • a spectral parameter (SP) that is relative band power (RBP) may be denoted SP RBP . It may further denote the frequency band: SPRBP, band. It may further denote the region: SPRBP, band, region. (1.2) Peak frequency (PF).
  • the parameter peak frequency is a frequency of the peak (maximal) signal in a frequency band and particular (limited) to a region. It is preferably determined from the power spectral density (PSD) curve.
  • PSD power spectral density
  • the peak frequency is determined by measuring peak frequency in the frequency band for each EEG electrode in the region, and combining the measured peak frequencies into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured signal powers.
  • peak frequency is preferably determined from an epoch peak frequency of each epoch of the EEG signal of the EEG electrode(s) of the region.
  • the multiple values of epoch peak frequencies across the session are combined into a single scalar value.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all an epoch peak frequencies in the session.
  • the PF parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of PF parameters at multiple different time points, and therefore does not include a time-based component.
  • Peak frequency and methods for measurement of peak frequency are known in the art, for instance, from Sörnmo et al 2005.
  • a spectral parameter (SP) that is relative peak frequency (PF) may be denoted SP PF . It may further denote the frequency band: SP PF, fb . It may further denote the region: SP PF, fb, region .
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured signal powers. If a peak is detected at the edge of a frequency band - possibly due to a linear trend in the curve or the absence of a true peak - the value is preferably not included.
  • the PSD curves may be determined using a multitaper estimation method and corrected for a 1/f distribution by multiplying each point by the corresponding frequency (e.g. Bronez, 1992).
  • peak power is preferably determined from an epoch peak power of each epoch of the EEG signal of the EEG electrode(s) of the region.
  • the multiple values of epoch peak powers across the session are combined into a single scalar value.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all an epoch peak powers in the session.
  • the PP parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of PP parameters at multiple different time points, and therefore does not include a time-based component.
  • the parameter peak bandwidth is a width of the peak signal in a frequency band and particular (limited) to a region.
  • the peak bandwidth is determined by measuring peak bandwidth in the frequency band for each EEG electrode in the region, and combining the measured peak bandwidths into a single scalar value.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1 st and 3 rd quartile)) of all the measured signal bandwidths. If a peak is detected at the edge of a frequency band - possibly due to a linear trend in the curve or the absence of a true peak - the value is preferably not included.
  • the PSD curves may be determined using a multitaper estimation method and corrected for a 1/f distribution based on the 1/f exponent (see below in parameter 1.5).
  • the BW parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of BW parameters at multiple different time points, and therefore does not include a time-based component.
  • Peak Bandwidth (BW) and methods for measurement of Peak Bandwidth (BW) are known in the art, for instance, from Bronez, 1992.
  • a spectral parameter (SP) that is Peak Bandwidth (BW) may be denoted BW. It may further denote the frequency band: BW, fb.
  • the parameter 1/f exponent is a slope (exponent) of the decay of the power spectral density (PSD) curve particular (limited) to a region.
  • the 1/f exponent is determined by measuring 1/f exponent for each EEG electrode in the region, and combining the measured 1/f exponents into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all the measured offsets.
  • the PSD curves may be determined using a multitaper estimation method.
  • the EXP parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of EXP parameters at multiple different time points, and therefore does not include a time- based component.
  • Peak frequency and methods for measurement of peak frequency are known in the art, for instance, from Medel et al 2023.
  • a spectral parameter (SP) that is 1/f exponent (EXP) may be denoted SPEXP. It may further denote the frequency band: SP EXP, fb . It may further denote the region: SP EXP, fb, region . (1.6) Offset (OFF).
  • the parameter offset is an amplitude of the constant offset of the power spectral density (PSC) curve particular to a region.
  • the offset is determined by measuring the offset for each EEG electrode in the region, and combining the measured offsets into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured offsets.
  • the PSD curves may be determined using a multitaper estimation method. Offset (OFF) and methods for measurement of offset (OFF) are known in the art.
  • the OFF parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of OFF parameters at multiple different time points, and therefore does not include a time- based component.
  • a spectral parameter (SP) that is offset (OFF) may be denoted OFF. It may further denote the frequency band: OFF, fb. It may further denote the region: OFF, fb, region. It is appreciated that the one or more multiple different spectral parameters may be measured in the subject, optionally each measured in a different frequency band and/or in a different region.
  • Functional connectivity is determined from the EEG signals to include information on how the brain regions communicate with each other.
  • a functional connectivity parameter may be a connectivity strength (CS) functional connectivity parameter.
  • a CS functional connectivity parameter is limited to a region which may be one of: FL, FR, BL, BR, AR.
  • the CS functional parameter is determined by measuring a strength of connection (connectivity strength, CS) between all electrode pairs in the region, and combining the measured connectivity strengths into a single scalar value. The combining may be performed by a statistical method e.g.
  • the connectivity strength between pairs of electrodes may be determined by several different methods. Examples of methods known in the art for determining connectivity strength include: Imaginary coherence (IMCOH), Phase locking value (PLV), Partial directed coherence (PDC), and Amplitude envelope correlation (AEC). These have been described in at least Bastos et al, 2016, and Briels et al 2020.
  • Connectivity strength (CS) functional connectivity parameter and methods for its measurement are generally known in the art, for instance, in Sporns 2022. An example of how connectivity strength (CS) functional connectivity parameter is determined is provided in FIG.2.
  • a plurality of different EEG electrodes are represented (1 to 5), and connectivity strengths (CS) between them labelled.
  • a value of the connectivity strength (CS) functional connectivity parameter may be determined as an average of the individual connectivity strengths (CS).
  • CS functional connectivity parameter is preferably determined from an epoch CS functional connectivity parameter of each epoch of the EEG signal(s) of the EEG electrode(s) of the region.
  • the multiple values of epoch CS functional connectivity parameters across the session are combined into a single scalar value.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all epoch CS functional connectivity parameters in the session.
  • FCPCS connectivity strength
  • a functional connectivity parameter may be a global efficiency (GE) functional connectivity parameter.
  • a GE functional connectivity parameter is limited to a region which may be one of: FL, FR, BL, BR, AR.
  • the GE functional parameter is determined by measuring an integration (togetherness) of connections (connectivity integration, CI) between all EEG electrode pairs in the region, and combining the measured connectivity integrations (CI) into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) this of all the measured connectivity integrations.
  • the connectivity integration between a particular pair of EEG electrodes is determined from the connection strengths between all EEG electrodes pairs (e.g. a to g) in the region, and determining the shortest path length (connection strength-1) between the particular pair of EEG electrodes which may be a direct path (e.g. a-b) or an indirect path via one or more other EEG electrodes (e.g. a-c-b) based on the a summation of CI values along the path which gives the lowest value.
  • Global efficiency is know in the art Examples of methods known in the art for determining connectivity strength and hence connectivity integration (which is an inverse of connectivity strength) include: Imaginary coherence (IMCOH), Phase locking value (PLV), Partial directed coherence (PDC), Amplitude envelope correlation (AEC). These have been described in at least Bastos et al, 2016, and Briels et al 2020.
  • IMCOH Imaginary coherence
  • PLC Phase locking value
  • AEC Amplitude envelope correlation
  • a plurality of different EEG electrodes are represented (1 to 5), and connectivity integration (CI) (CS -1 from FIG.2) between them labelled.
  • CI (1,2) 1.25
  • CI (1,5) 10,
  • CI (2,3) 1.43
  • CI (2,4) 2.5
  • CI (2,5) 2.5
  • the shortest path length between EEG electrode pairs is determined according to the path which yield the smallest summation of CI values along that path.
  • SPL shortest path length
  • GE functional connectivity parameter is preferably determined from an epoch GE functional connectivity parameter of each epoch of the EEG signal(s) of the EEG electrode(s) of the region.
  • the multiple values of epoch GE functional connectivity parameters across the session are combined into a single scalar value.
  • the combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g.
  • FCP function connectivity parameter
  • FCPGE function connectivity parameter
  • CS connectivity strength
  • CI connectivity integration
  • Pairs of EEG signals (x and y) are considered, one from EEG signal channel x, and one from EEG signal channel y.
  • Each EEG channel contains an EEG signal from a different EEG electrode, and the electrodes are located in different positions on the head.
  • Imaginary coherence (IMCOH) Connectivity strength of an EEG electrode pair may be determined from imaginary coherence (IMCOH) of the EEG electrode pairs.
  • the imaginary coherence (IMCOH) is computed from a normalized cross-spectrum between two signals (e.g. Nolte, 2004) according to Eq.1: where S xy (f) is the cross-spectral density between signals x and y, and S xx (f) and S yy (f) are auto-spectral densities of signals x and y respectively.
  • S xy (f) is the cross-spectral density between signals x and y
  • S xx (f) and S yy (f) are auto-spectral densities of signals x and y respectively.
  • the connectivity strength (IMCOH) for an x,y pair is determined from the value Cxy(f)
  • connectivity integration for an x,y pair is determined from the value (Cxy(f)) -1 .
  • Phase value (PLV) Connectivity strength of an EEG electrode pair may be determined from a phase locking value (PLV) of the EEG electrode pairs.
  • the parameter phase locking value (PLV) is the most common phase synchrony measure and aims to detect signals with a phase difference that is stable over time.
  • the PLV is the absolute value of the averaged phase difference, expressed as a complex unit-length vector (e.g. Lachaux, 1999) according to Eq.2: Where PLVxy is a phase locking value between signals x and y, ⁇ x(t) and ⁇ y(t) are the phases of signals x and y respectively at time-point t, T is the total length of the signals, and i is the imaginary unit.
  • the connectivity strength (PLV) for an x,y pair is determined from the value PLV xy
  • connectivity integration for an x,y pair is determined from the value (PLV xy )- 1.
  • PDC Partial directed coherence
  • the parameter partial directed coherence (PDC) uses the Fourier transform of an autoregressive model to quantify the fraction of the spectral power of signal x that contributes to the future of signal y. PDC detects only direct interactions between signals and is normalized with respect to the total outflow (e.g. Baccalá, 2001) according to Eq.3: Where PDCxy(f) is the partial directed coherence between signals x and y, evaluated at frequency f.
  • a xy (f) is the element at position (x,y) taken from the matrix A(f), which is the Fourier transform of the coefficient matrix from an autoregressive model fitted to signals x and y (e.g. Baccalá, 2001).
  • the partial directed coherence (PDC) for an x,y pair is determined from the value PDCxy(f), and connectivity integration for an x,y pair is determined from the value (PDCxy(f)) -1 .
  • the (measurement or training) parameter set preferably comprises at least one parameter (A to I) from Table 1.
  • the (measurement or training) parameter set may consist of at least one parameter (A to I) selected from Table 1.
  • Parameter Parameter Parameter Parameter Type Frequency Preferred Abbreviation Name band Region (optional) A RBP delta Spectral, relative Delta FR SPRBP,delta,(FR) band power B GE functional Functional Delta all regions FCPGE,delta,(AR) connectivity delta connectivity, global band efficiency C CS functional Functional Delta BR & FL FCPCS,delta,(BR,FL) connectivity delta connectivity, band connectivity strength D EXP Spectral, 1/f Broadband FL SPEXP,bd, (FL) exponent E GE functional Functional Broadband FR FCPGE,bd,(FR) connectivity connectivity, global broad band efficiency F PF broadband Spectral, peak Broadband BL SPPF, bd, (BL) frequency G RBP gamma Spectral, relative Gamma FR SPRBP,gamma,(FR) band power H PF theta Spectral, peak Theta FR SPPF,theta,(FR) frequency I RBP theta Spectral, relative Theta FR
  • the measurement data of the test subject comprises a measurement parameter set.
  • the measurement parameter set preferably comprises at least parameter A in Table 1.
  • the measurement parameter set preferably comprises at least parameters A and B in Table 1.
  • the measurement parameter set preferably comprises at least parameters A to C in Table 1.
  • the measurement parameter set preferably comprises at least parameters A to D in Table 1.
  • the measurement parameter set preferably comprises at least parameters A to E in Table 1.
  • the measurement parameter set preferably comprises at least parameters A to F in Table 1.
  • the measurement parameter set preferably comprises at least parameters A to G in Table 1.
  • the measurement parameter set preferably comprises at least parameters A to H in Table 1.
  • the measurement parameter set preferably comprises at least parameters A to I in Table 1.
  • the likelihood of epilepsy in the test subject may be determined by comparing the measured parameter(s) with a healthy reference parameter(s).
  • the likelihood of epilepsy in the test subject may be determined by comparing the measurement parameter set of the test subject with the healthy parameter set.
  • the likelihood of epilepsy in the test subject may be determined by comparing a value of a parameter (e.g. one of A to I) for the test subject with the same (healthy reference) parameter of the healthy parameter set.
  • the comparison may show no deviation or a deviation (an increase, or a decrease) compared to a healthy subject. Where the deviation is an increase or a decrease, depending on the parameter, the test subject has an increased likelihood of epilepsy.
  • Table 2 sets out the deviation (increase or decrease) in value for each parameter for a test subject compared with a healthy subject that is linked to an increased likelihood of epilepsy in the test subject.
  • RBP delta SP RBP,delta,(FR) Increase B GE functional connectivity delta band
  • FCPGE,delta,(AR) Increase C CS functional connectivity delta band
  • FCPCS,delta,(BR,FL) Increase D EXP broadband SP EXP, bd,
  • FL) Increase E GE functional connectivity broad band
  • FCP GE,bd,(FR) Increase F PF broadband SP PF, bd,
  • BL Decrease G RBP gamma SP RBP,gamma,(FR) Decrease H PF theta SPPF,theta,(FR) Decrease I RBP theta SPRBP,theta,(FR) Increase Table 2: List of parameters and deviations.
  • the deviation is a deviation compared with healthy reference.
  • a deviation indicating an increase means that when there is an increase in the parameter value for the test subject compared with the same healthy reference parameter, the test subject has an increased likelihood of epilepsy.
  • a deviation indicating a decrease means that when there is a decrease in the parameter value for the test subject compared with the same healthy reference parameter, the test subject has an increased likelihood of epilepsy.
  • One or more of the parameters are able to predict epilepsy in a test subject that has suffered a first attack. Where multiple test parameters have been measured in the test subject, the likelihood may be determined based on the deviation (absolute difference between the test parameter values and the healthy reference values. A weighted combination of these deviations (differences) results in an epilepsy risk index (ERI).
  • ERP epilepsy risk index
  • the healthy parameter set comprises one or more healthy reference parameters (e.g. one of healthy reference parameters A to I).
  • a healthy reference parameter is a statistical indicator of that parameter in a healthy population. More in particular, a healthy reference parameter (e.g. healthy reference parameter A) is determined by measuring the same parameter (e.g.
  • the healthy parameter set comprises one or more healthy reference parameters each determined from a statistical indicator of a parameter in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A, determined from a statistical indicator of parameter A in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A and B, each determined from a statistical indicator of parameter A and B respectively in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A to C, each determined from a statistical indicator of parameter A to C respectively in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A to D, each determined from a statistical indicator of parameter A to D respectively in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A to E, each determined from a statistical indicator of parameter A to E respectively in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A to F, each determined from a statistical indicator of parameter A to F respectively in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A to G, each determined from a statistical indicator of parameter A to G respectively in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A to H, each determined from a statistical indicator of parameter A to H respectively in Table 1 in a healthy population.
  • the healthy parameter set preferably comprises at least healthy reference parameter A to I, each determined from a statistical indicator of parameter A to I respectively in Table 1 in a healthy population.
  • exemplary healthy reference parameter values are provided in Table 3. These values may be calculated by the skilled person using the methods described herein applied to a group of healthy subjects. It is understood that the reference values are given for general guidance, and may be refined depending on the population group (e.g. child, adult).
  • a probabilistic machine learning model, PMLM may optionally be used to determine the likelihood of the test subject having epilepsy.
  • Multiple training datasets are used to train the PMLM, wherein each training dataset of a training subject comprises a training parameter set and a training tag.
  • a training method for training a probabilistic machine learning model, PMLM, for determining a likelihood of a subject having epilepsy comprising, the training method comprising: - receiving a plurality of training data sets, wherein: -each training data set has been acquired from a training subject and contains: - a training parameter set that has been measured in the training subject in a resting state, and - a training tag (e.g.1 or 0) that is indication of presence (1) or absence (0) of epilepsy in the training subject; - training the PMLM using the plurality of training data sets, wherein: - an input to the PMLM is the training parameter set of the data training set, and an output of PMLM is a pending tag (probabilistic); -
  • a nested cross-validation scheme may be used to train and adjust the PMLM.
  • the PMLM is adjusted based on a performance metric.
  • a performance metric is a Brier Score (BrS).
  • a training tag is translated into a 1 or 0 (1 - epilepsy present or 0 - epilepsy absent in the training subject).
  • the BrS is used to compare the pending tag - the probabilistic output of the PMLM - to the training tag, using for instance, Eq.5: wherein p i the pending tag i.e. predicted PMLM output probability for training subject i. It may have a value between 0 and 1; o i the training tag for training subject i.
  • PMLM may have a value of 1 (epilepsy present) or 0 (epilepsy absent) in the training subject; N is the number of training subjects.
  • the PMLM is penalised more if the predicted probability (pi) is further away from the training tag value (oi). As the predictions get closer to the training tags the BrS becomes smaller, so training is based on minimising the Brier Score.
  • Any machine learning models that can give a graded output score and trained to predict probabilities may be used as a PMLM. Examples of suitable candidates for PMLM include Logistic Regression, Support Vector Machine, and Na ⁇ ve Bayes.
  • PMLMs and nested cross-validation schemes have been described in Ferro et al., Comparing Probabilistic Forecasting Systems with the Brier Score, Weather and Forecasting; T. Proix et al., Forecasting seizure risk in adults with focal epilepsy: a development and validation study; B.
  • the trained PMLM may be calibrated to further improve the performance.
  • the goal of probability calibration is to fit a simple model to the predicted probabilities and cancel this model out to bring the probabilities closer to the diagonal line on the calibration curve.
  • Calibration may be achieved by any method of the art, such as Platt’s scaling or isotonic regression.
  • Platt scaling, a sigmoid function is used as regression model.
  • Isotonic regression a free-form line is fitted to the calibration curve that is non-decreasing (i.e., the line is always increasing).
  • Each training data set has been acquired from a training subject.
  • a training dataset comprises - a training parameter set that has been measured in the training subject, and - a training tag (e.g.1 or 0) that is indication of presence (1) or absence (0) of epilepsy in the training subject.
  • the training parameter set comprises one or more of the parameters of Table 1.
  • the training parameter set preferably comprises at least parameter A in Table 1.
  • the training parameter set preferably comprises at least parameters A and B in Table 1.
  • the training parameter set preferably comprises at least parameters A to C in Table 1.
  • the training parameter set preferably comprises at least parameters A to D in Table 1.
  • the training parameter set preferably comprises at least parameters A to E in Table 1.
  • the training parameter set preferably comprises at least parameters A to F in Table 1.
  • the training parameter set preferably comprises at least parameters A to G in Table 1.
  • the training parameter set preferably comprises at least parameters A to H in Table 1.
  • the training parameter set preferably comprises at least parameters A to I in Table 1.
  • the parameters present in the training parameter set are preferably the same as the parameters present in the measurement parameter set. For instance, where the trained PMLM has been trained using a training parameter set comprising parameters A to D, the measurement parameter set using as test input to the trained PMLM also comprises parameters A to D measured from the test subject. Typically, the number of training subjects is 200 or more.
  • the determining from the measurement data the likelihood of the test subject having epilepsy may comprise using the trained PMLM.
  • a method for determining a likelihood of a test subject having epilepsy comprises: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state, wherein the measurement parameter set comprises at least Parameter A of Table 1; - using the measurement data as an input for a trained PMLM, the trained PMLM trained according to the training method described herein; - outputting from the trained PMLM, a likelihood of the test subject having epilepsy.
  • deterministic machine learning is used to predict whether a subject has dysfunction or not.
  • the inventors have found that deterministic machine learning applied to the problem of epilepsy prediction produces a very low level of predictive accuracy.
  • Example 2 a deterministic machine learning model
  • PMLM probabilistic machine learning model
  • Example 1 a sensitivity of 36% and specificity of 77%, which is in line with the current diagnostic yield of EEG in clinical practice based on visual analysis of the EEG signals, with sensitivity ranging between 25-56% and specificity between 78-98% (Smith et al., 2005). It did not, however, provide an improvement over the current accuracies, and its performance therefore remains unsatisfactory.
  • the PMLM has been developed in the field of meteorology to predict non-deterministic events such as weather forecasting. Because the method quantifies the uncertainty of predictions, it lends itself well to event forecasting - quantifying how likely it is that an event will happen within a specific timeframe. Assigning a subject to a diagnostic group, however, is not considered a routine application of a PMLM because a diagnosis is neither considered to be non-deterministic (either a subject belongs to a disease group or does not) nor a prediction about a future event.
  • PMLM can be applied to the problem of epilepsy likelihood, and it provides an accurate prediction, even though PMLM is not being used to predict an event.
  • highly predictive parameters (parameter A to H) in Table 1) that can be measured in the resting state of the subject, namely, there is no need to try to induce epileptic activity in the subject.
  • Diagnostic labels for subjects have not previously been considered to be non-deterministic, because the subject either belongs to one group or they do not and - given enough evidence - the subject can be assigned to the correct group without uncertainty. Switching to a graded output (likelihood) that returns a probability index of a diagnostic label is not routine, yet the inventors have found that it produces more predictive accuracy compared with the deterministic approach.
  • the graded output provided by the present method quantifies the uncertainty of the prediction, which can be used to decide on the type of treatment e.g. less invasive (drug) or more invasive (surgical), as a function of the likelihood of the epilepsy.
  • the methods mentioned herein, wherein: - the measurement parameter set comprises at least parameter A in Table 1, or - the measurement parameter set comprises at least parameters A and B in Table 1 - the measurement parameter set comprises at least parameters A to C in Table 1 - the measurement parameter set comprises at least parameters A to D in Table 1 - the measurement parameter set comprises at least parameters A to E in Table 1 - the measurement parameter set comprises at least parameters A to F in Table 1 - the measurement parameter set comprises at least parameters A to G in Table 1 - the measurement parameter set comprises at least parameters A to H in Table 1 - the measurement parameter set comprises at least parameters A to I in Table 1 may be applied to one or more methods as described in the aspects below.
  • a method of determining a likelihood of a test subject having epilepsy as described elsewhere herein, which method further comprises: - comparing a value(s) of measurement parameter set with value(s) of healthy parameter set, - finding a deviation or no deviation of the value(s) of measurement parameter set with value(s) of healthy parameter set; and - attributing said finding of deviation or no deviation to a likelihood of the test subject having epilepsy.
  • a method for the selection of a prophylactic or therapeutic treatment for epilepsy comprises: - exposing one or more subjects to the prophylactic or therapeutic treatment, - comparing a value(s) of measurement parameter set with value(s) of healthy parameter set in the test subject prior to and after exposing to the subject to the prophylactic or therapeutic treatment.
  • the selection of a prophylactic or therapeutic treatment is determined based on a deviation or no deviation of the value(s) of measurement parameter set after before and after the exposing.
  • a method of assessing an efficacy of a therapeutic treatment further comprises: - exposing the one or more test subjects to the prophylactic or therapeutic treatment, - comparing a value(s) of measurement parameter set of the one or more test subjects prior to and after exposing to the subject to the prophylactic or therapeutic treatment.
  • the efficacy of the therapeutic treatment is determined based on a deviation or no deviation of the value(s) prior to and after the exposing.
  • a system for determining a likelihood of a subject having epilepsy configured to carry out the method as described herein.
  • the system may include computing device (described below) and a plurality of EEG electrodes.
  • a wearable headset comprising a plurality of EEG electrodes for determining a likelihood of a subject having epilepsy, configured to carry out the method as described herein.
  • the wearable headset may comprise a computing device (described below) or connection (wireless or via one or more cables) to a computing device.
  • the plurality of EEG electrodes is operatively connected to the computing device.
  • operatively connected it is meant that the computing device is able to receive signals from the plurality of EEG electrodes in order to carry out a method as described herein.
  • the method is typically automatic, meaning there is no intervention or monitoring by the practitioner.
  • the method described herein is a computer implemented method. The method is in vitro, more in particular ex vivo. The method is an offline method.
  • the method may be performed on stored measurement data.
  • a computing device or system configured for performing a method as described herein.
  • a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform a method as described herein.
  • a computer readable medium having stored thereon a computer program (product) having instructions which when executed by a computing device or system cause the computing device or system to perform (each of the steps of) the method as described herein.
  • a data stream which is representative of a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform (each of the steps of) the method as described herein.
  • the method may be performed using a standard computer system such as an Intel Architecture IA-32 based computer system 2, and implemented as programming instructions of one or more software modules stored on non-volatile (e.g. hard disk or solid-state drive) storage associated with the corresponding computer system.
  • a standard computer system such as an Intel Architecture IA-32 based computer system 2
  • non-volatile (e.g. hard disk or solid-state drive) storage associated with the corresponding computer system.
  • non-volatile e.g. hard disk or solid-state drive
  • ASICs application-specific integrated circuits
  • the method or system may produce an output that is: - displayed on a screen, or - saved to a file.
  • the present method of system may be regarded as a method for measuring or detecting the one or more parameters.
  • a system or kit is provided for use in a method described herein.
  • a use of a system is provided for use in a method described herein, wherein said device comprises at least one EEG electrode.
  • a use of a system is provided for use in a method described herein, wherein said device comprises an EEG system.
  • the system or kit or system or kit for use is a method may further comprise one or more of: - a computer system for executing the method described herein; - an EEG system.
  • the feature selection was based on an ANOVA F-test where at each iteration a comparison was made between parameter values in epilepsy patients and non-epilepsy patients. The parameter with the lowest p-value was selected. In addition, parameters with a strong correlation with the newly selected feature (r > 0.80) were removed from the list of possible candidates for future iterations. Table 4 lists the top 15 selected parameters, with the associated p-value and effect size (d), and whether the feature is increased or decreased in epilepsy patients.
  • the effect size is a more representative indication of the distance between the distributions of the two groups (i.e., epilepsy vs non-epilepsy) and is computed as the difference of the means, normalized by the grouped standard deviation.
  • the effect size may be calculated according to Eq.6: [Eq.6] where ⁇ 1 / ⁇ 2 and ⁇ 1 / ⁇ 2 are the mean and standard deviation of the parameter values of the two patient groups.
  • ANOVA Parameter Abbreviation Effect p-value Deviation Coefficient Ranking name size 1 (A) RBP delta SPRBP,delta,(FR) 0.55 4.76e- inc 1.10 (FR) 13 2 (B) GE functional FCPGE,delta,(BR), PLV 0.44 7.60e- inc 1.85 connectivity 09 delta band (BR) PLV 3 (C) CS functional FCPCS,delta,(BR), 0.43 1.26e- inc 2.09 * 10 -2 connectivity IMCOH 08 delta band (BR) IMCOH 4 (D) EXP (FL) SPEXP,bd, (FL) 0.39 5.39e- inc 2.75 * 10 -2 08 5 (F) PF broadband SPPF, bd, (BL) 0.39 2.81e- dec 6.62 * 10
  • the letter in brackets (A to I) in the first column is the parameter name in Table 1.
  • the Deviation column indicates whether an increase (inc) or decrease (dec) in the parameter value compared with a non-epileptic subject was evident in the epileptic subject. The effect size, p-value and coefficient (weight) per parameter are also indicated.
  • PMLM probabilistic machine learning model
  • the model was optimised to predict the conditional probabilities of the training subject belonging to a certain class (subject suffering from epilepsy or subject not suffering from epilepsy). This was quantified with the Brier Score (BrS), which compared the output probability from the Logistic Regression model with the actual probability.
  • the actual probability for a subject suffering from epilepsy was 1, and the actual probability for a subject not suffering from epilepsy was 0.
  • the score was computed according to the following formula (Eq.7): with N being the number of training subjects, pi the output probability from the model, and oi the actual (observed) probability of training subject i.
  • the Brier Score becomes lower as the predictions pi correspond better with the observations oi, so during training the BrS is minimized.
  • the aim of the model training is hence to lower the BrS.
  • the nested cross-validation comprised two 5-fold cross-validation schemes.
  • the data was split into 5 parts (also called folds) of equal size and each containing the same proportion of epilepsy and non-epilepsy patients.
  • one fold is set aside to be used as test set, the other four folds are taken together and again divided into 5 equal folds for the inner cross-validation.
  • the inner cross-validation is used for optimizing and calibrating the classifier
  • the outer cross-validation is used to test the performance on unseen data.
  • FIGs.4-5 show the performance when multiple PMLMs were trained, one PMLM for each individual training parameter of Table 4.
  • the name of the graph corresponds to the ANOVA ranking (e.g.1) in Table 4 and to the training parameter used to train that PMLM.
  • the next most predictive PMLMs are shown in order (FIG.4 to 5, #2 to #15).
  • FIG.6-7 (#1 to #16) shows the performance when multiple PMLMs were trained, each PMLM trained with an increasing number of different parameters. From the name of the graph (#1 to #16), it can be determined the number of training parameters and which different training parameters were used to train the PMLM.
  • FIG.6 #3 used 3 different training parameters which had ANOVA rankings 1 to 3 in Table 4.
  • FIG.6 #4 used 4 different training parameters which had ANOVA rankings 1 to 4 in Table 4 etc.
  • BS calibration curve and Brier Score
  • FIG.8-9 shows the performance when multiple PMLMs were trained, each with an increasing number of different parameters, but when the regional specificity of the parameter was removed. Thus, each parameter was averaged over all regions (AR) (and thus not confined to one quadrant of the scalp FL, FR, BL, BR).
  • FIG.8 #3 used 3 training parameters which had ANOVA rankings 1 to 3 in Table 4.
  • FIG.8 #4 used 4 training parameters which had ANOVA rankings 1 to 4 in Table 4 etc.
  • the performance of each PMLM is comparable to the results obtained in FIG.4-5.
  • the results show that a parameter defining specific spatial regions might be preferable but it is not essential.
  • a ranked list of the most predictive parameters could be determined, which separately or in combination, are indicative of a likelihood of epilepsy in a test subject. The list is presented in Table 1.
  • the PMLM was trained, it was calibrated to further improve the accuracy of the probability predictions.
  • Platt’s scaling method was used, where unwanted distortions in the calibration curve of the PMLM were reduced by fitting a sigmoid function to the curve and rescaling the curve to the desired trend (Niculescu-Mizil, 2005).
  • a part of the large retrospective dataset of EEGs mentioned in Example 1 was also used to validate the PMLM.
  • the predicted probabilities on the test folds were put together and visualized with a calibration curve as shown in FIG.10.
  • the calibration curve was generated from a calibrated PMLM trained using 12 parameters (the 12 highest ranked parameters in Table 4).
  • the calibration curve of the trained model shows a good correspondence between the predicted probabilities and the actual fraction of positives.
  • Example 2 A deterministic machine learning model to classify epilepsy subjects vs non-epilepsy subjects was trained. The first steps (parameters, standardization, iterative feature selection) were the same as in Example 1. A 5-fold cross-validation scheme was used to train and evaluate a Na ⁇ ve Bayes classifier. The classifier was optimized based on the area under the receiver operating curve (AUC).
  • the performance of the trained model was expressed in terms of sensitivity and specificity, and a confusion matrix that shows the predicted labels vs the true labels.
  • the resulting performance of the deterministic machine learning model is shown in FIG.11. It achieved an AUC of 0.60, sensitivity of 36% and specificity of 77%, which is in line with the current diagnostic yield of EEG in clinical practice based on visual analysis of the EEG signals, with sensitivity ranging between 25-56% and specificity between 78-98% (Smith et al., 2005). It did not, however, provide an improvement over the current accuracies.
  • Example 3 A new dataset was used to validate the calibration curve of Example 1.
  • the new data set contained 137 patients, each of whom had presented as potentially having first attack of epilepsy, and had not been used to train the PMLM of Example 1.
  • 137 patients 89 patients (epilepsy group) were later diagnosed to have epilepsy based on a follow-up study of at least 2 years duration, and 48 were shown to be suffering from other conditions (control groups having a later diagnosis of a “cardiovascular” or “other” health related problem) and not epilepsy.
  • the EEG recordings were collected for several hours (long-term EEG), compared with Example 1 where the EEG recordings were collected for 15 minutes (routine EEG). Twenty parameters were measured in each patient and applied to the calibration curve of Example 1.
  • PF delta PF delta
  • PF delta PF delta
  • BL PF delta
  • PF delta PF delta
  • PF theta BL
  • PF theta PF theta
  • FR PP theta
  • BW theta BL
  • FR BW alpha
  • RBP alpha BR
  • PF beta1 FR
  • BW beta1 FL
  • RBP beta1 FR
  • PF beta2 BW beta2
  • BW beta2 BW beta2
  • BW beta2 BW 1-30Hz
  • OFF FL
  • the results are shown in FIGs.12 and 13.
  • the epilepsy group showed a predicted probably of epilepsy (0.65) greater than the control group (cardiovascular (0.6) and other (0.59)).
  • calibration curves show good correspondence between the new dataset (long term EEG) with the dataset used in Example 1 to validate the calibration curve (routine term EEG), and both calibrations curves are close to the diagonal.
  • the model exposed to a completely different dataset, still reliably predicts, based on spectral and/or connectivity features from a resting state EEG recording, what the probability is that a patient has/will develop epilepsy.

Landscapes

  • Health & Medical Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • General Health & Medical Sciences (AREA)
  • Pathology (AREA)
  • Physics & Mathematics (AREA)
  • Biophysics (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Artificial Intelligence (AREA)
  • Neurology (AREA)
  • Veterinary Medicine (AREA)
  • Animal Behavior & Ethology (AREA)
  • Surgery (AREA)
  • Molecular Biology (AREA)
  • Neurosurgery (AREA)
  • Physiology (AREA)
  • Psychiatry (AREA)
  • Mathematical Physics (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Signal Processing (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Primary Health Care (AREA)
  • Fuzzy Systems (AREA)
  • Psychology (AREA)
  • Evolutionary Computation (AREA)
  • Epidemiology (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

Provided is a method or system method for determining a likelihood of a test subject having epilepsy after having suffered a possible epileptic seizure, comprising: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state wherein the measurement parameter set comprises Parameter A of Table 1, - determining from the measurement data the likelihood of the test subject having epilepsy.

Description

METHOD AND SYSTEM FOR PREDICTION OF EPILEPSY Field of the invention The present invention is in a field of diagnosing epilepsy in a subject after a suspected first seizure using an electroencephalogram (EEG). Background to the invention Epilepsy is a chronic disorder of the brain, characterized by the unprovoked recurrence of seizures. It affects approximately 1% of the global population, making it the most common neurological disorder after migraine. Diagnosis and management of patients who experienced a first seizure with transient neurological deficit or loss of consciousness remains a challenging task for the consulting neurologist, evidenced by the fact that not more than 50% of patients are diagnosed appropriately (Krumholz, Neurology, 2007). Errors committed in epilepsy diagnosis and seizure classification are known to lead to inappropriate decisions on the use or choice of anti-epileptic drugs and to other serious patient management errors. Differential diagnosis of epilepsy encompasses psychogenic events, transient ischemic attacks, migraine, cardiac events, etc. Clinical history, electroencephalography (EEG), and neuroimaging (more specifically MRI) are the fundamental modalities to achieve a correct diagnosis. However, 50%-99% of the MRIs and up to 80% of scalp EEGs (Bouma, European Journal of Neurology, 2016) may be negative, despite an underlying epileptic disorder. EEG plays a central role in the diagnosis and management of epilepsy and the presence of transient events such as interictal epileptiform discharges (IEDs) is one of the diagnostic hallmarks of epilepsy. Photic stimulation, hyperventilation and a period of attempted sleep are often included in 20-30min routine EEG recordings to increase the occurrence probability of these IEDs. Despite these extra tests, the sensitivity of recording an IED with routine EEG remains low (29-55%). Longer recordings either with prolonged recording times or multiple EEG recordings can increase the yield of IEDs but are impractical for clinical application (Burkholder, Neurology, 2016). This invention uses advanced EEG analysis techniques to extract parameters that objectively describe the characteristics of the EEG signals and combines them into an epilepsy risk index. In view of the problems in the art, it is an aim of the present invention to overcome the following problems: - Low diagnostic accuracy (more specifically low sensitivity) of routine EEG recordings, caused mainly by the low chance of recording epileptic activity; - Lack of a tool that extracts quantitative EEG measures that objectively represent the underlying characteristics in resting state EEG signals (beyond visual interpretation); - Lack of a computer-assisted diagnostic tool that provides an interpretable prediction of epilepsy diagnosis. Summary of the invention Provided is a method for determining a likelihood of a test subject having epilepsy after having suffered a possible epileptic seizure, comprising: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state wherein the measurement parameter set comprises Parameter A of Table 1, - determining from the measurement data the likelihood of the test subject having epilepsy. The method is a computer-implemented method. Parameter A is relative band power, RBP, of a power spectral density, PSD, curve limited to a delta frequency band and limited to a front right region of the head of the subject. According to one aspect, an increase in Parameter A of the measurement parameter set compared with a healthy reference Parameter A is indicative of an increased likelihood of the test subject having epilepsy. According to a further aspect, the measurement parameter set further comprises one or more of Parameter B, Parameter C, Parameter D, wherein each parameter is derived from the EEG recording of the test subject in the resting state. Parameter B is a global efficiency functional connectivity parameter that is a statistical combination of (SPLs)-1 limited to a delta frequency band, in all regions of the head of the subject, wherein the SPL is a shortest path length between EEG electrode pairs in all regions of the head of the subject determined according to the path which yields a smallest summation of inverse connectivity strengths, CS-1, values along that path. Parameter C is a connectivity strength functional connectivity parameter that is a statistical combination of connectivity strengths, CS, limited to a delta frequency band between EEG electrodes pairs limited to a back right, BR, and/or front left, FL, region of the head of the subject. Parameter D is a 1/f exponent parameter that is a statistical combination of slope of a decay of a power spectral density, PSD, curve limited to a broadband frequency band for each EEG electrode limited to a front left (FL), region of the head of the subject. According to a further aspect, an increase in Parameter B of the measurement parameter set compared with a healthy reference Parameter B is indicative of an increased likelihood of the test subject having epilepsy, and/or an increase in Parameter C of the measurement parameter set compared with a healthy reference Parameter C is indicative of an increased likelihood of the test subject having epilepsy, and/or an increase in Parameter D of the measurement parameter set compared with a healthy reference Parameter D is indicative of an increased likelihood of the test subject having epilepsy. According to a further aspect, the measurement parameter set further comprises one or more of Parameter E, Parameter F, Parameter G, Parameter H, wherein each parameter is derived from the EEG recording of the test subject in the resting state. Parameter E is a global efficiency functional connectivity parameter that is a statistical combination of (SPLs)-1 limited to a broadband frequency band, and limited to a front right (FR), of the head of the subject, wherein the SPL is a shortest path length between EEG electrode pairs in all regions of the head of the subject determined according to the path which yields a smallest summation of inverse connectivity strengths, CS-1, values along that path. Parameter F is a peak frequency parameter that is a statistical combination of frequency of a peak (maximal) signal of a power spectral density, PSD, curve limited to a broadband frequency band for each EEG electrode limited to a back left, BL, region of the head of the subject. Parameter G is relative band power, RBP, of a power spectral density, PSD, curve limited to a gamma frequency band limited to a front right region of the head of the subject. Parameter H is a peak frequency parameter that is a statistical combination of frequency of a peak (maximal) signal of a power spectral density, PSD, curve limited to a theta frequency band for each EEG electrode limited to a front right, FR, region of the head of the subject. According to a further aspect: a decrease in Parameter E of the measurement parameter set compared with a healthy reference Parameter E is indicative of an increased likelihood of the test subject having epilepsy, and/or a decrease in Parameter F of the measurement parameter set compared with a healthy reference Parameter F is indicative of an increased likelihood of the test subject having epilepsy, and/or a decrease in Parameter G of the measurement parameter set compared with a healthy reference Parameter G is indicative of an increased likelihood of the test subject having epilepsy an increase in Parameter H of the measurement parameter set compared with a healthy reference Parameter H is indicative of an increased likelihood of the test subject having epilepsy. According to a further aspect, the measurement parameter set comprises Parameter A, Parameter B, Parameter C, Parameter D, wherein each parameter is derived from the EEG recording of the test subject in a resting state. Further provided is a training method for training a probabilistic machine learning model, PMLM, for determining a likelihood of a test subject having epilepsy comprising, the training method comprising: - receiving a plurality of training data sets, wherein: -each training data set has been acquired from a training subject and contains: - a training parameter set that has been measured in the training subject, and - a training tag that is indication of presence or absence of epilepsy in the training subject; - training the PMLM using the plurality of training data sets, wherein: - an input to the PMLM is the training parameter set of the data training set, and an output of PMLM is a pending tag; - the pending tag is compared with the training tag of the training data set, and - the PMLM is adjusted so that the pending tag approaches the training tag of the training set. According to an aspect, the determining is performed using the measurement data as an input to a trained PMLM, trained according to the training method described herein. According to a further aspect, the determining is performed by comparing the measurement parameter with a healthy parameter set. Further provided is a computing device or system configured for performing the method as described herein. Further provided is a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform the method as described herein. Further provided is a computer readable medium having stored thereon a computer program having instructions which when executed by a computing device or system cause the computing device or system to perform the method as described herein. Further provided is a data stream which is representative of a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform the method as described herein. Further provided is a wearable headset comprising a plurality of EEG electrodes for determining a likelihood of a subject having epilepsy, configured to carry out the method as described herein. Further provided is a wearable headset for determining a likelihood of a subject having epilepsy, comprising a computing device and plurality of EEG electrodes operatively connected to the computing device, wherein the computing device is configured to carry out the method as described herein. Figure Legends FIG.1 shows an exemplary arrangement of EEG electrodes on a head of a subject, showing standard electrode locations. FIG. 2 shows a node map containing locations of different EEG electrodes (nodes) and connection strength between the different nodes. FIG. 3 shows a node map containing locations of different EEG electrodes (nodes) and connectivity integration between the different nodes. FIG.4-5 Graphs #1 to #16 show experimental data that are calibration curves for multiple probabilistic machine learning models (#1 to #15), each trained on only one but different parameter. The Brier Score (BrS) of each model is shown in the bottom right corner of the curve. FIG.6-7 Graphs #1 to #16 show experimental data that are calibration curves for multiple probabilistic machine learning models (#1 to #15), each trained with an increasing number of parameters. FIG.8-9 Graphs #1 to #16 show experimental data that are calibration curves for multiple probabilistic machine learning models (#1 to #15), each trained with an increasing number of parameters, but wherein the regional specificity of the parameter has been removed. FIG.10 shows a calibration curve for a finally trained model. FIG.11 shows a confusion matrix and evaluation data for a deterministic machine learning model. FIG. 12 show box plots of predicted probably of epilepsy-, cardiac- and other-diagnosed subjects using the finally trained model of Example 1. FIG.13 shows a calibration curves for data used in Examples 1 and 3. Detailed description of the invention Before the present system and method of the invention are described, it is to be understood that this invention is not limited to particular systems and methods or combinations described, since such systems and methods and combinations may, of course, vary. It is also to be understood that the terminology used herein is not intended to be limiting, since the scope of the present invention will be limited only by the appended claims. As used herein, the singular forms "a", "an", and "the" include both singular and plural referents unless the context clearly dictates otherwise. The terms "comprising", "comprises" and "comprised of" as used herein are synonymous with "including", "includes" or "containing", "contains", and are inclusive or open-ended and do not exclude additional, non-recited members, elements or method steps. It will be appreciated that the terms "comprising", "comprises" and "comprised of" as used herein comprise the terms "consisting of", "consists" and "consists of". The recitation of numerical ranges by endpoints includes all numbers and fractions subsumed within the respective ranges, as well as the recited endpoints. The term "about" or “approximately” as used herein when referring to a measurable value such as a parameter, an amount, a temporal duration, and the like, is meant to encompass variations of +/-10% or less, preferably +/-5% or less, more preferably +/-1% or less, and still more preferably +/-0.1% or less of and from the specified value, insofar such variations are appropriate to perform in the disclosed invention. It is to be understood that the value to which the modifier "about" or “approximately” refers is itself also specifically, and preferably, disclosed. Whereas the terms “one or more” or “at least one”, such as one or more or at least one member(s) of a group of members, is clear per se, by means of further exemplification, the term encompasses inter alia a reference to any one of said members, or to any two or more of said members, such as, e.g., any ≥3, ≥4, ≥5, ≥6 or ≥7 etc. of said members, and up to all said members. All references cited in the present specification are hereby incorporated by reference in their entirety. In particular, the teachings of all references herein specifically referred to are incorporated by reference. Unless otherwise defined, all terms used in disclosing the invention, including technical and scientific terms, have the meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. By means of further guidance, term definitions are included to better appreciate the teaching of the present invention. In the following passages, different aspects of the invention are defined in more detail. Each aspect so defined may be combined with any other aspect or aspects unless clearly indicated to the contrary. In particular, any feature indicated as being preferred or advantageous may be combined with any other feature or features indicated as being preferred or advantageous. Reference throughout this specification to “one embodiment” or “an embodiment” means that a particular feature, structure or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, appearances of the phrases “in one embodiment” or “in an embodiment” in various places throughout this specification are not necessarily all referring to the same embodiment, but may. Furthermore, the particular features, structures or characteristics may be combined in any suitable manner, as would be apparent to a person skilled in the art from this disclosure, in one or more embodiments. Furthermore, while some embodiments described herein include some but not other features included in other embodiments, combinations of features of different embodiments are meant to be within the scope of the invention, and form different embodiments, as would be understood by those in the art. For example, in the appended claims, any of the claimed embodiments can be used in any combination. In the present description of the invention, reference is made to the accompanying drawings that form a part hereof, and in which are shown by way of illustration only of specific embodiments in which the invention may be practiced. Parenthesized or emboldened reference numerals affixed to respective elements merely exemplify the elements by way of example, with which it is not intended to limit the respective elements. Unless otherwise indicated, all figures and drawings in this document are not to scale and are chosen for the purpose of illustrating different embodiments of the invention. In particular the dimensions of the various components are depicted in illustrative terms only, and no relationship between the dimensions of the various components should be inferred from the drawings, unless so indicated. It is to be understood that other embodiments may be utilised and structural or logical changes may be made without departing from the scope of the present invention. The following detailed description, therefore, is not to be taken in a limiting sense, and the scope of the present invention is defined by the appended claims. The present invention provides a (computer-implemented) method for determining a likelihood of a test subject having epilepsy comprising: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state, wherein the measurement parameter set comprises at least parameter A from Table 1, - determining from the measurement data the likelihood of the test subject having epilepsy. The test subject is a subject for whom a likelihood of epilepsy is sought, based on the measurement data comprising the measurement parameter set. The test subject is typically human. Preferably, the test subject has suffered a first attack. The measurement data is typically acquired within 1 to 4 weeks of the first attack. The healthy subject is one that has suffered a first attack, after which a healthy dataset comprising a healthy parameter set was acquired. The healthy dataset was typically acquired within 1 to 4 weeks of the first attack. Subsequent to the first attack, during a follow-up period it could be established that the healthy subject was not suffering from epilepsy. The follow-up period was typically at least 2 years. In a population of healthy subjects, a healthy reference parameter may be measured that is a statistical indicator (e.g. order parameter (1st quartile, median, or 3rd quartile) or average) of that parameter for the population of healthy subjects. The measurement parameter set of the test subject may be compared with a healthy reference dataset, and depending on the deviation from a corresponding parameter values, a likelihood of the test subject having epilepsy can be determined. The training subject is a subject from whom a training dataset has been obtained comprising a training parameter set and a training tag. The training subject is one that has suffered a first attack, after which the training dataset was acquired. The training dataset was typically acquired within 1 to 4 weeks of the first attack. Subsequent to the first attack, during a follow- up period it could be established whether the training subject was suffering from epilepsy (tagged epileptic training subject) or was not suffering from epilepsy (tagged healthy training subject). The follow-up period was typically at least 2 years. Where a probabilistic machine learning model, PMLM, is used to determine from the measurement data the likelihood of the subject having epilepsy, a trained PMLM has been trained using multiple training datasets, each training dataset from a different training subject. The first attack is a possible epileptic seizure, characterized by a transient neurological deficit. This neurological deficit can present itself in different ways, including but not limited to loss or alteration of consciousness, uncontrolled movements, and/or sensory hallucinations. The first attack is the initial event based on which the medical doctor considers the possible diagnosis of epilepsy. Epilepsy is a disease of the brain defined according the International League Against Epilepsy (ILAE) (Fischer, 2014) by any of the following conditions: - at least two unprovoked (or reflex) seizures occurring >24 h apart, - one unprovoked (or reflex) seizure and a probability of further seizures similar to the general recurrence risk (at least 60%) after two unprovoked seizures, occurring over the next 10 years, - diagnosis of an epilepsy syndrome Epilepsy is considered to be resolved for individuals who had an age-dependent epilepsy syndrome but are now past the applicable age or those who have remained seizure-free for the last 10 years, with no seizure medicines for the last 5 years. The EEG recording is preferably a conventional EEG recording obtained using a conventional EEG system in which a plurality of electrodes are disposed at different locations on the head of the (test or training) subject so as to record the electrical activity of the brain. Electrode locations and names may be specified according to the International 10–20 system. Each electrode is connected to one input of a differential amplifier (one amplifier per pair of electrodes); a common system reference electrode is connected to the other input of each differential amplifier. The amplifiers amplify the voltage between the active electrode and the reference, and the signal from a channel is digitized via an analog- to-digital converter. Each active electrode provides an EEG signal to a different channel. The plurality of electrodes may be incorporated into a wearable headpiece such as a cap. Exemplary EEG systems of the art include NicoleteOne EEG (Natus), eego (ANT Neuro), RMS Super Spec 32 Channel EEG Machine (RMS), NS-EEG-D1 (Neurostyle), Brain Quick (Micromed), Easy EEG (BrainCapture), Neurofax / Cerebair / airEEG (Nihon Kohden), Arc EEG (Cadwell). An exemplary arrangement of EEG electrodes on a head of a subject is shown in FIG.1. The labelling convention shown in FIG.1 is standard and known in the art. The EEG recording of the (test or healthy or training) subject is recorded during an EEG session. The EEG may last around 10-30 minutes of continuous recording. During this time there might be blocks where the subject is exposed to stimuli such as photic stimulation (i.e., stroboscopic light flashes) and/or hyperventilation to increase the odds of recording epileptic activity. During EEG recording of the (test or healthy or training) subject, the subject is in a resting state. By resting state, it is meant that the subject is under restful conditions such as sitting or lying, and the subject is not exposed to stimuli (i.e., there is no attempt to induce epileptic activity). The eyes may be open or closed. The EEG recording of the (test or training or healthy) subject, comprising multiple EEG signal channels, is preferably stored. The EEG signal from each channel is preferably processed using one or more steps in the following order: - filtering with a high-pass filter (cut-off frequency 1Hz); - filtering with a 50Hz or 60Hz notch filter; - detecting and removing eye movement and eye blink artifacts. They may be detected and removed automatically using independent component analysis (ICA) and matching the components to pre-selected templates. The ICA components were removed if the correlation between the component and one of the templates was larger than 0.95. - average referencing. In average referencing an average of the EEG signals is subtracted from each channel. After average referencing, the overall amplitude across all channels sums up essentially to zero at each time point. - dividing into epochs (3s to 5s long). - omitting a beginning portion (e.g. first minute) and end portion (e.g. final minute) from the EEG recording. This is because artifacts may be present in these portions. - omitting stimulation-containing epochs from the EEG recording. Stimulation (e.g. photic) may have an unwanted effect on the EEG signal, epochs containing a stimulation periods are removed. - omitting artifact-containing epochs from the EEG recording. Artifacts may be detected automatically, and artifact-containing epochs removed automatically using an Autoreject algorithm (e.g. Jas, 2017). One or more parameters is measured in a test subject and the measurement parameter set comprises the one or more parameters. One or more parameters may be measured in a population of healthy subjects, and the healthy parameter set (where used) comprises the one or more healthy reference parameters, each healthy reference parameter being a statistical indication of that parameter measured in the population. One or more parameters may be used to populate the training parameter set (where used). Each parameter is determined from one or more EEG signal channels. A parameter may be: - a spectral parameter, - a functional connectivity parameter. A parameter may be particular (limited) to a frequency band. In EEG measurements in the art, the frequency bands are typically delta (Δ, 1-4 Hz), theta (θ, 4-8 Hz), alpha (α, 8-12 Hz), beta1 (β1, 12-20 Hz), beta2 (β2, 20-30 Hz), gamma (γ, 30-100 Hz), and broadband (bd, 1-100 Hz). A frequency band may encompass a range of two or more of delta, theta, alpha, beta1, beta2, gamma. The range is typically continuous. For example, a frequency band delta-beta2 is in the range 1-30 Hz. A parameter may be limited to a frequency band. A parameter may or may not be particular to a region of the head of the (test or training) subject. The region corresponds to the location of the EEG electrode(s) on the subject (excluding reference). Typical regions, as known in the art, are front left (FL), front right (FR), back left (BL), back right (BR), all regions (AR), one or more of AR (XR). Each region (FL, FR, BL, BR, XR) contains at least one EEG electrode, preferably 2 to 10 EEG electrodes. All region (AR) contains the total of all electrodes (excluding reference). As an example, with an arrangement of electrodes as depicted in FIG.1, FL contains Fp1, F7, F3, Fz, T3, C3, Cz; FR contains Fp2, F8, F4, Fz, T4, C4, Cz; BL contains T3, C3, Cz, T5, P3, Pz, O1; and BR contains T4, C4, Cz, T6, P4, Pz, O2. Within a session, an epoch parameter value is preferably determined from each epoch (of multiple epochs) of the EEG signal(s). An epoch is a continuous subperiod (e.g.3 to 5 seconds long) of the EEG signal(s). As mentioned elsewhere, some epochs may be discarded, for instance: - those epochs at the beginning (e.g. first minute) and/or end portion (e.g. final minute) from the EEG recording; - those epochs containing a stimulation event; - those epochs containing on or more artefacts. The multiple epoch parameter values (of the same parameter type) obtained from the multiple epochs are combined to arrive at the parameter value that is a single scalar value for the session. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all epoch parameter values of the session. A spectral parameter may be one of the following (1.1 to 1.6): (1.1) Relative band power (RBP). The parameter relative band power is a measure of signal power (signal area, power spectral density (PSD) curve) in a frequency band and particular (limited) to a region. The relative band power is determined by measuring signal power (signal area) in the frequency band for each EEG electrode in the region, and combining the measured signal powers into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured signal powers of the region. Within a session, relative band power is preferably determined from an epoch relative band power of each epoch of the EEG signal of the EEG electrode(s) of the region. The multiple values of epoch relative band powers across the session are combined into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all an epoch relative band powers in the session. The RBP parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of RBP parameters at multiple different time points, and therefore does not include a time- based component. Relative band power and methods for measurement of relative band power are known in the art, for instance, from Laman et al 2005, Wang et al 2016 and Krishnan et al 2020. A spectral parameter (SP) that is relative band power (RBP) may be denoted SPRBP. It may further denote the frequency band: SPRBP, band. It may further denote the region: SPRBP, band, region. (1.2) Peak frequency (PF). The parameter peak frequency is a frequency of the peak (maximal) signal in a frequency band and particular (limited) to a region. It is preferably determined from the power spectral density (PSD) curve. The peak frequency is determined by measuring peak frequency in the frequency band for each EEG electrode in the region, and combining the measured peak frequencies into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured signal powers. Within a session, peak frequency is preferably determined from an epoch peak frequency of each epoch of the EEG signal of the EEG electrode(s) of the region. The multiple values of epoch peak frequencies across the session are combined into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all an epoch peak frequencies in the session. The PF parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of PF parameters at multiple different time points, and therefore does not include a time-based component. Peak frequency and methods for measurement of peak frequency are known in the art, for instance, from Sörnmo et al 2005.A spectral parameter (SP) that is relative peak frequency (PF) may be denoted SPPF. It may further denote the frequency band: SPPF, fb. It may further denote the region: SPPF, fb, region. wer is a power (area) of the peak (maximal) signal in a frequency band and particular (limited) to a region. It is preferably determined from the power spectral density (PSD) curve. The peak power is determined by measuring peak power in the frequency band for each EEG electrode in the region, and combining the measured peak powers into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured signal powers. If a peak is detected at the edge of a frequency band - possibly due to a linear trend in the curve or the absence of a true peak - the value is preferably not included. The PSD curves may be determined using a multitaper estimation method and corrected for a 1/f distribution by multiplying each point by the corresponding frequency (e.g. Bronez, 1992). Within a session, peak power is preferably determined from an epoch peak power of each epoch of the EEG signal of the EEG electrode(s) of the region. The multiple values of epoch peak powers across the session are combined into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all an epoch peak powers in the session. The PP parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of PP parameters at multiple different time points, and therefore does not include a time-based component.
Figure imgf000016_0001
. The parameter peak bandwidth is a width of the peak signal in a frequency band and particular (limited) to a region. It is preferably determined from the power spectral density (PSD) curve. The peak bandwidth is determined by measuring peak bandwidth in the frequency band for each EEG electrode in the region, and combining the measured peak bandwidths into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured signal bandwidths. If a peak is detected at the edge of a frequency band - possibly due to a linear trend in the curve or the absence of a true peak - the value is preferably not included. The PSD curves may be determined using a multitaper estimation method and corrected for a 1/f distribution based on the 1/f exponent (see below in parameter 1.5). (e.g. Bronez, 1992). The BW parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of BW parameters at multiple different time points, and therefore does not include a time-based component. Peak Bandwidth (BW) and methods for measurement of Peak Bandwidth (BW) are known in the art, for instance, from Bronez, 1992. A spectral parameter (SP) that is Peak Bandwidth (BW) may be denoted BW. It may further denote the frequency band: BW, fb. It may further denote the region: BW, fb, region. (1.5) 1/f exponent (EXP). The parameter 1/f exponent is a slope (exponent) of the decay of the power spectral density (PSD) curve particular (limited) to a region. The 1/f exponent is determined by measuring 1/f exponent for each EEG electrode in the region, and combining the measured 1/f exponents into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all the measured offsets. The PSD curves may be determined using a multitaper estimation method. The EXP parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of EXP parameters at multiple different time points, and therefore does not include a time- based component. Peak frequency and methods for measurement of peak frequency are known in the art, for instance, from Medel et al 2023. A spectral parameter (SP) that is 1/f exponent (EXP) may be denoted SPEXP. It may further denote the frequency band: SPEXP, fb. It may further denote the region: SPEXP, fb, region. (1.6) Offset (OFF). The parameter offset is an amplitude of the constant offset of the power spectral density (PSC) curve particular to a region. The offset is determined by measuring the offset for each EEG electrode in the region, and combining the measured offsets into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all the measured offsets. The PSD curves may be determined using a multitaper estimation method. Offset (OFF) and methods for measurement of offset (OFF) are known in the art. The OFF parameter is a stationary parameter, namely, it is determined from the power spectral density (PSD) curve at one moment in time, or is preferably a statistical combination of OFF parameters at multiple different time points, and therefore does not include a time- based component. A spectral parameter (SP) that is offset (OFF) may be denoted OFF. It may further denote the frequency band: OFF, fb. It may further denote the region: OFF, fb, region. It is appreciated that the one or more multiple different spectral parameters may be measured in the subject, optionally each measured in a different frequency band and/or in a different region. Functional connectivity is determined from the EEG signals to include information on how the brain regions communicate with each other. Functional connectivity is determined from EEG signals from all pairs of electrodes within a region, and is a scalar value indicating direct or indirect strength of connection between different electrode pairs. Many connectivity metrics exist focussing on different possible aspects of functional relationships between the considered signals (e.g. Sakkalis, 2011). A functional connectivity parameter may be a connectivity strength (CS) functional connectivity parameter. A CS functional connectivity parameter is limited to a region which may be one of: FL, FR, BL, BR, AR. The CS functional parameter is determined by measuring a strength of connection (connectivity strength, CS) between all electrode pairs in the region, and combining the measured connectivity strengths into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) of all the measured connectivity strengths for the region. The connectivity strength between pairs of electrodes may be determined by several different methods. Examples of methods known in the art for determining connectivity strength include: Imaginary coherence (IMCOH), Phase locking value (PLV), Partial directed coherence (PDC), and Amplitude envelope correlation (AEC). These have been described in at least Bastos et al, 2016, and Briels et al 2020. Connectivity strength (CS) functional connectivity parameter and methods for its measurement are generally known in the art, for instance, in Sporns 2022. An example of how connectivity strength (CS) functional connectivity parameter is determined is provided in FIG.2. In FIG.2, a plurality of different EEG electrodes (nodes) are represented (1 to 5), and connectivity strengths (CS) between them labelled. In the figure CS (1,2) = 0.8, CS (1,3) = 0.1, CS (1,4) = 0.5, CS (1,5) = 0.1, CS (2,3) = 0.7, CS (2,4) = 0.4, CS (2,5) = 0.4, CS (3,4) = 0.1, CS (3,5) = 0.2, CS (4,5) = 0.8. A value of the connectivity strength (CS) functional connectivity parameter may be determined as an average of the individual connectivity strengths (CS). Connectivity strength (CS) functional connectivity parameter value = (0.8 + 0.1 + 0.5 + 0.1 + 0.7 + 0.4 + 0.4 + 0.1 + 0.2 + 0.8)/10 = 0.41. Within a session, CS functional connectivity parameter is preferably determined from an epoch CS functional connectivity parameter of each epoch of the EEG signal(s) of the EEG electrode(s) of the region. The multiple values of epoch CS functional connectivity parameters across the session are combined into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all epoch CS functional connectivity parameters in the session. It is an aspect that not all epochs are used, but a selection is made where epochs (e.g.100) with the largest relative band power are used. For example, when calculating CS functional connectivity parameter in the alpha band, only the epochs with the largest relative alpha power may be selected to determine the CS functional connectivity parameter. This selection scheme allows a better extraction of the connectivity patterns by looking at the epochs that best represent the activity in that frequency band. A function connectivity parameter (FCP) that is connectivity strength (CS) may be denoted FCPCS. It may further denote the frequency band: FCPCS, band. It may further denote the region: FCPCS, band, region. It may further denote the CS measurement method: FCPCS, band, region, method. A functional connectivity parameter may be a global efficiency (GE) functional connectivity parameter. A GE functional connectivity parameter is limited to a region which may be one of: FL, FR, BL, BR, AR. The GE functional parameter is determined by measuring an integration (togetherness) of connections (connectivity integration, CI) between all EEG electrode pairs in the region, and combining the measured connectivity integrations (CI) into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g. 1st and 3rd quartile)) this of all the measured connectivity integrations. The connectivity integration between a particular pair of EEG electrodes (e.g. a and b) is determined from the connection strengths between all EEG electrodes pairs (e.g. a to g) in the region, and determining the shortest path length (connection strength-1) between the particular pair of EEG electrodes which may be a direct path (e.g. a-b) or an indirect path via one or more other EEG electrodes (e.g. a-c-b) based on the a summation of CI values along the path which gives the lowest value. Global efficiency is know in the art Examples of methods known in the art for determining connectivity strength and hence connectivity integration (which is an inverse of connectivity strength) include: Imaginary coherence (IMCOH), Phase locking value (PLV), Partial directed coherence (PDC), Amplitude envelope correlation (AEC). These have been described in at least Bastos et al, 2016, and Briels et al 2020. Global efficiency (GE) functional connectivity parameter and methods for its measurement are generally known in the art, for instance, in Sporns 2022 and https://en.wikipedia.org/wiki/ Efficiency_(network_science). An example of how global efficiency (GE) functional connectivity parameter is determined is provided in FIG.3. In FIG.3, a plurality of different EEG electrodes (nodes) are represented (1 to 5), and connectivity integration (CI) (CS-1 from FIG.2) between them labelled. In the figure CI (1,2) = 1.25, CI (1,3) = 10, CI (1,4) = 2, CI (1,5) = 10, CI (2,3) = 1.43, CI (2,4) = 2.5, CI (2,5) = 2.5, CI (3,4) = 10, CI (3,5) = 5, CI (4,5) = 1.25. The shortest path length between EEG electrode pairs is determined according to the path which yield the smallest summation of CI values along that path. For instance, the shortest path length between (1)-(3) is not direct (sum CI=10) but is (1)-(2)-(3) (sum CI=0.8+0.7 = 1.5). In the FIG.3 shortest path length (SPL) are SPL (1,2) = 1.25, SPL (1,3) = 2.68, SPL (1,4) = 2, SPL (1,5) = 2.25, SPL (2,3) = 1.43, SPL (3,4) = 2.5, SPL (4,5) = 2.5, SPL (3,4) = 3.93, SPL (3,5) = 3.93, SPL (4,5) = 1.25. A value of the global efficiency (GE) functional connectivity parameter may be determined as an average of the individual (SPLs-1). Global efficiency (GE) functional connectivity parameter value = (1/1.25 + 1/2.68 + 1/2 + 1/2.25 + 1/1.43 + 1/2.5 + 1/2.5 + 1/3.93 + 1/3.93 + 1/1.25)/10)/10 = 0.49. Within a session, GE functional connectivity parameter is preferably determined from an epoch GE functional connectivity parameter of each epoch of the EEG signal(s) of the EEG electrode(s) of the region. The multiple values of epoch GE functional connectivity parameters across the session are combined into a single scalar value. The combining may be performed by a statistical method e.g. taking an average, taking an order statistic (e.g. median, or another quantile (e.g.1st and 3rd quartile)) of all epoch GE functional connectivity parameters in the session. It is an aspect that not all epochs are used, but a selection is made where 100 epochs with the largest relative band power are used. For example, when calculating GE functional connectivity parameter in the alpha band, only the epochs with the largest relative alpha power may be selected to determine the GE functional connectivity parameter. This selection scheme allows a better extraction of the connectivity patterns by looking at the epochs that best represent the activity in that frequency band. A function connectivity parameter (FCP) that is global efficiency (GE) may be denoted FCPGE. It may further denote the frequency band: FCPGE, band. It may further denote the region: FCPGE, band, region. It may further denote the CI measurement method: FCPGE, band, region, method. Particular examples of ways to calculate connectivity strength (CS) and connectivity integration (CI, which is CS-1) are provided below. Pairs of EEG signals (x and y) are considered, one from EEG signal channel x, and one from EEG signal channel y. Each EEG channel contains an EEG signal from a different EEG electrode, and the electrodes are located in different positions on the head. (2.1) Imaginary coherence (IMCOH): Connectivity strength of an EEG electrode pair may be determined from imaginary coherence (IMCOH) of the EEG electrode pairs. The imaginary coherence (IMCOH) is computed from a normalized cross-spectrum between two signals (e.g. Nolte, 2004) according to Eq.1:
Figure imgf000021_0001
where Sxy(f) is the cross-spectral density between signals x and y, and Sxx(f) and Syy(f) are auto-spectral densities of signals x and y respectively. By taking the imaginary part of the coherence, instantaneous interactions i.e., spurious connections due to volume conduction are removed. The connectivity strength (IMCOH) for an x,y pair is determined from the value Cxy(f), and connectivity integration for an x,y pair is determined from the value (Cxy(f))-1. (2.2) Phase
Figure imgf000022_0001
value (PLV) Connectivity strength of an EEG electrode pair may be determined from a phase locking value (PLV) of the EEG electrode pairs. The parameter phase locking value (PLV) is the most common phase synchrony measure and aims to detect signals with a phase difference that is stable over time. The PLV is the absolute value of the averaged phase difference, expressed as a complex unit-length vector (e.g. Lachaux, 1999) according to Eq.2:
Figure imgf000022_0002
Where PLVxy is a phase locking value between signals x and y, ^x(t) and ^y(t) are the phases of signals x and y respectively at time-point t, T is the total length of the signals, and i is the imaginary unit. The connectivity strength (PLV) for an x,y pair is determined from the value PLVxy, and connectivity integration for an x,y pair is determined from the value (PLVxy)- 1. (2.3) Partial directed coherence (PDC) The parameter partial directed coherence (PDC) uses the Fourier transform of an autoregressive model to quantify the fraction of the spectral power of signal x that contributes to the future of signal y. PDC detects only direct interactions between signals and is normalized with respect to the total outflow (e.g. Baccalá, 2001) according to Eq.3:
Figure imgf000022_0003
Where PDCxy(f) is the partial directed coherence between signals x and y, evaluated at frequency f. Axy(f) is the element at position (x,y) taken from the matrix A(f), which is the Fourier transform of the coefficient matrix from an autoregressive model fitted to signals x and y (e.g. Baccalá, 2001). The partial directed coherence (PDC) for an x,y pair is determined from the value PDCxy(f), and connectivity integration for an x,y pair is determined from the value (PDCxy(f))-1. The (measurement or training) parameter set preferably comprises at least one parameter (A to I) from Table 1. The (measurement or training) parameter set may consist of at least one parameter (A to I) selected from Table 1. Parameter Parameter Parameter Type Frequency Preferred Abbreviation Name band Region (optional) A RBP delta Spectral, relative Delta FR SPRBP,delta,(FR) band power B GE functional Functional Delta all regions FCPGE,delta,(AR) connectivity delta connectivity, global band efficiency C CS functional Functional Delta BR & FL FCPCS,delta,(BR,FL) connectivity delta connectivity, band connectivity strength D EXP Spectral, 1/f Broadband FL SPEXP,bd, (FL) exponent E GE functional Functional Broadband FR FCPGE,bd,(FR) connectivity connectivity, global broad band efficiency F PF broadband Spectral, peak Broadband BL SPPF, bd, (BL) frequency G RBP gamma Spectral, relative Gamma FR SPRBP,gamma,(FR) band power H PF theta Spectral, peak Theta FR SPPF,theta,(FR) frequency I RBP theta Spectral, relative Theta FR SPRBP,theta,(FR) band power Table 1: Parameters derived from EEG measurement of a test or training subject. Further descriptions of and/or methods for measurement of Parameter A are known in the art, for instance, from Laman et al 2005, Wang et al 2016 and Krishnan et al 2020. Further descriptions of and/or methods for measurement of Parameter B are known in the art, for instance, from Sporns 2022 and https://en.wikipedia.org/wiki/ Efficiency_(network_science), Bastos et al, 2016, and Briels et al 2020. Further descriptions of and/or methods for measurement of Parameter C are known in the art, for instance, from Sporns 2022, Bastos et al, 2016, and Briels et al 2020. Further descriptions of and/or methods for measurement of Parameter D are known in the art, for instance, from Medel et al 2023. Further descriptions of and/or methods for measurement of Parameter E are known in the art, for instance, from Sporns 2022, https://en.wikipedia.org/wiki/ Efficiency_(network_science), Bastos et al, 2016, and Briels et al 2020. Further descriptions of and/or methods for measurement of Parameter F are known in the art, for instance, from Sörnmo et al 2005. Further descriptions of and/or methods for measurement of Parameter G are known in the art, for instance, from Laman et al 2005, Wang et al 2016 and Krishnan et al 2020. Further descriptions of and/or methods for measurement of Parameter H are known in the art, for instance, from Sörnmo et al 2005. Further descriptions of and/or methods for measurement of Parameter I are known in the art, for instance, from Laman et al 2005, Wang et al 2016 and Krishnan et al 2020. The measurement data of the test subject comprises a measurement parameter set. The measurement parameter set preferably comprises at least parameter A in Table 1. The measurement parameter set preferably comprises at least parameters A and B in Table 1. The measurement parameter set preferably comprises at least parameters A to C in Table 1. The measurement parameter set preferably comprises at least parameters A to D in Table 1. The measurement parameter set preferably comprises at least parameters A to E in Table 1. The measurement parameter set preferably comprises at least parameters A to F in Table 1. The measurement parameter set preferably comprises at least parameters A to G in Table 1. The measurement parameter set preferably comprises at least parameters A to H in Table 1. The measurement parameter set preferably comprises at least parameters A to I in Table 1. The likelihood of epilepsy in the test subject may be determined by comparing the measured parameter(s) with a healthy reference parameter(s). The likelihood of epilepsy in the test subject may be determined by comparing the measurement parameter set of the test subject with the healthy parameter set. In particular, the likelihood of epilepsy in the test subject may be determined by comparing a value of a parameter (e.g. one of A to I) for the test subject with the same (healthy reference) parameter of the healthy parameter set. The comparison may show no deviation or a deviation (an increase, or a decrease) compared to a healthy subject. Where the deviation is an increase or a decrease, depending on the parameter, the test subject has an increased likelihood of epilepsy. The greater the deviation from the healthy parameter set, the higher the likelihood of epilepsy. Table 2 sets out the deviation (increase or decrease) in value for each parameter for a test subject compared with a healthy subject that is linked to an increased likelihood of epilepsy in the test subject. Parameter Parameter Name Abbreviation Deviation A RBP delta SPRBP,delta,(FR) Increase B GE functional connectivity delta band FCPGE,delta,(AR) Increase C CS functional connectivity delta band FCPCS,delta,(BR,FL) Increase D EXP broadband SPEXP, bd, (FL) Increase E GE functional connectivity broad band FCPGE,bd,(FR) Increase F PF broadband SPPF, bd, (BL) Decrease G RBP gamma SPRBP,gamma,(FR) Decrease H PF theta SPPF,theta,(FR) Decrease I RBP theta SPRBP,theta,(FR) Increase Table 2: List of parameters and deviations. The deviation is a deviation compared with healthy reference. A deviation indicating an increase means that when there is an increase in the parameter value for the test subject compared with the same healthy reference parameter, the test subject has an increased likelihood of epilepsy. A deviation indicating a decrease means that when there is a decrease in the parameter value for the test subject compared with the same healthy reference parameter, the test subject has an increased likelihood of epilepsy. One or more of the parameters are able to predict epilepsy in a test subject that has suffered a first attack. Where multiple test parameters have been measured in the test subject, the likelihood may be determined based on the deviation (absolute difference between the test parameter values and the healthy reference values. A weighted combination of these deviations (differences) results in an epilepsy risk index (ERI). An example of an equation (Eq.4) thereof is provided below. ^^^ = ∑^ ^^(^^ − ^^) [Eq.4] where Hi and Ti are the healthy and test values respectively of parameter i, and ai are the weight coefficients of parameter i. The weight coefficients may be determined based on the training dataset. Possible values for the weight coefficients are listed in Table 4. The healthy parameter set comprises one or more healthy reference parameters (e.g. one of healthy reference parameters A to I). A healthy reference parameter is a statistical indicator of that parameter in a healthy population. More in particular, a healthy reference parameter (e.g. healthy reference parameter A) is determined by measuring the same parameter (e.g. parameter A) in multiple subjects of a population of healthy subjects, and obtaining a statistical indicator (e.g. order parameter (1st quartile, median, or 3rd quartile) or average) of that measured parameter for the population of healthy subjects. Typically, the number of subjects in the population of healthy subjects is 200 or more. The healthy parameter set comprises one or more healthy reference parameters each determined from a statistical indicator of a parameter in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A, determined from a statistical indicator of parameter A in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A and B, each determined from a statistical indicator of parameter A and B respectively in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A to C, each determined from a statistical indicator of parameter A to C respectively in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A to D, each determined from a statistical indicator of parameter A to D respectively in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A to E, each determined from a statistical indicator of parameter A to E respectively in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A to F, each determined from a statistical indicator of parameter A to F respectively in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A to G, each determined from a statistical indicator of parameter A to G respectively in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A to H, each determined from a statistical indicator of parameter A to H respectively in Table 1 in a healthy population. The healthy parameter set preferably comprises at least healthy reference parameter A to I, each determined from a statistical indicator of parameter A to I respectively in Table 1 in a healthy population. As a guidance, exemplary healthy reference parameter values are provided in Table 3. These values may be calculated by the skilled person using the methods described herein applied to a group of healthy subjects. It is understood that the reference values are given for general guidance, and may be refined depending on the population group (e.g. child, adult). Parameter Parameter Name Abbreviation Healthy reference parameter value A RBP delta SPRBP,delta,(AR) 5.10 * 10-2 ± 4.87 * 10-2 B GE functional FCPGE,delta,(AR) IMCOH: 2.08 * 10-2 ± 9.47 * 10-3 connectivity delta PLV: 3.01 * 10-1 ± 3.80 * 10-2 band PDC: 1.37 * 10-1 ± 1.43 * 10-2 AEC: 2.02 * 10-1 ± 2.66 * 10-2 C CS functional FCPCS,delta,(AR) IMCOH: 3.04 * 10-2 ± 1.40 * 10-2 connectivity delta PLV: 4.67 * 10-1 ± 6.79 * 10-2 band PDC: 1.86 * 10-1 ± 2.17 * 10-2 AEC: 3.38 * 10-1 ± 4.45 * 10-2 D EXP SPEXP,bd, (FL) 1.10 ± 0.43 E GE functional FCPGE,bd,(AR) IMCOH: 1.55 * 10-2 ± 6.52 * 10-3 connectivity broad PLV: 3.03 * 10-1 ± 2.69 * 10-2 band PDC: 1.31 * 10-1 ± 1.43 * 10-2 AEC: 9.95 * 10-2 ± 2.28 * 10-2 F PF broadband SPPF, bd, (AR) 9.65 ± 4.11 G RBP gamma SPRBP,gamma,(AR) 3.89 * 10-1 ± 1.37 * 10-1 H PF theta SPPF,theta,(AR) 7.09 ± 8.64 * 10-1 I RBP theta SPRBP,theta,(AR) 8.11 * 10-2 ± 6.72 * 10-2 Table 3: List of parameters and exemplary healthy reference values when the parameter is measured for all regions (AR). A probabilistic machine learning model, PMLM, may optionally be used to determine the likelihood of the test subject having epilepsy. Multiple training datasets are used to train the PMLM, wherein each training dataset of a training subject comprises a training parameter set and a training tag. Provided herein is a training method for training a probabilistic machine learning model, PMLM, for determining a likelihood of a subject having epilepsy comprising, the training method comprising: - receiving a plurality of training data sets, wherein: -each training data set has been acquired from a training subject and contains: - a training parameter set that has been measured in the training subject in a resting state, and - a training tag (e.g.1 or 0) that is indication of presence (1) or absence (0) of epilepsy in the training subject; - training the PMLM using the plurality of training data sets, wherein: - an input to the PMLM is the training parameter set of the data training set, and an output of PMLM is a pending tag (probabilistic); - the pending tag (probabilistic) is compared with the training tag (e.g.1 or 0) of the training data set, and - the PMLM is adjusted so that the pending tag (probabilistic) approaches the training tag (e.g.1 or 0) of the training set. In more detail, a nested cross-validation scheme may be used to train and adjust the PMLM. The PMLM is adjusted based on a performance metric. One example of such a performance metric is a Brier Score (BrS). A training tag is translated into a 1 or 0 (1 - epilepsy present or 0 - epilepsy absent in the training subject). The BrS is used to compare the pending tag - the probabilistic output of the PMLM - to the training tag, using for instance, Eq.5:
Figure imgf000028_0001
wherein pi the pending tag i.e. predicted PMLM output probability for training subject i. It may have a value between 0 and 1; oi the training tag for training subject i. It may have a value of 1 (epilepsy present) or 0 (epilepsy absent) in the training subject; N is the number of training subjects. Using the nested cross-validation scheme, the PMLM is penalised more if the predicted probability (pi) is further away from the training tag value (oi). As the predictions get closer to the training tags the BrS becomes smaller, so training is based on minimising the Brier Score. Any machine learning models that can give a graded output score and trained to predict probabilities (e.g., using the BrS) may be used as a PMLM. Examples of suitable candidates for PMLM include Logistic Regression, Support Vector Machine, and Naïve Bayes. Examples of performance metrics include the Brier Score mentioned above, LogLoss (also called cross-entropy loss), and the Brier Skill Score. The Brier Skill Score (BrSS) is an adaptation of the Brier Score, where the score is normalized compared to a random predictor (BrSS = 1-BrS/BrSref, with BrSref being the Brier Score of a random predictor). Other examples of PMLMs and nested cross-validation schemes have been described in Ferro et al., Comparing Probabilistic Forecasting Systems with the Brier Score, Weather and Forecasting; T. Proix et al., Forecasting seizure risk in adults with focal epilepsy: a development and validation study; B. Wissel et al., Early identification of epilepsy surgery candidates: A multicenter, machine learning study. The trained PMLM may be calibrated to further improve the performance. The goal of probability calibration is to fit a simple model to the predicted probabilities and cancel this model out to bring the probabilities closer to the diagonal line on the calibration curve. Calibration may be achieved by any method of the art, such as Platt’s scaling or isotonic regression. In Platt’s scaling, a sigmoid function is used as regression model. In Isotonic regression a free-form line is fitted to the calibration curve that is non-decreasing (i.e., the line is always increasing). Each training data set has been acquired from a training subject. A training dataset comprises - a training parameter set that has been measured in the training subject, and - a training tag (e.g.1 or 0) that is indication of presence (1) or absence (0) of epilepsy in the training subject. The training parameter set comprises one or more of the parameters of Table 1. The training parameter set preferably comprises at least parameter A in Table 1. The training parameter set preferably comprises at least parameters A and B in Table 1. The training parameter set preferably comprises at least parameters A to C in Table 1. The training parameter set preferably comprises at least parameters A to D in Table 1. The training parameter set preferably comprises at least parameters A to E in Table 1. The training parameter set preferably comprises at least parameters A to F in Table 1. The training parameter set preferably comprises at least parameters A to G in Table 1. The training parameter set preferably comprises at least parameters A to H in Table 1. The training parameter set preferably comprises at least parameters A to I in Table 1. The parameters present in the training parameter set are preferably the same as the parameters present in the measurement parameter set. For instance, where the trained PMLM has been trained using a training parameter set comprising parameters A to D, the measurement parameter set using as test input to the trained PMLM also comprises parameters A to D measured from the test subject. Typically, the number of training subjects is 200 or more. The determining from the measurement data the likelihood of the test subject having epilepsy may comprise using the trained PMLM. A method for determining a likelihood of a test subject having epilepsy comprises: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state, wherein the measurement parameter set comprises at least Parameter A of Table 1; - using the measurement data as an input for a trained PMLM, the trained PMLM trained according to the training method described herein; - outputting from the trained PMLM, a likelihood of the test subject having epilepsy. Typically deterministic machine learning is used to predict whether a subject has dysfunction or not. However, the inventors have found that deterministic machine learning applied to the problem of epilepsy prediction produces a very low level of predictive accuracy. In a retrospective study, the inventors trained and tested a deterministic machine learning model (Example 2) and a probabilistic machine learning model (PMLM) (Example 1) to classify epilepsy subjects vs non-epilepsy subjects. The deterministic machine learning model achieved a sensitivity of 36% and specificity of 77%, which is in line with the current diagnostic yield of EEG in clinical practice based on visual analysis of the EEG signals, with sensitivity ranging between 25-56% and specificity between 78-98% (Smith et al., 2005). It did not, however, provide an improvement over the current accuracies, and its performance therefore remains unsatisfactory. The PMLM on the other hand performed very well, evidenced by a calibration curve close to the diagonal and a low Brier Score (BrS = 0.23). The PMLM has been developed in the field of meteorology to predict non-deterministic events such as weather forecasting. Because the method quantifies the uncertainty of predictions, it lends itself well to event forecasting - quantifying how likely it is that an event will happen within a specific timeframe. Assigning a subject to a diagnostic group, however, is not considered a routine application of a PMLM because a diagnosis is neither considered to be non-deterministic (either a subject belongs to a disease group or does not) nor a prediction about a future event. The inventors have found that PMLM can be applied to the problem of epilepsy likelihood, and it provides an accurate prediction, even though PMLM is not being used to predict an event. In addition, the inventors have found highly predictive parameters (parameter A to H) in Table 1) that can be measured in the resting state of the subject, namely, there is no need to try to induce epileptic activity in the subject. Diagnostic labels for subjects have not previously been considered to be non-deterministic, because the subject either belongs to one group or they do not and - given enough evidence - the subject can be assigned to the correct group without uncertainty. Switching to a graded output (likelihood) that returns a probability index of a diagnostic label is not routine, yet the inventors have found that it produces more predictive accuracy compared with the deterministic approach. The graded output provided by the present method quantifies the uncertainty of the prediction, which can be used to decide on the type of treatment e.g. less invasive (drug) or more invasive (surgical), as a function of the likelihood of the epilepsy. The methods mentioned herein, wherein: - the measurement parameter set comprises at least parameter A in Table 1, or - the measurement parameter set comprises at least parameters A and B in Table 1 - the measurement parameter set comprises at least parameters A to C in Table 1 - the measurement parameter set comprises at least parameters A to D in Table 1 - the measurement parameter set comprises at least parameters A to E in Table 1 - the measurement parameter set comprises at least parameters A to F in Table 1 - the measurement parameter set comprises at least parameters A to G in Table 1 - the measurement parameter set comprises at least parameters A to H in Table 1 - the measurement parameter set comprises at least parameters A to I in Table 1 may be applied to one or more methods as described in the aspects below. According to one aspect, a method of determining a likelihood of a test subject having epilepsy as described elsewhere herein, which method further comprises: - comparing a value(s) of measurement parameter set with value(s) of healthy parameter set, - finding a deviation or no deviation of the value(s) of measurement parameter set with value(s) of healthy parameter set; and - attributing said finding of deviation or no deviation to a likelihood of the test subject having epilepsy. According to one aspect, a method for the selection of a prophylactic or therapeutic treatment for epilepsy is provided, which method comprises: - exposing one or more subjects to the prophylactic or therapeutic treatment, - comparing a value(s) of measurement parameter set with value(s) of healthy parameter set in the test subject prior to and after exposing to the subject to the prophylactic or therapeutic treatment. The selection of a prophylactic or therapeutic treatment is determined based on a deviation or no deviation of the value(s) of measurement parameter set after before and after the exposing. According to one aspect, a method of assessing an efficacy of a therapeutic treatment is provided, which method further comprises: - exposing the one or more test subjects to the prophylactic or therapeutic treatment, - comparing a value(s) of measurement parameter set of the one or more test subjects prior to and after exposing to the subject to the prophylactic or therapeutic treatment. The efficacy of the therapeutic treatment is determined based on a deviation or no deviation of the value(s) prior to and after the exposing. Further provided is a system for determining a likelihood of a subject having epilepsy, configured to carry out the method as described herein. The system may include computing device (described below) and a plurality of EEG electrodes. Further provided is a wearable headset comprising a plurality of EEG electrodes for determining a likelihood of a subject having epilepsy, configured to carry out the method as described herein. The wearable headset may comprise a computing device (described below) or connection (wireless or via one or more cables) to a computing device. The plurality of EEG electrodes is operatively connected to the computing device. By operatively connected, it is meant that the computing device is able to receive signals from the plurality of EEG electrodes in order to carry out a method as described herein. The method is typically automatic, meaning there is no intervention or monitoring by the practitioner. The method described herein is a computer implemented method. The method is in vitro, more in particular ex vivo. The method is an offline method. The method may be performed on stored measurement data. Further provided is a computing device or system configured for performing a method as described herein. Further provided is a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform a method as described herein. Further provided is a computer readable medium having stored thereon a computer program (product) having instructions which when executed by a computing device or system cause the computing device or system to perform (each of the steps of) the method as described herein. Further provided is a data stream which is representative of a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform (each of the steps of) the method as described herein. The method may be performed using a standard computer system such as an Intel Architecture IA-32 based computer system 2, and implemented as programming instructions of one or more software modules stored on non-volatile (e.g. hard disk or solid-state drive) storage associated with the corresponding computer system. However, it will be apparent that at least some of the steps of any of the described processes could alternatively be implemented, either in part or in its entirety, as one or more dedicated hardware components, such as gate configuration data for one or more field programmable gate arrays (FPGAs), or as application-specific integrated circuits (ASICs), for example. The method or system may produce an output that is: - displayed on a screen, or - saved to a file. The present method of system may be regarded as a method for measuring or detecting the one or more parameters. According to one aspect, a system or kit is provided for use in a method described herein. According to one aspect, a use of a system is provided for use in a method described herein, wherein said device comprises at least one EEG electrode. According to one aspect, a use of a system is provided for use in a method described herein, wherein said device comprises an EEG system. The system or kit or system or kit for use is a method may further comprise one or more of: - a computer system for executing the method described herein; - an EEG system. Example 1 EEG recordings were acquired from multiple subjects, each of which had suffered a first attack, and based on a follow-up study of at least 2 years duration could be classified as suffering from epilepsy or not. From the EEG recordings per subject multiple parameters were derived that was a spectral parameter or a functional connectivity parameter. In total 385 parameters were computed for each subject, the most important of which are shown in Table 4. The parameter values were standardized by taking a z-score. As known in the art, ^ = (^ − ^)/^, where z is the z-score of the parameter value x, and ^ and ^ are the mean and standard deviation of the parameter values across all subjects. Subsequently, an iterative feature selection scheme was applied. The feature selection was based on an ANOVA F-test where at each iteration a comparison was made between parameter values in epilepsy patients and non-epilepsy patients. The parameter with the lowest p-value was selected. In addition, parameters with a strong correlation with the newly selected feature (r > 0.80) were removed from the list of possible candidates for future iterations. Table 4 lists the top 15 selected parameters, with the associated p-value and effect size (d), and whether the feature is increased or decreased in epilepsy patients. The effect size is a more representative indication of the distance between the distributions of the two groups (i.e., epilepsy vs non-epilepsy) and is computed as the difference of the means, normalized by the grouped standard deviation. The effect size may be calculated according to Eq.6: [Eq.6] where ^1/ ^2 and ^1/ ^2 are the mean and standard deviation of the parameter values of the two patient groups. ANOVA Parameter Abbreviation Effect p-value Deviation Coefficient Ranking name size 1 (A) RBP delta SPRBP,delta,(FR) 0.55 4.76e- inc 1.10 (FR) 13 2 (B) GE functional FCPGE,delta,(BR), PLV 0.44 7.60e- inc 1.85 connectivity 09 delta band (BR) PLV 3 (C) CS functional FCPCS,delta,(BR), 0.43 1.26e- inc 2.09 * 10-2 connectivity IMCOH 08 delta band (BR) IMCOH 4 (D) EXP (FL) SPEXP,bd, (FL) 0.39 5.39e- inc 2.75 * 10-2 08 5 (F) PF broadband SPPF, bd, (BL) 0.39 2.81e- dec 6.62 * 10-3 (BL) 07 6 (E) GE functional FCPGE,bd,(FR), AEC 0.35 2.43e- inc -1.94 connectivity 06 broadband (BR) AEC 7 (G) RBP gamma SPRBP,gamma,(FR) 0.34 5.15e- dec -0.36 (FR) 06 8 (H) PF theta SPPF,theta,(FR) 0.34 7.12e- dec -2.79 * 10-2 (FR) 06 9 (C) GE functional FCPGE,delta,(BL), PLV 0.34 7.32e- inc -2.13 * 10-2 connectivity 06 delta band (BL) PLV 10 (D) CS functional FCPCS,delta,(FL), PLV 0.33 7.80e- inc -3.94 * 10-3 connectivity 06 delta band (FL) PLV 11 (C) GE functional FCPGE,delta,(FL), PLV 0.32 2.12e- inc 4.94 * 10-2 connectivity 05 delta band (FL) PLV ANOVA Parameter Abbreviation Effect p-value Deviation Coefficient Ranking name size 12 (I) RBP theta SPRBP,theta,(FR) 0.31 2.76e- inc -0.67 (FR) 05 13 (C) GE functional FCPGE,delta,(FR), PLV 0.29 1.15e- inc -4.13 * 10-2 connectivity 04 delta band (FR) PLV 14 PF beta1 (BR) SPPF, beta1, (BR) 0.29 1.17e- dec -2.72 * 10-2 04 15 PP delta SPPP,delta,(FR) 0.28 2.01e- inc -6.25 * 10-4 04 16 (C) GE functional FCPGE,delta,(BR), AEC 0.27 2.82e- inc 1.34 connectivity 04 delta band (BR) AEC Table 4: Top 16 features selected with the iterative feature selection scheme, and ranked by ANOVA ranking. The letter in brackets (A to I) in the first column is the parameter name in Table 1. The Deviation column indicates whether an increase (inc) or decrease (dec) in the parameter value compared with a non-epileptic subject was evident in the epileptic subject. The effect size, p-value and coefficient (weight) per parameter are also indicated. Based on the data in Table 4, it was evident that epilepsy patients can be characterised by slowing of the EEG signal and increased connectivity and network integration in the lower frequencies (i.e. an increase in parameter 1: delta band). Subsequently each parameter of Table 4 was used separately or cumulatively to train a probabilistic machine learning model (PMLM), from which it could be determined which parameter(s) had the most predictive power. A retrospective dataset of 914 EEGs, recorded from patients who arrived at the epilepsy unit after a first insult and for whom the diagnosis of epilepsy was initially unclear was utilised. For all patients, diagnostic labels after a follow-up period of 2 years or more (i.e., during this period clinicians were able to prove whether the patient suffered from epilepsy or not) were assigned. These labels were used to perform a group analysis on the parameters, and train a probabilistic machine learning model, PMLM. A nested cross-validation scheme was used to train and test the probabilistic machine learning model, PMLM, that was a Logistic Regression model. During training, the model was optimised to predict the conditional probabilities of the training subject belonging to a certain class (subject suffering from epilepsy or subject not suffering from epilepsy). This was quantified with the Brier Score (BrS), which compared the output probability from the Logistic Regression model with the actual probability. The actual probability for a subject suffering from epilepsy was 1, and the actual probability for a subject not suffering from epilepsy was 0. The score was computed according to the following formula (Eq.7):
Figure imgf000038_0001
with N being the number of training subjects, pi the output probability from the model, and oi the actual (observed) probability of training subject i. The Brier Score becomes lower as the predictions pi correspond better with the observations oi, so during training the BrS is minimized. The aim of the model training is hence to lower the BrS. The nested cross-validation comprised two 5-fold cross-validation schemes. First, the data was split into 5 parts (also called folds) of equal size and each containing the same proportion of epilepsy and non-epilepsy patients. In each iteration of the outer split, one fold is set aside to be used as test set, the other four folds are taken together and again divided into 5 equal folds for the inner cross-validation. The inner cross-validation is used for optimizing and calibrating the classifier, the outer cross-validation is used to test the performance on unseen data. FIGs.4-5 (#1 to #16) show the performance when multiple PMLMs were trained, one PMLM for each individual training parameter of Table 4. The name of the graph (e.g. #1) corresponds to the ANOVA ranking (e.g.1) in Table 4 and to the training parameter used to train that PMLM. Delta (FR) bandpower (FIG.4, #1) appears to be the most predictive (BrS = 0.2293) with good performance across a relatively wide range of possible probabilities. The next most predictive PMLMs are shown in order (FIG.4 to 5, #2 to #15). FIG.6-7 (#1 to #16) shows the performance when multiple PMLMs were trained, each PMLM trained with an increasing number of different parameters. From the name of the graph (#1 to #16), it can be determined the number of training parameters and which different training parameters were used to train the PMLM. For instance, FIG.6 #3 used 3 different training parameters which had ANOVA rankings 1 to 3 in Table 4. For instance, FIG.6 #4 used 4 different training parameters which had ANOVA rankings 1 to 4 in Table 4 etc. As expected, there was an improvement in calibration curve and Brier Score (BS) with each addition, until a total of 12 parameters had been reached. After this, the performance became worse again, which might indicate overfitting on the training data for more than 12 features. FIG.8-9 shows the performance when multiple PMLMs were trained, each with an increasing number of different parameters, but when the regional specificity of the parameter was removed. Thus, each parameter was averaged over all regions (AR) (and thus not confined to one quadrant of the scalp FL, FR, BL, BR). From name of the PMLM (#1 to #16), it can be determined the number of training parameters and which different training parameters were used to train the PMLM. For instance, FIG.8 #3 used 3 training parameters which had ANOVA rankings 1 to 3 in Table 4. For instance, FIG.8 #4 used 4 training parameters which had ANOVA rankings 1 to 4 in Table 4 etc. The performance of each PMLM is comparable to the results obtained in FIG.4-5. The results show that a parameter defining specific spatial regions might be preferable but it is not essential. Based on the data in FIGs.4 to 9, a ranked list of the most predictive parameters could be determined, which separately or in combination, are indicative of a likelihood of epilepsy in a test subject. The list is presented in Table 1. After the PMLM was trained, it was calibrated to further improve the accuracy of the probability predictions. Platt’s scaling method was used, where unwanted distortions in the calibration curve of the PMLM were reduced by fitting a sigmoid function to the curve and rescaling the curve to the desired trend (Niculescu-Mizil, 2005). A part of the large retrospective dataset of EEGs mentioned in Example 1 was also used to validate the PMLM. The predicted probabilities on the test folds were put together and visualized with a calibration curve as shown in FIG.10. The calibration curve was generated from a calibrated PMLM trained using 12 parameters (the 12 highest ranked parameters in Table 4). The calibration curve of the trained model shows a good correspondence between the predicted probabilities and the actual fraction of positives. This is also evidenced by a low Brier Score for the PMLM (BrS = 0.23). In other words, the model reliably predicts, based on spectral and/or connectivity features from a resting state EEG recording, what the probability is that a patient has/will develop epilepsy. Example 2 A deterministic machine learning model to classify epilepsy subjects vs non-epilepsy subjects was trained. The first steps (parameters, standardization, iterative feature selection) were the same as in Example 1. A 5-fold cross-validation scheme was used to train and evaluate a Naïve Bayes classifier. The classifier was optimized based on the area under the receiver operating curve (AUC). The performance of the trained model was expressed in terms of sensitivity and specificity, and a confusion matrix that shows the predicted labels vs the true labels. The resulting performance of the deterministic machine learning model is shown in FIG.11. It achieved an AUC of 0.60, sensitivity of 36% and specificity of 77%, which is in line with the current diagnostic yield of EEG in clinical practice based on visual analysis of the EEG signals, with sensitivity ranging between 25-56% and specificity between 78-98% (Smith et al., 2005). It did not, however, provide an improvement over the current accuracies. The trained PMLM on the other hand (FIG. 10) performed significantly better, evidenced by a calibration curve close to the diagonal and a low Brier Score (BrS = 0.23). Example 3 A new dataset was used to validate the calibration curve of Example 1. The new data set contained 137 patients, each of whom had presented as potentially having first attack of epilepsy, and had not been used to train the PMLM of Example 1. Of the 137 patients, 89 patients (epilepsy group) were later diagnosed to have epilepsy based on a follow-up study of at least 2 years duration, and 48 were shown to be suffering from other conditions (control groups having a later diagnosis of a “cardiovascular” or “other” health related problem) and not epilepsy. The EEG recordings were collected for several hours (long-term EEG), compared with Example 1 where the EEG recordings were collected for 15 minutes (routine EEG). Twenty parameters were measured in each patient and applied to the calibration curve of Example 1. The parameters (in no particular order were): PF delta (BR), PF delta (BL), PF delta, PF theta (BL), PF theta (FR), PP theta (FL), BW theta (BL), BW theta (FR), BW alpha, RBP alpha (BR), PF beta1 (FR), BW beta1 (FL), RBP beta1 (FR), PF beta2 (BR), BW beta2 (BR), BW 1-30Hz (FR), RBP 1-30Hz, OFF (FL). The results are shown in FIGs.12 and 13. In FIG.12 the epilepsy group showed a predicted probably of epilepsy (0.65) greater than the control group (cardiovascular (0.6) and other (0.59)). The differences are highly significant and the p-values are far below 0.05 (0.6e-7 and 0.2e-10). In FIG.13, calibration curves show good correspondence between the new dataset (long term EEG) with the dataset used in Example 1 to validate the calibration curve (routine term EEG), and both calibrations curves are close to the diagonal. In other words, the model, exposed to a completely different dataset, still reliably predicts, based on spectral and/or connectivity features from a resting state EEG recording, what the probability is that a patient has/will develop epilepsy.
References - Baccalá LA, Sameshima K. Partial directed coherence: a new concept in neural structure determination. Biol Cybern.2001 Jun;84(6):463-74. doi: 10.1007/PL00007990. PMID: 11417058. - Bastos AM, Schoffelen JM, A Tutorial Review of Functional Connectivity Analysis Methods and Their Interpretational Pitfalls, Front. Syst. Neurosci., 08 January 2016, Volume 9 - 2015, doi: 10.3389/fnsys.2015.00175. - Bouma, H.K., Labos, C., Gore, G.C., Wolfson, C. and Keezer, M.R. (2016), The diagnostic accuracy of routine electroencephalography after a first unprovoked seizure. Eur J Neurol, 23: 455-463. https://doi.org/10.1111/ene.12739. - Briels CT, Schoonhoven DN, Stam CJ, de Waal H, Scheltens P & Gouw AA, Reproducibility of EEG functional connectivity in Alzheimer’s disease, Alzheimer's Research & Therapy, volume 12, Article number: 68 (2020), doi: 10.1186/s13195- 020-00632-3. - Bronez T. P., "On the performance advantage of multitaper spectral analysis," in IEEE Transactions on Signal Processing, vol.40, no.12, pp.2941-2946, Dec.1992, doi: 10.1109/78.175738. - Burkholder DB, Britton JW, Rajasekaran V, Fabris RR, Cherian PJ, Kelly-Williams KM, So EL, Nickels KC, Wong-Kisiel LC, Lagerlund TD, Cascino GD, Worrell GA, Wirrell EC. Routine vs extended outpatient EEG for the detection of interictal epileptiform discharges. Neurology.2016 Apr 19;86(16):1524-30. doi: 10.1212/WNL.0000000000002592. Epub 2016 Mar 16. PMID: 26984946; PMCID: PMC4836883. - Fisher RS, Acevedo C, Arzimanoglou A, Bogacz A, Cross JH, Elger CE, Engel J Jr, Forsgren L, French JA, Glynn M, Hesdorffer DC, Lee BI, Mathern GW, Moshé SL, Perucca E, Scheffer IE, Tomson T, Watanabe M, Wiebe S. ILAE official report: a practical clinical definition of epilepsy. Epilepsia.2014 Apr;55(4):475-82. doi: 10.1111/epi.12550. Epub 2014 Apr 14. PMID: 24730690. - Jas M, Engemann DA, Bekhti Y, Raimondo F, Gramfort A. Autoreject: Automated artifact rejection for MEG and EEG data. Neuroimage.2017 Oct 1;159:417-429. doi: 10.1016/j.neuroimage.2017.06.030. Epub 2017 Jun 20. PMID: 28645840; PMCID: PMC7243972. - Krishnan P, Yaacob S, Krishnan AP, Rizon M, Ang CK, EEG based Drowsiness Detection using Relative Band Power and Short-time Fourier Transform, Journal of Robotics, Networking and Artificial Life, Vol.7(3); December (2020), pp.147–151, doi: 10.2991/jrnal.k.200909.001. - Krumholz A, Wiebe S, Gronseth G, Shinnar S, Levisohn P, Ting T, Hopp J, Shafer P, Morris H, Seiden L, Barkley G, French J; Quality Standards Subcommittee of the American Academy of Neurology; American Epilepsy Society. Practice Parameter: evaluating an apparent unprovoked first seizure in adults (an evidence-based review): report of the Quality Standards Subcommittee of the American Academy of Neurology and the American Epilepsy Society. Neurology.2007 Nov 20;69(21):1996- 2007. doi: 10.1212/01.wnl.0000285084.93652.43. PMID: 18025394. - Laman DM, Wieneke GH.†; van Duijn Hans, Veldhuizen RJ, van Huffelen, AC, QEEG Changes During Carotid Clamping in Carotid Endarterectomy: Spectral Edge Frequency Parameters and Relative Band Power Parameters, Journal of Clinical Neurophysiology 22(4):p 244-252, August 2005. doi: 10.1097/01.WNP.0000167931.83516.CF.Lachaux JP, Rodriguez E, Martinerie J, Varela FJ. Measuring phase synchrony in brain signals. Hum Brain Mapp. 1999;8(4):194-208. doi: 10.1002/(sici)1097-0193(1999)8:4<194::aid-hbm4>3.0.co;2- c. PMID: 10619414; PMCID: PMC6873296. - Medel V, Irani M, Crossley N, Ossandón T & Boncompte G, Complexity and 1/f slope jointly reflect brain states, Scientific Reports volume 13, Article number: 21700 (2023); doi: 10.1038/s41598-023-47316-0. - Niculescu-Mizil, A., & Caruana, R. (2005). Predicting good probabilities with supervised learning. Proceedings of the 22nd International Conference on Machine Learning - ICML ’05. https://doi.org/10.1145/1102351.1102430. - Nolte G, Bai O, Wheaton L, Mari Z, Vorbach S, Hallett M. Identifying true brain interaction from EEG data using the imaginary part of coherency. Clin Neurophysiol. 2004 Oct;115(10):2292-307. doi: 10.1016/j.clinph.2004.04.029. PMID: 15351371. - Sakkalis V. Review of advanced techniques for the estimation of brain connectivity. measured with EEG/MEG. Comput Biol Med.2011 Dec;41(12):1110-7. doi: 10.1016/j.compbiomed.2011.06.020. Epub 2011 Jul 26. PMID: 21794851. - Smith SJ. EEG in the diagnosis, classification, and management of patients with epilepsy. J Neurol Neurosurg Psychiatry.2005 Jun;76 Suppl 2(Suppl 2):ii2-7. doi: 10.1136/jnnp.2005.069245. PMID: 15961864; PMCID: PMC1765691. - Sörnmo L, Laguna P, Bioelectrical Signal Processing in Cardiac and Neurological Applications - A volume in Biomedical Engineering, Book, 2005, doi: 10.1016/B978- 0-12-437552-9.X5000-4. - Sporns O, Graph theory methods: applications in brain networks, Dialogues in Clinical Neuroscience Volume 20, 2018 - Issue 2, pages 111-121. doi: doi.org/10.31887/DCNS.2018.20.2/osporns. - Wang Y, Sokhadze EM, El-Baz AS, Xiaoli L, Sears L, Casanova MF, Allan T, Relative Power of Specific EEG Bands and Their Ratios during Neurofeedback Training in Children with Autism Spectrum Disorder, Front. Hum. Neurosci., 14 January 2016, Sec. Brain Imaging and Stimulation, Volume 9 - 2015, doi: 10.3389/fnhum.2015.00723.

Claims

Claims 1. A computer-implemented method for determining a likelihood of a test subject having epilepsy after having suffered a possible epileptic seizure, comprising: - receiving measurement data comprising a measurement parameter set derived from an EEG recording of the test subject in a resting state wherein the measurement parameter set comprises Parameter A; and - determining from the measurement data the likelihood of the test subject having epilepsy, wherein parameter A is relative band power, RBP, of a power spectral density, PSD, curve limited to a delta frequency band and limited to a front right region of the head of the subject. 2. The computer-implemented method according to claim 1, wherein an increase in Parameter A of the measurement parameter set compared with a healthy reference Parameter A is indicative of an increased likelihood of the test subject having epilepsy. 3. The computer-implemented method according to claim 1 or 2, wherein the measurement parameter set further comprises one or more of Parameter B, Parameter C, Parameter D, wherein each parameter is derived from the EEG recording of the test subject in the resting state; wherein parameter B is a global efficiency functional connectivity parameter that is a statistical combination of (SPLs)-1 limited to a delta frequency band, in all regions of the head of the subject, wherein the SPL is a shortest path length between EEG electrode pairs in all regions of the head of the subject determined according to the path which yields a smallest summation of inverse connectivity strengths, CS-1, values along that path; wherein parameter C is a connectivity strength functional connectivity parameter that is a statistical combination of connectivity strengths, CS, limited to a delta frequency band between EEG electrodes pairs limited to a back right, BR, and/or front left, FL, region of the head of the subject; and wherein parameter D is a 1/f exponent parameter that is a statistical combination of slope of a decay of a power spectral density, PSD, curve limited to a broadband frequency band for each EEG electrode limited to a front left (FL), region of the head of the subject. 4. The computer-implemented method according to claim 3, wherein: an increase in Parameter B of the measurement parameter set compared with a healthy reference Parameter B is indicative of an increased likelihood of the test subject having epilepsy; and/or an increase in Parameter C of the measurement parameter set compared with a healthy reference Parameter C is indicative of an increased likelihood of the test subject having epilepsy; and/or an increase in Parameter D of the measurement parameter set compared with a healthy reference Parameter D is indicative of an increased likelihood of the test subject having epilepsy. 5. The computer-implemented method according to any one of claims 1 to 4, wherein the measurement parameter set further comprises one or more of Parameter E, Parameter F, Parameter G, Parameter H, wherein each parameter is derived from the EEG recording of the test subject in the resting state; wherein parameter E is a global efficiency functional connectivity parameter that is a statistical combination of (SPLs)-1 limited to a broadband frequency band, and limited to a front right (FR), of the head of the subject, wherein the SPL is a shortest path length between EEG electrode pairs in all regions of the head of the subject determined according to the path which yields a smallest summation of inverse connectivity strengths, CS-1, values along that path; wherein parameter F is a peak frequency parameter that is a statistical combination of frequency of a peak (maximal) signal of a power spectral density, PSD, curve limited to a broadband frequency band for each EEG electrode limited to a back left, BL, region of the head of the subject; wherein parameter G is relative band power, RBP, of a power spectral density, PSD, curve limited to a gamma frequency band limited to a front right region of the head of the subject; wherein parameter H is a peak frequency parameter that is a statistical combination of frequency of a peak (maximal) signal of a power spectral density, PSD, curve limited to a theta frequency band for each EEG electrode limited to a front right, FR, region of the head of the subject. 6. The computer-implemented method according to claim 5, wherein a decrease in Parameter E of the measurement parameter set compared with a healthy reference Parameter E is indicative of an increased likelihood of the test subject having epilepsy; and/or a decrease in Parameter F of the measurement parameter set compared with a healthy reference Parameter F is indicative of an increased likelihood of the test subject having epilepsy; and/or a decrease in Parameter G of the measurement parameter set compared with a healthy reference Parameter G is indicative of an increased likelihood of the test subject having epilepsy; an increase in Parameter H of the measurement parameter set compared with a healthy reference Parameter H is indicative of an increased likelihood of the test subject having epilepsy. 7. The computer-implemented method according to any one of claim 1 to 6, wherein the measurement parameter set comprises Parameter A, Parameter B, Parameter C, Parameter D, wherein each parameter is derived from the EEG recording of the test subject in a resting state. 8. The method according to any one of claims 1 to 7, wherein the determining is performed by comparing the measurement parameter with a healthy parameter set. 9. A computer-implemented training method for training a probabilistic machine learning model, PMLM, for determining a likelihood of a test subject having epilepsy comprising, the training method comprising: - receiving a plurality of training data sets, wherein: -each training data set has been acquired from a training subject and contains: - a training parameter set that has been measured in the training subject, and - a training tag that is indication of presence or absence of epilepsy in the training subject; - training the PMLM using the plurality of training data sets, wherein: - an input to the PMLM is the training parameter set of the data training set, and an output of PMLM is a pending tag; - the pending tag is compared with the training tag of the training data set, and - the PMLM is adjusted so that the pending tag approaches the training tag of the training set. 10. The computer-implemented method according to any one of claims 1 to 9, wherein the determining is performed using the measurement data as an input to a trained PMLM, trained according to the training method according to claim 9. 11. A computing device or system configured for performing the method of any one of claims 1 to 10. 12. A computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform the method of any one of claims 1 to 10. 13. A computer readable medium having stored thereon a computer program having instructions which when executed by a computing device or system cause the computing device or system to perform the method of any one of claims 1 to 10. 14. A data stream which is representative of a computer program or computer program product having instructions which when executed by a computing device or system cause the computing device or system to perform the method of any one of claims 1 to 10. 15. A wearable headset for determining a likelihood of a subject having epilepsy, comprising a computing device and plurality of EEG electrodes operatively connected to the computing device, wherein the computing device is configured to carry out the method of any one of claims 1 to 10.
PCT/EP2024/070335 2023-07-18 2024-07-18 Method and system for prediction of epilepsy Pending WO2025017101A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
EP23185994.3 2023-07-18
EP23185994 2023-07-18

Publications (1)

Publication Number Publication Date
WO2025017101A1 true WO2025017101A1 (en) 2025-01-23

Family

ID=87418945

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2024/070335 Pending WO2025017101A1 (en) 2023-07-18 2024-07-18 Method and system for prediction of epilepsy

Country Status (1)

Country Link
WO (1) WO2025017101A1 (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150216436A1 (en) * 2012-09-07 2015-08-06 Children's Medical Center Corporation Detection of epileptogenic brains with non-linear analysis of electromagnetic signals
US20160183828A1 (en) * 2014-12-30 2016-06-30 I-Shou University Electronic apparatus for establishing prediction model based on electroencephalogram
US20210267530A1 (en) * 2018-06-29 2021-09-02 Mybrain Technologies Multiclass classification method for the estimation of eeg signal quality

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150216436A1 (en) * 2012-09-07 2015-08-06 Children's Medical Center Corporation Detection of epileptogenic brains with non-linear analysis of electromagnetic signals
US20160183828A1 (en) * 2014-12-30 2016-06-30 I-Shou University Electronic apparatus for establishing prediction model based on electroencephalogram
US20210267530A1 (en) * 2018-06-29 2021-09-02 Mybrain Technologies Multiclass classification method for the estimation of eeg signal quality

Non-Patent Citations (24)

* Cited by examiner, † Cited by third party
Title
BACCALÁ LASAMESHIMA K.: "Partial directed coherence: a new concept in neural structure determination", BIOL CYBERN., vol. 84, no. 6, June 2001 (2001-06-01), pages 463 - 74, XP055759914
BASTOS AMSCHOFFELEN JM: "A Tutorial Review of Functional Connectivity Analysis Methods and Their Interpretational Pitfalls", FRONT. SYST. NEUROSCI., vol. 9, 8 January 2016 (2016-01-08)
BOUMA, EUROPEAN JOURNAL OF NEUROLOGY, 2016
BOUMA, H.KLABOS, C.GORE, G.CWOLFSON, C.KEEZER, M.R.: "The diagnostic accuracy of routine electroencephalography after a first unprovoked seizure", EUR J NEUROL, vol. 23, 2016, pages 455 - 463, XP072030416, DOI: 10.1111/ene.12739
BRIELS CTSCHOONHOVEN DNSTAM CJDE WAAL HSCHELTENS PGOUW AA: "Reproducibility of EEG functional connectivity in Alzheimer's disease", ALZHEIMER'S RESEARCH & THERAPY, vol. 12, 2020
BRONEZ T. P.: "On the performance advantage of multitaper spectral analysis", IEEE TRANSACTIONS ON SIGNAL PROCESSING, vol. 40, no. 12, December 1992 (1992-12-01), pages 2941 - 2946, XP000941539, DOI: 10.1109/78.175738
BURKHOLDER DBBRITTON JWRAJASEKARAN VFABRIS RRCHERIAN PJKELLY-WILLIAMS KMSO ELNICKELS KCWONG-KISIEL LCLAGERLUND TD: "Routine vs extended outpatient EEG for the detection of interictal epileptiform discharges", NEUROLOGY, vol. 86, no. 16, 19 April 2016 (2016-04-19), pages 1524 - 30
DE STEFANO PIA ET AL: "Increased delta power as a scalp marker of epileptic activity: a simultaneous scalp and intracranial electroencephalography study", EUROPEAN JOURNAL OF NEUROLOGY, vol. 29, no. 1, 1 January 2022 (2022-01-01), GB, pages 26 - 35, XP093211368, ISSN: 1351-5101, Retrieved from the Internet <URL:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9293335/pdf/ENE-29-26.pdf> DOI: 10.1111/ene.15106 *
FISHER RSACEVEDO CARZIMANOGLOU ABOGACZ ACROSS JHELGER CEENGEL J JRFORSGREN LFRENCH JAGLYNN M: "ILAE official report: a practical clinical definition of epilepsy", EPILEPSIA, vol. 55, no. 4, April 2014 (2014-04-01), pages 475 - 82, XP055206122, DOI: 10.1111/epi.12550
JAS MENGEMANN DABEKHTI YRAIMONDO FGRAMFORT A.: "Autoreject: Automated artifact rejection for MEG and EEG data", NEUROIMAGE, vol. 159, 1 October 2017 (2017-10-01), pages 417 - 429, XP085258102, DOI: 10.1016/j.neuroimage.2017.06.030
KRISHNAN PYAACOB SKRISHNAN APRIZON MANG CK: "EEG based Drowsiness Detection using Relative Band Power and Short-time Fourier Transform", JOURNAL OF ROBOTICS, NETWORKING AND ARTIFICIAL LIFE, vol. 7, no. 3, December 2020 (2020-12-01), pages 147 - 151
KRUMHOLZ AWIEBE SGRONSETH GSHINNAR SLEVISOHN PTING THOPP JSHAFER PMORRIS HSEIDEN L: "Practice Parameter", AMERICAN EPILEPSY SOCIETY, article "Quality Standards Subcommittee of the American Academy of Neurology"
LAMAN DMWIENEKE GH.TVAN DUIJN HANSVELDHUIZEN RJVAN HUFFELEN, AC: "QEEG Changes During Carotid Clamping in Carotid Endarterectomy: Spectral Edge Frequency Parameters and Relative Band Power Parameters", JOURNAL OF CLINICAL NEUROPHYSIOLOGY, vol. 22, no. 4, pages 244 - 252
MATOS JOÃO ET AL: "Diagnosis of Epilepsy with Functional Connectivity in EEG after a Suspected First Seizure", vol. 9, no. 11, 14 November 2022 (2022-11-14), pages 690, XP093105765, Retrieved from the Internet <URL:https://biblio.ugent.be/publication/01GRNKBZ9KBMEMW4Q5RN4PB9AP/file/01GRNKDDHBFJ197WCN8RWH5WRH.pdf> DOI: 10.3390/bioengineering9110690 *
MEDEL VIRANI MCROSSLEY NOSSANDON TBONCOMPTE G: "Complexity and 1/f slope jointly reflect brain states", SCIENTIFIC REPORTS, vol. 13, 2023
NICULESCU-MIZIL, A.CARUANA, R.: "Predicting good probabilities with supervised learning", PROCEEDINGS OF THE 22ND INTERNATIONAL CONFERENCE ON MACHINE, 2005
NOLTE GBAI OWHEATON LMARI ZVORBACH SHALLETT M.: "Identifying true brain interaction from EEG data using the imaginary part of coherency", CLIN NEUROPHYSIOL., vol. 115, no. 10, October 2004 (2004-10-01), pages 2292 - 307, XP004559973
SAKKALIS V.: "Review of advanced techniques for the estimation of brain connectivity", COMPUT BIOL MED., vol. 41, no. 12, December 2011 (2011-12-01), pages 1110 - 7, XP028119327, DOI: 10.1016/j.compbiomed.2011.06.020
SMITH SJ.: "EEG in the diagnosis, classification, and management of patients with epilepsy", J NEUROL NEUROSURG PSYCHIATRY., vol. 76, June 2005 (2005-06-01)
SÖRNMO LLAGUNA P: "Bioelectrical Signal Processing in Cardiac and Neurological Applications - A volume in Biomedical Engineering", BOOK, 2005
SPORNS O: "Graph theory methods: applications in brain networks", DIALOGUES IN CLINICAL NEUROSCIENCE, vol. 20, 2018, pages 111 - 121
THANGAVEL PRASANTH ET AL: "Improving automated diagnosis of epilepsy from EEGs beyond IEDs", JOURNAL OF NEURAL ENGINEERING, IOP PUBLISHING, BRISTOL, GB, vol. 19, no. 6, 24 November 2022 (2022-11-24), XP020469103, ISSN: 1741-2560, [retrieved on 20221124], DOI: 10.1088/1741-2552/AC9C93 *
VARELA FJ.: "Measuring phase synchrony in brain signals", HUM BRAIN MAPP., vol. 8, no. 4, 1999, pages 194 - 208, XP055720986, DOI: 10.1002/(SICI)1097-0193(1999)8:4<194::AID-HBM4>3.0.CO;2-C
WANG YSOKHADZE EMEL-BAZ ASXIAOLI LSEARS LCASANOVA MFALLAN T: "Relative Power of Specific EEG Bands and Their Ratios during Neurofeedback Training in Children with Autism Spectrum Disorder", FRONT. HUM. NEUROSCI., vol. 9, 14 January 2016 (2016-01-14)

Similar Documents

Publication Publication Date Title
Zavecz et al. NREM sleep as a novel protective cognitive reserve factor in the face of Alzheimer's disease pathology
Babaeeghazvini et al. Brain structural and functional connectivity: A review of combined works of diffusion magnetic resonance imaging and electro-encephalography
Franciotti et al. Default network is not hypoactive in dementia with fluctuating cognition: an Alzheimer disease/dementia with Lewy bodies comparison
Bernardo et al. Visual and semi-automatic non-invasive detection of interictal fast ripples: A potential biomarker of epilepsy in children with tuberous sclerosis complex
Jacobs et al. Hemodynamic changes preceding the interictal EEG spike in patients with focal epilepsy investigated using simultaneous EEG-fMRI
Jelic et al. Quantitative electroencephalography in mild cognitive impairment: longitudinal changes and possible prediction of Alzheimer’s disease
Van Diessen et al. Are high frequency oscillations associated with altered network topology in partial epilepsy?
Gupta et al. Space–time network connectivity and cortical activations preceding spike wave discharges in human absence epilepsy: a MEG study
Bosma et al. Disturbed functional brain networks and neurocognitive function in low-grade glioma patients: a graph theoretical analysis of resting-state MEG
Kisler et al. Abnormal alpha band power in the dynamic pain connectome is a marker of chronic pain with a neuropathic component
Westlake et al. Resting state alpha-band functional connectivity and recovery after stroke
Franciotti et al. Cortical rhythms reactivity in AD, LBD and normal subjects: a quantitative MEG study
Abreu et al. Ballistocardiogram artifact correction taking into account physiological signal preservation in simultaneous EEG-fMRI
Fujiwara et al. Ictal MEG onset source localization compared to intracranial EEG and outcome: improved epilepsy presurgical evaluation in pediatrics
Miao et al. Using ictal high-frequency oscillations (80–500 Hz) to localize seizure onset zones in childhood absence epilepsy: a MEG study
US20090054801A1 (en) Method and device for determining depressive disorders by measuring bioelectromagnetic signals of the brain
IL191088A (en) System and method for prediction of cognitive decline
Dubovik et al. Adaptive reorganization of cortical networks in Alzheimer’s disease
Koshiyama et al. Abnormal phase discontinuity of alpha-and theta-frequency oscillations in schizophrenia
McMackin et al. Dysfunction of attention switching networks in amyotrophic lateral sclerosis
Wu et al. Quantify neuromagnetic network changes from pre-ictal to ictal activities in absence seizures
Lodder et al. Automated EEG analysis: characterizing the posterior dominant rhythm
van Houdt et al. Are epilepsy-related fMRI components dependent on the presence of interictal epileptic discharges in scalp EEG?
Waldman et al. Ripple oscillations in the left temporal neocortex are associated with impaired verbal episodic memory encoding
Yi et al. Capturing the abnormal brain network activity in early Parkinsons disease with mild cognitive impairment based on dynamic functional connectivity

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 24745731

Country of ref document: EP

Kind code of ref document: A1