[go: up one dir, main page]

WO2022101368A1 - Diagnostic clinique du trouble déficitaire de l'attention avec hyperactivité (tdah) chez les adultes - Google Patents

Diagnostic clinique du trouble déficitaire de l'attention avec hyperactivité (tdah) chez les adultes Download PDF

Info

Publication number
WO2022101368A1
WO2022101368A1 PCT/EP2021/081442 EP2021081442W WO2022101368A1 WO 2022101368 A1 WO2022101368 A1 WO 2022101368A1 EP 2021081442 W EP2021081442 W EP 2021081442W WO 2022101368 A1 WO2022101368 A1 WO 2022101368A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
adhd
machine learning
clinical
outcome
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.)
Ceased
Application number
PCT/EP2021/081442
Other languages
English (en)
Other versions
WO2022101368A9 (fr
Inventor
Marios ADAMOU
Grigoris ANTONIOU
Tianhua Chen
Ilias TACHMAZIDIS
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.)
South West Yorkshire Partnership Nhs Foundation Trust
University of Huddersfield
Original Assignee
South West Yorkshire Partnership Nhs Foundation Trust
University of Huddersfield
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 South West Yorkshire Partnership Nhs Foundation Trust, University of Huddersfield filed Critical South West Yorkshire Partnership Nhs Foundation Trust
Priority to GB2305959.5A priority Critical patent/GB2614862A/en
Publication of WO2022101368A1 publication Critical patent/WO2022101368A1/fr
Publication of WO2022101368A9 publication Critical patent/WO2022101368A9/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • 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
    • 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
    • G16H10/00—ICT specially adapted for the handling or processing of patient-related medical or healthcare data
    • G16H10/20—ICT specially adapted for the handling or processing of patient-related medical or healthcare data for electronic clinical trials or questionnaires
    • 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
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients

Definitions

  • ADHD Attention Deficit Hyperactivity Disorder
  • the present disclosure relates to a computer system and method for clinical diagnosis of ADHD in adults.
  • ADHD Attention-deficit/hyperactivity disorder
  • a method of configuring a diagnostic tool for diagnosing ADHD in adults comprising: generating a user interface for a diagnostic computer system, the user interface comprising a set of input fields, each for receiving a value of a clinical parameter, the value indicative of a patient’s mental health state, wherein the set of input fields has been selected based on a determination of clinical parameters which have been found to be statistically significant in the diagnosis of ADHD in adults.
  • a method of determining the statistical relevance of a clinical parameter may comprise using knowledge of a human expert skilled in the diagnosis of ADHD in adults. Another method may comprise training a machine learning model using a collection of clinical parameters which have been determined to have some bearing in the diagnosis of ADHD in adults, and extracting features from the trained machine learning model using a machine learning feature selection process which determines features that are statistically relevant to decision outcomes of the machine learning model.
  • the machine learning model is a decision tree model which is trained on patient case data where the clinical parameters and the ADHD diagnosis ( positive or negative ) is known . Once trained , the model implements a decision tree to determine a diagnostic outcome for a patient with known clinical parameters but for whom the diagnostic outcome is unknown.
  • a method of diagnosing ADHD comprising: receiving input patient case data via a user interface of a computer system, the input data comprising a set of indicators, each indicator comprising a value indicative of a patient’s mental health state; applying at least a respective subset of the indicators to each of two models of the computer system, the two models comprising a machine learning predictive model and a knowledge model; determining using the predictive machine learning model a first diagnostic outcome based on the indicators; determining using the knowledge model a second diagnostic outcome based on the indicators; combining the first and second outcomes to generate a decision output which indicates whether the indicators are indicative of a positive diagnosis of ADHD, whether the indicators are indicative of a negative diagnosis of ADHD or whether the indicators indicate that expert opinion, for example through further, more in-depth assessment, is required.
  • a computer implemented method of processing clinical data comprising: receiving input patient case data via a user interface of a computer system, the input data comprising a set of indicators, each indicator comprising a value indicative of a patient’s mental health state; applying at least a respective subset of the indicators to each of two models of the computer system, the two models comprising a machine learning predictive model and a knowledge model; determining using the predictive machine learning model a first diagnostic outcome based on the indicators; determining using the knowledge model a second diagnostic outcome based on the indicators; combining the first and second outcomes to generate a decision output which indicates whether the indicators are indicative of a positive diagnosis of ADHD, whether the indicators are indicative of a negative diagnosis of ADHD or whether the indicators indicate that expert opinion, for example through further, more in-depth assessment, is required.
  • the subset of indicators applied to each model has preferably been selected based on a determination of which indicators are statistically relevant when used in that particular model.
  • the subset of indicators applied to the machine learning predictive model differs from, but may overlap with, the subset of indicators applied to the knowledge model.
  • the set of indicators comprise at least one indicator which represents a score derived from a structured questionnaire, for example a DIVA questionnaire.
  • the indicators may further comprise one or more of the following: an indicator of a depression state (PHQ9), an indication of an anxiety state (GAD), an indication of a bipolar state (MDQ), an indication of alcohol use (AUDIT), an indication of substance abuse (DAST), an indication of brain injury (HELPS), and an indication of a personality disorder (IOWA).
  • PHQ9 an indication of a depression state
  • GAD an indication of an anxiety state
  • MDQ an indication of a bipolar state
  • AUDIT indication of alcohol use
  • DAST substance abuse
  • HLPS brain injury
  • IOWA personality disorder
  • Each subset may include one or more of the above.
  • the models may be such that a value for the indicator is considered by the respective model only if it exceeds certain thresholds associated with those particular state indicators.
  • the invention further provides, in another aspect, a method of operating a diagnostic tool comprising entering into respective fields of a graphical user interface input patient case data comprising clinical parameters indicative of a state of a patient’s mental health, and causing the clinical parameters to be evaluated by a computer system comprising: a machine learning model which has been trained on clinical data pertaining to
  • ADHD diagnosed adult patients the clinical data comprising a set of clinical parameters, the machine learning model having been trained on the clinical data to predict as an outcome, optionally with a confidence score, a positive diagnosis of ADHD or a negative diagnosis of ADHD based on corresponding input clinical parameters; a computer implemented knowledge model configured to implement a sequence of rules, the rules assessing the input clinical parameters and generating as an outcome either a positive diagnosis of ADHD, a negative diagnosis of ADHD or an indication that expert input is required; and an output module connected to receive an outcome from the machine learning model and an outcome from the knowledge model and to generate the output decision based on the outcome of the machine learning model and the outcome of the knowledge model.
  • the confidence score may indicate the level of confidence that the machine learning model associates with the predicted outcome .
  • the clinical parameters may comprise indicators as defined above .
  • a computer system for performing a clinical diagnosis of ADHD in adults comprising: a machine learning model which has been trained on clinical data pertaining to ADHD diagnosed adult patients, the clinical data comprising a set of clinical parameters, the machine learning model having been trained on the clinical data to predict as an outcome, optionally along with a confidence score, a positive diagnosis of ADHD or a negative diagnosis of ADHD based on corresponding input clinical parameters; a computer implemented knowledge model configured to implement a sequence of rules, the rules assessing the input clinical parameters and generating as an outcome either a positive diagnosis of ADHD, a negative diagnosis of ADHD or an indication that expert input is required; a user interface configured to receive a set of input clinical parameters corresponding to a patient to be diagnosed and to supply those input parameters to each of the machine learning model and the knowledge model and to present at least an output decision to a user; and an output module connected to receive an outcome from the machine learning model and an outcome from the knowledge model and to generate the output decision based on the outcome of the machine learning model and
  • the output module may be configured to generate as the output decision a positive diagnosis of ADHD when the outcome of the machine learning model and the outcome of the knowledge model both indicate a positive diagnosis.
  • the output module may be configured to generate the output decision as a negative diagnosis when the machine learning model and the knowledge model both indicate a negative diagnosis of ADHD.
  • the output module may be configured to generate as the output decision that expert input is required when the outcome of the machine learning model differs from the outcome of the knowledge model.
  • the user interface may be configured to display to a user the outcomes of one or both of the machine learning model and the knowledge model, optionally associated with a reasoning output of the respective model.
  • the sequence of rules of the knowledge model may comprise a first rule which compares at least one clinical parameter score pertaining to a structured questionnaire carried out by a clinician with a corresponding threshold and indicates a negative diagnosis if the clinical parameter score is less than the threshold.
  • the sequence may comprise a second rule which determines whether multiple clinical parameters are indicated relative to respective thresholds and if so, generates an outcome that expert opinion is needed.
  • the sequence may comprise a third rule which compares the at least one clinical parameter score with a corresponding threshold and which indicates a positive diagnosis of ADHD if the clinical parameter score exceeds the corresponding threshold.
  • Figure 1 is a schematic diagram showing a diagnostic computer system for diagnosing ADHD in adults
  • Figure 2 illustrates a user interface for the computer system
  • Figure 3 illustrates a flowchart of the diagnostic method carried out by the computer system
  • Figure 4 illustrates a flowchart of a rule -based model
  • Figure 5 illustrates an experimental configuration for determining a best machine learning model
  • Figure 6 illustrates an example of weighted text analysis of an unstructured medical report.
  • ADHD Attention Deficit Hyperactivity Disorder
  • the present inventors have applied Al (Artificial Intelligence) to make it possible to support the clinical diagnosis of ADHD based on the analysis of relevant data. Their analysis studied data of adult patients who underwent diagnosis within the past few years. The inventors have developed a computer system and method of diagnosis using a hybrid approach, consisting of two different models: a machine learning model obtained by training on data of past cases, and a knowledge model capturing the expertise of medical experts through knowledge engineering. The resulting outputs of the hybrid approach have an accuracy of 95% on data currently available, and has to date been validated in a clinical environment on the basis of around 500 cases.
  • Al Artificial Intelligence
  • FIG. 1 illustrates a computer system according to one embodiment of the invention.
  • a computer system 100 provides a machine learning (ML) model 102 and a knowledge model 104.
  • the computer system comprises a graphical user interface 106.
  • the output of the ML model 102 is supplied to an output module 108 of the computer system 100.
  • the output of the knowledge model 104 is also supplied to the output module 108.
  • the output module 108 provides a decision output 110 to be displayed on the graphical user interface 106. The manner in which the decision output 110 is generated by the output module 108 is described later.
  • the output of the ML model 103 may also be supplied directly to the GUI 106, equally the output 105 of the knowledge module may also be supplied to the GUI 106.
  • the outputs 103 and 105 are displayed to a user using the GUI 106. It is not, however, necessary in all embodiments for the outputs of the ML model and the knowledge model to be displayed on the GUI 106. However, it has been found that clinicians prefer to use diagnostic support systems and methods where they have more visibility into the operation of such systems and methods when these systems and methods are computerised.
  • case data 112 pertaining to a patient to be diagnosed is inputted at the user interface 106.
  • a user can enter inputs into the user interface as described in more detail later.
  • a respective subset of the inputs is applied to each of the ML model 102 and knowledge model 104.
  • the subsets of case data 112, which is inputted into the system 100, is supplied to the ML model 102 and the knowledge model 104 such that each model can operate on the case data input which has been supplied to it and generate a respective output.
  • the output 103 of the ML model 102 indicates whether a diagnosis of ADHD is indicated or not indicated.
  • the output 105 of the knowledge model has three states: whether a diagnosis of ADHD is indicated or not indicated, and whether expert opinion should be sought.
  • both the ML model 102 and the knowledge model 104 have outputs indicating that a diagnosis of ADHD is indicated (that is, they are both “yes”), the decision output 110 of the output module 108 is indicated as “yes”.
  • the decision output 110 of the output module is indicated as “no”.
  • the decision output 110 of the output module provides an indication to seek expert opinion. If the state of the output of the knowledge model is that expert opinion is needed, this is applied as the decision output. That is, the system may advise a viewer of the display of the GUI 106 that there is a disparity between the models, and whether an expert opinion should be sought.
  • the ‘expert opinion’ option may be replaced or augmented by a further automated step using information stored in a memory 130 accessed by the output module, as described later.
  • Figure 2 depicts the user interface 106 configured to receive input patient case data 112, the user interface 106 comprising a plurality of data input fields 107 into which the input case data 112 is entered.
  • Each data input field 107 has an input field identifier 109, the input field identifier 109 indicating which input case data 112 belongs in which data input field 107.
  • Particular data input fields 107 of the user interface 106 may be configured to receive input case data 112 which identifies a patient.
  • a particular data input field 107 may be configured to receive input of a patient’s surname, NHS number, date of birth and/or any other input case data 112 capable of, wholly or in part, identifying the patient.
  • Other data input fields 107 may be configured to receive input case data 112 which is indicative of a patient’s health.
  • another data input field 107 may be configured to receive a patient’s average score on a questionnaire; the questionnaire score may be indicative of a patient’s mental health state, and therefore may be indicative of certain patient brain chemistry or hormone levels.
  • An exemplary set of suitable questionnaires could be as follows.
  • Conners’ ADHD Rating Scales could be used: a self reported questionnaire which considers a patient’s behavioural patterns, work or schoolwork, and social life in order to facilitate ADHD diagnoses; Conners’ ADHD rating is outlined in “(CAARS): technical manual, MHS North Tonawanda (1999).” Another such questionnaire could be the Drug Abuse Screening Test (D AST-10), a short “Yes/No” response questionnaire used to categorise the extent of a patient’s abuse of drugs; further details of the DAST-10 are given in “Skinner, H.A.: Addictive behaviours (1982).” Another suitable questionnaire could be the Iowa Personality
  • IOWA Alcohol Use Disorders Identification Test
  • AUDIT Alcohol Use Disorders Identification Test
  • MDQ Mood Disorder Questionnaire
  • the GAD-7 measuring generalized anxiety, could be used; the details of which are outlined in “Spitzer, R.L et al: A Brief Measure for Assessing Generalised Anxiety Disorder (2006).”
  • PHQ-9 Patient Health Questionnaire which measures the severity of depression, could be used, as could the HELPS 3 brain injury screening tool as detailed in “Picard, M. et al: HELPS (1991).”
  • the clinical interviews could be both structured and unstructured.
  • the structured interviews could be made using the Diagnostic Interview for ADHD in adults (DIVA), detailed in “Ramos-Quiroga, J.A. et al: Criteria and concurrent validity of DIVA (2019).”
  • the unstructured interviews could be captured from the text of final medical reports also provided, as discussed in more depth later.
  • data from the scores of the Sainsbury’s Risk Assessment Tool detailed in “Morgan, S.: Clinical Risk Management (2000)”
  • results from the objective measurement of ADHD symptoms obtained using the QbTest detailed in “Reh, V. et al: Behavioural Assessment of core ADHD Symptoms (2015),” could be used. It has been determined however that such additional data may not improve the accuracy of diagnosis.
  • Each item of input case data may represent a clinical parameter which may be used in a diagnosis of ADHD.
  • the inventors have determined that the particular set of clinical parameters offered to a clinician for entry into the user interface described herein represents a set of parameters which provide an accurate diagnostic outcome in over 95% of cases.
  • the set of parameters which is used therefore, enables a clinician to limit the input that they are making, and further limit the observations that need to be taken from a patient. This significantly improves the efficiency of a clinician both in taking patient information, and in using the diagnostic tool.
  • the input parameters described herein have been selected using the knowledge of experts to determine which parameters should be used in a rule-based flow of the knowledge model, and using a feature selection process to extract statistically relevant features from a trained machine learning model.
  • the fields of the user interface guide the clinician to enter values for all the required parameters , to ensure the consistent and accurate operation of the diagnostic tool .
  • the system uses at least a subset of input case data 112 to determine an overall assessment output 110, also referred to as a decision output, which may be displayed within a data output field 113.
  • the overall assessment output 110 is calculated via a hybrid approach comprising at least a knowledge-based (KR) model and a machine learning (ML) model.
  • the KR model is configured to, independently of the ML model, generate a KR diagnosis output 105 and a KR reasoning output 117 based at least upon a subset of the input case data 112.
  • the KR diagnosis output 105 indicates whether the KR model deduces that the patient has ADHD, doesn’t have ADHD, or requires further consultation.
  • the KR reasoning output 117 provides reasoning for the deduction based on the KR model’s interpretation of at least a subset of the input case data 112.
  • the ML model is configured to, independently of the KR model, generate a binary ML diagnosis output 103 and a ML reasoning output 119 based at least upon a subset of the input case data 112.
  • the ML model has been trained on case data to develop a decision tree to apply when the tool is used in diagnosis .
  • the ML diagnosis output 103 indicates whether the ML model deduces that the patient has ADHD or not, and the ML reasoning output 119 provides reasoning for the deduction based on the ML model’s interpretation of at least a subset of the input case data 112.
  • the ML output may be provided along with a confidence score.
  • the system combines the KR diagnosis output 105 and the ML diagnosis output 103 to generate the overall assessment output 110.
  • the system also produces an overall reasoning output 111, which, based at least upon the KR diagnosis output 105, the ML diagnosis output 103, and the predefined set of rules, provides reasoning for the hybrid model’s overall assessment output 110.
  • a particular data input field 107 may have one or more associated threshold values 115. These thresholds are used by the models as described later, and may or may not be displayed to a user of the user interface.
  • Each data output field 113 has an associated output field identifier 114, the associated output field identifier 114 indicating what each instance of output data 103, 105, 110, 111, 117 and 119 means, namely: the overall assessment diagnosis, overall reasoning output, KR diagnosis output, KR reasoning output, ML diagnosis output and ML reasoning output respectively.
  • Figure 3 is a flow chart outlining the diagnostic process.
  • the flow chart illustrates the subsets of the user input 112 being processed independently by the KR 104 and ML 102 models, the outputs 103, 105 thereof being processed by the hybrid algorithm using a predefined set of rules to generate an overall assessment diagnosis 110 and overall reasoning output 111.
  • the diagnosis outputs 103, 105 and reasoning outputs 117, 119 of each path and the diagnosis and reasoning outputs of the overall assessment 110, 111 are then displayed through the user interface 106 within a data output field 113 identified by an output field identifier 114.
  • the present diagnostic system uses clinical data collected from a NHS service, which delivers a clinical pathway that is compliant with NICE recommendations (i.e. gold standard), in a decision tool that can automate the process of making a diagnosis.
  • the clinical data exemplified herein is from an NHS specialist mental health provider in the form of screening questionnaires and clinical interviews, which are routinely collected when a new patient is referred.
  • knowledge-based systems aim to represent knowledge explicitly via tools such as if-then rules, which allow such a system to reason about how it reaches a conclusion and to provide explanation of its reasoning to the user.
  • An innovative approach is described herein which improves over a system where only machine learning-based approaches were used.
  • the present approach simultaneously employs machine learning and a knowledge-based approach in a hybrid manner.
  • the ML model 102 is a prediction model which has been trained based on machine learning using clinical data.
  • the experimental evaluation described herein demonstrates that by applying machine learning to input case data, a diagnostic accuracy of 85% can be achieved.
  • the knowledge model is created by capturing knowledge from medical experts and representing it in the form of rules that may conflict with each other. Used alone, the model seeks to give yes/no answers for clear-cut cases, while referring the remainder for further assessment by medical experts. This concept of referral in unclear cases is retained herein and allows the diagnostic system to serve as a diagnostic decision support tool that increases the productivity of a clinical team, not as a way to reduce or replace the clinical team. That is, the outputs from the two models are combined with the effect that, where they are in disagreement, patients are referred to medical experts. On this basis, the system achieves an accuracy of 95%.
  • a National Health Service specialist mental health provider (South West Yorkshire Partnership NHS Foundation Trust-SWYPFT) made available for analysis all anonymized data for assessments made of ADHD patients in the period between 2014 and 2017. Overall, there were 69 such patients. For all these patients, the data contained information which included demographics and several validated self-reported screening questionnaires and clinical interviews.
  • Each patient has a patient record containing a client ID, which is used to group all entries related to a particular patient; see below.
  • each variable from the risk assessment data is binary: that is, it only takes ‘yes’ or ‘no’ for a given assessment question. Based on the inventors’ analysis, it has been determined that inclusion of all the main assessment data, and any of the risk assessment data does not improve the accuracy of the diagnosis. Instead, they have identified a set of statistically relevant factors on which to base the models.
  • a “leave- one-out cross validation” technique was utilised such that the learning method was trained on all the data except for one patient, and a prediction was made for this particular patient. The performance of each model was evaluated with an accuracy and AUC score. While it has been determined that the Decision Tree is the best model for clinical use, it will be appreciated that any suitable machine learning algorithm may be employed, for example Support Vector Machine, Naive Bayes, Decision Tree, K-nearest Neighbour, all of which have been considered as top 10 algorithms in data mining, as well as Logistic Regression and Random Forest.
  • a prediction model based on machine learning was developed using clinical data collected by an NHS specialist mental health provider and is translated into if-then rules represented in table 2:
  • DIVA scores are used in order to inform a decision; in particular, low DIVA scores indicate that ADHD should not be inferred, while high DIVA scores provide a clearer indication towards ADHD diagnosis.
  • each indicator such as D AST- 10, IOWA, AUDIT, MDQ, GAD-7, PHQ-9 and HELPS is also used in order to assess the extent to which patients are affected by substance abuse, personality disorder, alcohol use, bipolar disorder, anxiety, depression and brain injury, respectively.
  • such indicators need to be considered and weighted towards a decision as well, since the presence of high anxiety levels or personality disorder could lead to overlapping symptoms with ADHD.
  • Rule #1 If DIVA scores are below threshold, then the decision is ‘no’
  • Rule #2 If multiple indicators are present, then the decision is ‘expert’
  • Rule #3 If DIVA scores are above threshold, then the decision is ‘yes’
  • the knowledge model initially considers the applicability of rule #1, where for low DIVA scores the decision is ‘no’. However, in case DIVA scores are high (namely rule #1 is not applicable), the knowledge model needs first to evaluate the presence of multiple indicators. Assuming there are multiple indicators present, both rule #2 and rule #3 are applicable. However, rule #2 has a higher priority, thus the decision is ‘expert’. Note that only when neither rule #1 nor #2 are applicable, the decision is based on rule #3, namely the decision is ‘yes’. Note that an indicator is considered present by the knowledge model relative to (above or below) a particular threshold.
  • the knowledge model 104 comprises a computer program arranged to execute a knowledge - based algorithm on one or more processor.
  • the steps of the algorithm are shown in Figure 4.
  • the knowledge model has access to computer memory which holds information to be used by the steps of the algorithm.
  • the knowledge model initially considers the applicability of rule number 1, which is whether or not the DIVA scores are below threshold.
  • the computer program may access memory which holds the relevant DIVA parameters and their appropriate thresholds for performing the analysis of this step. That is, each DIVA parameter is associated with the particular threshold value in a table 200 stored in memory.
  • the threshold value for DIVA_child_IA may be 7.42
  • the DIVA_child_HI threshold may be 6.66
  • the DIVA_adult_IA threshold may be 7.51
  • the DIVA_adult_HI threshold may be 6.52. It will readily be appreciated that other thresholds may be utilised.
  • rule 1 may be implemented by applying Boolean algebra to the parameters. For example, rule 1 could state that:
  • rule No. 1 is not applicable, and the algorithm moves on to rule No. 2 at step S2.
  • rule No. 2 is implemented. That is, if multiple indicators are present based on their respective thresholds, then the output decision is “expert”.
  • the relevant indicators are stored in a table in memory 202, and may include some or all of the above indicators.
  • step S2 If at step S2 it is determined that there are not multiple indicators, rule No. 2 is considered not applicable and the algorithm moves to step S3 where rule No. 3 is applied. According to rule No. 3, if the DIVA scores are above a threshold, then the decision is YES.
  • a comprehensive knowledge model requires a wide range of rules and a carefully selected rule prioritisation. This can be achieved through a trial and error process where clinical experts and knowledge engineers bridge the gap between empirical knowledge and machine- readable representation of knowledge.
  • the quality of a developed knowledge model produced by the inventors is assessed herein based on existing data. It will be apparent that the knowledge model may be periodically re-evaluated as more data is made available through new patients.
  • Substance abuse Pain, Brain injury, Personality disorder.
  • Rule: If Condition, then Decision e.g. Rule #1 : If 0 ⁇ DIVA Childhood IA ⁇ 6 and 0 ⁇ DIVA Childhood HI ⁇ 6, then No. Rules are applied in the order Rule #1 to #6 until a decision is reached.
  • the results of the knowledge model 104 are combined with the results of the machine learning model 102.
  • the machine learning model provides yes/no answers
  • the knowledge model provides yes/no/expert answers.
  • a hybrid model can be developed by combining the two approaches. When both models agree on a yes/no answer, then this is the final answer. However, in some embodiments when the two models are in disagreement, then patients are referred to a medical expert. Table 5 summarizes all possible outcomes for the hybrid model.
  • another action may be taken based on predetermined rules prior to or in addition to referral to a medical expert.
  • rules could be stored in memory 130, corresponding to outcomes such as those laid out in table 5.
  • a modified version of table 5 could be stored in memory 130 to be accessed by the output module to determine different further outcomes.
  • the hybrid model is thus more robust since a yes/no answer is endorsed by both machine learning and knowledge models. Note that referring patients to medical experts is a valid and desirable outcome since the developed algorithm is designed to speed up the diagnosis process for clear-cut cases, thus leading to a higher throughput of cases per clinical expert who is already on the team; the aim is not to reduce or replace a clinical team.
  • Another benefit of the hybrid approach is that the machine learning model provides an outcome even for cases that are referred to clinical specialists - the latter can conduct their evaluation with the additional information that the ML model had a tendency to a particular outcome.
  • Table 6 summaries the performance on the main assessment report, which consisted of 27 variables. Most algorithms achieve accuracy in the range of 70%-80%, with the decision tree algorithm having accomplished the highest accuracy as highlighted in bold, followed by Random Forest and Naive Bayes. Note that for these three best-performing algorithms, no single algorithm significantly outperforms another.
  • the experiment utilises risk assessment data; this data evaluates a patient’s historical behaviour, which may be indicative of ADHD. Based on the additional 66 variables resulting from the risk assessment data, the aforementioned machine learning algorithms construct predictive models joining the main assessment and the risk assessment data; this results in a total of 94 variables.
  • the decision tree has been determined the best overall classifier in comparison to five popular alternatives, achieving two highest accuracies and one top AUC value.
  • the decision tree algorithm generates a set of if-then rules, with each rule providing a diagnosis specified by the condition.
  • the rule base is interpretable, offering a means to explain how a conclusion is derived, which is necessary for a data-driven model to be employed in practice. Such an explanation can be offered in the reasoning output - for example as indicated in field 119 of Figure 2. Note in that field that Rule 3 does not refer to Rule s of the knowledge model, but to a rule of the decision tree of the ML model.
  • risk assessment data helped to improve performance of the decision tree from 82.609% to 85.507% when used in conjunction with the main assessment data. In some embodiments, such risk assessment data could be utilised to generate the final model.
  • the knowledge model is based on if-then rules, encoding the knowledge of medical experts.
  • the best-performing machine learning algorithm namely the decision tree algorithm, generates a set of if-then rules as well.
  • the results of these two models are combined in the hybrid system for supporting an overall prediction of an ADHD diagnosis.
  • Table 8 shows how patients were classified by the three models. It is evident that in the ML model all patients are classified as either having ADHD or not having ADHD (YES/NO outcomes only), while in the KR model approximately 55% of patients are classified to a YES/NO outcome, with almost 45% of patients instead being referred to a medical expert. This in turn leads the Hybrid model to assign 50% of the patients a YES/NO outcome and refer the other 50% of the patients to a medical expert.
  • the Hybrid model will classify the minimum number of patients to a YES/NO outcome (as both KR and ML models must provide the same classification) and the maximum number of patients will be referred to a medical expert (those referred by the KR model as well as all outcome disagreements between KR and ML models). It is worth pointing out that a rate of 50% of patients being referred to a senior clinical specialist will significantly speed up the diagnosis process, considering the fact that currently all patients are assessed by a senior clinical specialist.
  • Table 9 provides the number of patients whose outcomes were misclassified, as well as the corresponding misclassification rate.
  • the ML model misclassified 4 (out of 39) patients as having ADHD and 4 (out of 30) patients as not having ADHD. However, since in the ML model all patients are assigned a YES/NO outcome, the misclassification rates are relatively low, namely 10.26% and 13.33% for YES and NO outcomes, respectively.
  • Both the KR and Hybrid models exhibit the same number of misclassified patients (2 patients as having ADHD and 1 patient as not having ADHD). However, the Hybrid model shows a higher misclassification rate compared to the KR model.
  • Table 10 presents the accuracy of each model (the highest accuracy is highlighted in bold).
  • Table 11 provides a detailed list of outcomes for misclassified patients (with misclassified outcomes highlighted in bold). It is evident that all 8 patients were misclassified by the ML model. In comparison, only 3 (out of 8) patients were misclassified by the KR model, with 3 additional patients being classified correctly (in terms of YES/NO outcomes) and 2 patients referred to a medical expert. In general, our expectation is that with more patients being assessed in the future, a number of patients will be misclassified by the KR model whilst being correctly classified by the ML model, and vice versa. Thus, the Hybrid model will provide the best overall results since it will only misclassify patients that are misclassified by both the KR and ML models, with referrals to medical expert (due to conflicts between KR and ML models) serving as valid outcomes.
  • the present inventors used a combination of Al technologies: a machine learning model that was trained from clinical data of past cases, provided by an NHS Trust, and a knowledge model representing domain knowledge captured from clinical experts.
  • the new diagnostic system takes as input the same clinical data routinely collected by the NHS upon referral, and comes up with three possible outcomes: has ADHD, does not have ADHD, or consult medical expert.
  • This approach leads to increased productivity and throughput of cases, significantly reducing waiting lists and speeding up diagnosis and, where needed, treatment.
  • the combined use of a knowledge based model and machine learning has the potential to combine the strengths of both approaches: the intuitiveness and easy transferability of the knowledge model, and the adaptability and of the machine learning model.
  • the clinical and economic benefits are also significant.
  • the system enables a study to be undertaken to determine what proportion of cases can be handled automatically, and what percentage will need consultation of a human expert.
  • the system further enables a triaging to be conducted at the point of diagnosis to identify difficult cases that must be dealt with by senior clinical experts, while the clear-cut cases can receive a recommended diagnosis by the Al algorithm, to be verified by a less senior clinician.
  • k-fold cross validation is used whereby a model is given a dataset of known data on which training is run and an independent dataset of unknown data against which the model is tested.
  • a given data set is divided into k subsets. Each time, one of the k subsets is used as the test set and the other k- 1 subsets are put together to form a training set. Then the average error across all k trials is computed.
  • the leave-one-out cross validation is utilised, which is k-fold cross validation taken to its logical extreme, with k equal to the number of data points in the set. That is to say, the learning method is trained on all the data except for one patient, and a prediction is made for that remaining patient. The average error is computed on the basis where each patient has been used to test the performance of a given model and has been used to evaluate the learning method.
  • the next decision to make is which particular machine learning method to use for the acquisition of a final model to use in practice.
  • explanations of how a machine-made conclusion is derived play a significant role in allowing clinicians to make unbiased and informed decisions based on the machine conclusion in combination with their medical knowledge.
  • a transparent model is preferred, such that the system is able to reason about how it reaches a conclusion and provide explanation of its reasoning to end users.
  • Performancewise, the model should be accurate enough to make correct decisions.
  • a total number of six popular machine learning algorithms were tested in an effort to select the best model for clinical use:
  • - Support Vector Machine is a sequential optimisation algorithm, which aims to construct a multidimensional hyper-plane that optimally discriminates between the two classes by maximising the margin of the two data clusters.
  • Logistic Regression is a statistical model for building linear logistic regression models, which models the probability of output in terms of input.
  • a cutoff value is chosen and classifiers inputs with probability greater than the cut-off as one class, below as the other class.
  • Bayesian theorem is a simple probabilistic learning classifier that assigns class labels to problem instances, represented as vectors of feature values, based on direct application of the Bayesian theorem with strong independence assumptions.
  • Random Forest is a powerful ensemble learning method by constructing a multiple of decision tree classifiers at training time and outputting the class that is most representative among all ensemble members.
  • - Decision Tree is an algorithm to generate a tree that begins with the original training set at the root node. On each iteration of the algorithm, it iterates though every unused attribute and selects one with the largest information gain to produce subsets of the data. It continues to recuse on each subset until all attributes have been used or no more additional gains obtained.
  • Figure 5 illustrates the general experimental pipeline of the underlying study.
  • clinical data which comprises the main assessment data, risk assessment data and associated medical notes
  • three different configurations were setup to properly utilize all available data.
  • the leave-one-out cross validation was adopted to train and test a given machine learning model, i.e., to train a model using information from any N-l patients out of N total patients, and test the established model using the remaining patient.
  • the performance of each model was evaluated with an accuracy and AUC score.
  • the medical note takes down a comprehensive personal record, which is utilized to recognize significant life events such as aggression, prison, bully problem, which may contribute to the ADHD diagnosis.
  • the experiment is carried out further, with a view to utilising the information embedded in the text medical notes, which record the details pertaining to the development of ADHD symptoms over the course of growing up.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Epidemiology (AREA)
  • Biomedical Technology (AREA)
  • Primary Health Care (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

L'invention concerne un système informatique et un procédé pour poser un diagnostic clinique de TDAH chez les adultes. Un modèle d'apprentissage automatique est entraîné sur des données cliniques se rapportant à des patients adultes avec un diagnostic de TDAH. Les données cliniques comprennent un ensemble de paramètres cliniques. Le modèle d'apprentissage automatique est entraîné sur les données cliniques pour prévoir en tant que résultat un diagnostic positif du TDAH ou un diagnostic négatif du TDAH sur la base de paramètres cliniques d'entrée correspondants. Un modèle de connaissances mis en œuvre par ordinateur est configuré pour mettre en œuvre une séquence de règles, les règles évaluant les paramètres cliniques d'entrée et générant en tant que résultat soit un diagnostic positif du TDAH, soit un diagnostic négatif du TDAH, soit une indication qu'une entrée d'expert est requise. Une interface utilisateur reçoit un ensemble de paramètres cliniques d'entrée correspondant à un patient à diagnostiquer, fournit ces paramètres d'entrée au modèle d'apprentissage automatique ainsi qu'au modèle de connaissances et présente une décision de sortie à un utilisateur. Un modèle de sortie reçoit un résultat du modèle d'apprentissage automatique et un résultat du modèle de connaissances et génère la décision de sortie sur la base du résultat du modèle d'apprentissage automatique et du résultat du modèle de connaissances.
PCT/EP2021/081442 2020-11-11 2021-11-11 Diagnostic clinique du trouble déficitaire de l'attention avec hyperactivité (tdah) chez les adultes Ceased WO2022101368A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
GB2305959.5A GB2614862A (en) 2020-11-11 2021-11-11 Clinical Diagnosis of Attention Deficit Hyperactivity Disorder (ADHD) in adults

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
GB2017793.7 2020-11-11
GBGB2017793.7A GB202017793D0 (en) 2020-11-11 2020-11-11 Clinical diagnosis of attention deficit hyperactivity disorder (ADHD) in adults

Publications (2)

Publication Number Publication Date
WO2022101368A1 true WO2022101368A1 (fr) 2022-05-19
WO2022101368A9 WO2022101368A9 (fr) 2022-08-11

Family

ID=74046477

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2021/081442 Ceased WO2022101368A1 (fr) 2020-11-11 2021-11-11 Diagnostic clinique du trouble déficitaire de l'attention avec hyperactivité (tdah) chez les adultes

Country Status (2)

Country Link
GB (3) GB202017793D0 (fr)
WO (1) WO2022101368A1 (fr)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116895367A (zh) * 2023-09-11 2023-10-17 北京智精灵科技有限公司 一种基于脑功能训练的多动症训练方案推送方法及系统
US20240099623A1 (en) * 2022-09-25 2024-03-28 Eric Saewon CHANG System and methods for diagnosing attention deficit hyperactivity disorder via machine learning and deep learning
CN118364729A (zh) * 2024-06-19 2024-07-19 江西师范大学 一种基于决策树的计算机化自适应诊断测试生成方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190019581A1 (en) * 2015-12-18 2019-01-17 Cognoa, Inc. Platform and system for digital personalized medicine
US20190147128A1 (en) * 2016-06-14 2019-05-16 360 Knee Systems Pty Ltd Graphical representation of a dynamic knee score for a knee surgery
WO2020198065A1 (fr) * 2019-03-22 2020-10-01 Cognoa, Inc. Procédés et dispositifs de thérapie numérique personnalisée

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190019581A1 (en) * 2015-12-18 2019-01-17 Cognoa, Inc. Platform and system for digital personalized medicine
US20190147128A1 (en) * 2016-06-14 2019-05-16 360 Knee Systems Pty Ltd Graphical representation of a dynamic knee score for a knee surgery
WO2020198065A1 (fr) * 2019-03-22 2020-10-01 Cognoa, Inc. Procédés et dispositifs de thérapie numérique personnalisée

Non-Patent Citations (7)

* Cited by examiner, † Cited by third party
Title
DEVELOPMENT OF THE AUDIT TEST, 1982
HIRSCHFIELD, R.M, THE MOOD DISORDER QUESTIONNAIRE, 2002
MORGAN, S., CLINICAL RISK MANAGEMENT, 2000
PICARD, M. ET AL., HELPS, 1991
RAMOS-QUIROGA, J.A. ET AL., CRITERIA AND CONCURRENT VALIDITY OF DIVA, 2019
REH, V ET AL., BEHAVIOURAL ASSESSMENT OF CORE ADHD SYMPTOM, 2015
SPITZER, R.L ET AL., A BRIEF MEASURE FOR ASSESSING GENERALISED ANXIETY DISORDER, 2006

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20240099623A1 (en) * 2022-09-25 2024-03-28 Eric Saewon CHANG System and methods for diagnosing attention deficit hyperactivity disorder via machine learning and deep learning
CN116895367A (zh) * 2023-09-11 2023-10-17 北京智精灵科技有限公司 一种基于脑功能训练的多动症训练方案推送方法及系统
CN116895367B (zh) * 2023-09-11 2023-12-22 北京智精灵科技有限公司 一种基于脑功能训练的多动症训练方案推送方法及系统
CN118364729A (zh) * 2024-06-19 2024-07-19 江西师范大学 一种基于决策树的计算机化自适应诊断测试生成方法

Also Published As

Publication number Publication date
GB2614862A (en) 2023-07-19
GB202017793D0 (en) 2020-12-23
GB202305959D0 (en) 2023-06-07
WO2022101368A9 (fr) 2022-08-11
GB202102103D0 (en) 2021-03-31

Similar Documents

Publication Publication Date Title
Tutun et al. An AI-based decision support system for predicting mental health disorders
Chen et al. Automatic diagnosis of attention deficit hyperactivity disorder using machine learning
Küpper et al. Identifying predictive features of autism spectrum disorders in a clinical sample of adolescents and adults using machine learning
Chen et al. Diagnosing attention-deficit hyperactivity disorder (ADHD) using artificial intelligence: a clinical study in the UK
WO2022101368A1 (fr) Diagnostic clinique du trouble déficitaire de l'attention avec hyperactivité (tdah) chez les adultes
Giesemann et al. Predicting dropout from psychological treatment using different machine learning algorithms, resampling methods, and sample sizes
Engl et al. Identifying population segments for effective intervention design and targeting using unsupervised machine learning: an end-to-end guide
Wu et al. Systematic review of machine learning in PTSD studies for automated diagnosis evaluation
Islam et al. Exploring post-COVID-19 health effects and features with advanced machine learning techniques
Megala et al. Experimental analysis of employee depression identification system using modified learning procedure with AI assistance
Zolfagharnasab et al. A novel rule-based expert system for early diagnosis of bipolar and Major Depressive Disorder
Jafari et al. Diagnosis of borderline personality disorder based on Cyberball social exclusion task and resting-state fMRI: using machine learning approach as an auxiliary tool
Karmani et al. Machine learning based tuberculosis (ML-TB) health predictor model: early TB health disease prediction with ML models for prevention in developing countries
Uddin et al. A user friendly anger and anxiety disorder prediction scheme using machine learning and a mobile application for mental healthcare
Shahid et al. Machine learning models to predict neuropsychiatric disorders in various brain tumors
Batsakis et al. Neuro intel: A system for clinical diagnosis of attention deficit hyperactivity disorder (adhd) using artificial intelligence
Sumathi et al. FUZZY CLUSTERING BASED BAYESIAN FRAMEWORK TO PREDICT MENTAL HEALTH PROBLEMS AMONG CHILDREN.
Wang et al. A deep learning predictive classifier for autism screening and diagnosis
Mohammed et al. A novel model for classification of Parkinson’s disease: accurately identifying patients for surgical therapy
Ashik et al. Predictive Healthcare Analytics: Mental Health, Vaccination Effects, and Patient Satisfaction
Busari et al. Improving PTSD Screening Tools with CatBoost: A Bias-Aware Machine Learning Approach
Choudhary et al. Early Phase Detection of Diabetes Mellitus Using Machine Learning
Abbas et al. Autism spectrum disorder detection in toddlers and adults using deep learning
Gaddam et al. Predictive Analytics in the Healthcare Industry: Machine Learning for Early Diagnosis
US20240339184A1 (en) Systems, methods, and apparatus to provide longitudinal patient data

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: 21810998

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 202305959

Country of ref document: GB

Kind code of ref document: A

Free format text: PCT FILING DATE = 20211111

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21810998

Country of ref document: EP

Kind code of ref document: A1