[go: up one dir, main page]

WO2025016490A1 - 心电图异常监测方法、装置、设备和存储介质 - Google Patents

心电图异常监测方法、装置、设备和存储介质 Download PDF

Info

Publication number
WO2025016490A1
WO2025016490A1 PCT/CN2024/119209 CN2024119209W WO2025016490A1 WO 2025016490 A1 WO2025016490 A1 WO 2025016490A1 CN 2024119209 W CN2024119209 W CN 2024119209W WO 2025016490 A1 WO2025016490 A1 WO 2025016490A1
Authority
WO
WIPO (PCT)
Prior art keywords
ecg
data
ppg
abnormality monitoring
user
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/CN2024/119209
Other languages
English (en)
French (fr)
Inventor
刘张代红
克利夫顿大卫
朱婷婷
卢磊
张元亭
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.)
Oxford University Suzhou Science & Technology Co Ltd
Original Assignee
Oxford University Suzhou Science & Technology Co Ltd
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 Oxford University Suzhou Science & Technology Co Ltd filed Critical Oxford University Suzhou Science & Technology Co Ltd
Publication of WO2025016490A1 publication Critical patent/WO2025016490A1/zh
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
    • A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316—Modalities, i.e. specific diagnostic methods
    • A61B5/318—Heart-related electrical modalities, e.g. electrocardiography [ECG]
    • A61B5/346—Analysis of electrocardiograms
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/02—Detecting, measuring or recording for evaluating the cardiovascular system, e.g. pulse, heart rate, blood pressure or blood flow
    • 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
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/74—Details of notification to user or communication with user or patient; User input means
    • A61B5/746—Alarms related to a physiological condition, e.g. details of setting alarm thresholds or avoiding false alarms

Definitions

  • the present application relates to the technical field of electrocardiogram monitoring, and in particular to an electrocardiogram abnormality monitoring method, device, equipment and storage medium.
  • Electrocardiogram is a widely used medical examination method to monitor/classify cardiac abnormalities (such as myocardial infarction, ventricular hypertrophy, heart failure, etc.).
  • the 12-lead ECG is a standard method for obtaining cardiac function information in primary care. It is measured by placing 10 electrodes on the skin surface of the chest and limbs to record the electrical activity of the heart.
  • the ECG waveform consists of a QRS complex representing ventricular depolarization, a P wave representing atrial depolarization, and a T wave representing ventricular repolarization. Such a waveform is an informative and reliable measurement method for reflecting cardiac function, and is therefore widely used in clinical practice for the diagnosis of heart diseases.
  • 12-lead ECG devices can measure ECG data more completely, but they are usually bulky and equipped with electrodes, a central unit, and accessories such as a display and keyboard. It is difficult for ordinary users to conveniently use 12-lead ECG devices to perform daily continuous monitoring of their own ECG.
  • portable ECG devices such as smart watches and fitness trackers are smaller in size but can only measure one ECG lead, making the measured ECG data incomplete.
  • portable ECG devices require some user-initiated actions, such as holding the probe to close the conductive circuit, in order to take measurements, so the user needs to choose the right time to actively initiate the ECG measurement.
  • the paroxysmal nature of cardiac abnormalities such as arrhythmias makes the timing of ECG measurements a challenge even for clinical experts. Therefore, It is also difficult to continuously collect and monitor ECG using portable ECG equipment.
  • a method, apparatus, device and storage medium for monitoring abnormal electrocardiogram (ECG) are provided.
  • a method for monitoring ECG abnormality comprising: acquiring photoplethysmography (PPG) data of a user; inputting the PPG data into a pre-trained ECG abnormality monitoring model to obtain an output value of the ECG abnormality monitoring model; wherein the ECG abnormality monitoring model is pre-trained to output an output value indicating whether ECG data corresponding to the input PPG data is estimated to have abnormal heart function based on the input PPG data; when the output value of the ECG abnormality monitoring model indicates that the ECG data is estimated to have abnormal heart function, causing a user device associated with the user to sound an alarm.
  • PPG photoplethysmography
  • an ECG abnormality monitoring device comprising: a data acquisition module for acquiring PPG data of a user; a model monitoring module for inputting the PPG data into a pre-trained ECG abnormality monitoring model to obtain an output value of the ECG abnormality monitoring model; wherein the ECG abnormality monitoring model is used to output an output value indicating whether the ECG data corresponding to the input PPG data is estimated to have abnormal heart function based on the input PPG data; and an abnormal alarm module for causing a user device associated with the user to issue an alarm when the output value of the ECG abnormality monitoring model indicates that the ECG data is estimated to have abnormal heart function.
  • an ECG abnormality monitoring device including a PPG data detector, an output device, a memory and a processor; the PPG data detector is used to monitor the PPG data of a user and transmit the measured PPG data to the processor; the memory stores a computer program; and when the processor executes the computer program, it is used to receive PPG data from the PPG data detector and implement the above method, so that when the output value of the ECG abnormality monitoring model indicates that the user's ECG data is estimated to have abnormal heart function, the output device is used to cause a user device associated with the user to issue an alarm.
  • a computer-readable storage medium on which a computer program is stored, characterized in that the computer program implements the above method when executed by a processor.
  • FIG1 is a diagram showing an application environment of an ECG abnormality monitoring method according to an embodiment
  • FIG2 is a schematic diagram of a flow chart of an ECG abnormality monitoring method in one embodiment
  • FIG3 is an example of a PPG data segment retained by preprocessing in one embodiment
  • FIG4 is an example of a data segment that is not filtered out by preprocessing in one embodiment
  • FIG5 is a schematic diagram of the architecture of an ECG abnormality monitoring model in one embodiment
  • FIG6 is a schematic diagram of the architecture of a residual-dual convolutional attention block in one embodiment
  • FIG7 is a schematic diagram of the architecture of a dual convolutional attention block in one embodiment
  • FIG8 is a schematic diagram of the architecture of a convolutional attention block in one embodiment
  • FIG9 is an example of ECG raw data segments of leads II, V and AVR and corresponding PPG raw data segments in one embodiment
  • FIG. 10 is an example of an ECG pre-processed data segment marked as normal and its corresponding PPG pre-processed data segment in one embodiment
  • FIG. 11 is an example of an ECG pre-processed data segment marked as abnormal and its corresponding PPG pre-processed data segment in one embodiment
  • FIG12 is a schematic diagram of the structure of an ECG abnormality monitoring device in one embodiment
  • FIG. 13 is a schematic diagram of the structure of an ECG abnormality monitoring device in one embodiment.
  • 12-lead ECG is the standard method used in clinical practice to monitor abnormal cardiac function.
  • existing ECG measurement devices are inconvenient to use.
  • 12-lead ECG devices are bulky and inconvenient for users to monitor daily, and portable ECG devices can usually only measure one ECG lead, and usually require users to "actively" sample ECG through the device, making it difficult to grasp the appropriate measurement time.
  • PPG is an optical technology that measures the cardiac cycle by detecting changes in blood volume.
  • PPG can be measured "passively" by, for example, a pulse oximeter, and is often embedded as a feature in wearable devices such as fitness bands and smart watches.
  • the PPG waveform consists of a systolic wave and a diastolic wave, which is simpler and smoother than the ECG waveform.
  • PPG can be conveniently used for long-term continuous monitoring of physiological parameters such as pulse and respiratory rate.
  • it is very difficult to classify or monitor cardiac abnormalities based on PPG because it is a peripheral measurement signal that indirectly monitors the operation of the heart, and its relatively smooth morphology reflects less cardiac information compared to ECG. Therefore, existing PPGs are mostly used to measure physiological parameters such as pulse rate, respiratory rate, and blood pressure.
  • This application proposes an ECG abnormality monitoring method, which uses the continuously monitored and ubiquitous PPG signal to monitor possible ECG abnormalities, including but not limited to common cardiac function abnormalities such as atrioventricular block, sinus rhythm abnormality, bundle branch block, atrial fibrillation, premature beats, etc., so as to prompt the user to perform an ECG examination to learn about cardiac risks.
  • common cardiac function abnormalities such as atrioventricular block, sinus rhythm abnormality, bundle branch block, atrial fibrillation, premature beats, etc.
  • the ECG abnormality monitoring method provided in the present application can be applied in the application environment as shown in FIG. 1 .
  • the terminal 102 communicates with the server 104 via a network.
  • the terminal 102 may be a terminal device capable of measuring PPG data.
  • the terminal 102 may be provided with a PPG data detector, such as a pulse oximeter.
  • the terminal 102 may be a portable wearable terminal device such as a smart watch, a smart bracelet, or a smart headset with a built-in PPG data detector.
  • the terminal 102 may also have other forms, such as a medical monitor equipped with a pulse oximeter, or a system consisting of a smart bracelet with a built-in pulse oximeter and a smart phone that wirelessly communicates with the smart bracelet, etc.
  • a medical monitor equipped with a pulse oximeter or a system consisting of a smart bracelet with a built-in pulse oximeter and a smart phone that wirelessly communicates with the smart bracelet, etc.
  • the server 104 may be implemented as an independent server or a server cluster consisting of multiple servers.
  • the ECG abnormality monitoring method of the present application may include a training phase and a testing phase.
  • the ECG abnormality monitoring model may be trained using a training data set to obtain a trained ECG abnormality monitoring model, the abnormalities including but not limited to common ECG lead abnormalities such as atrioventricular block, sinus rhythm abnormality, bundle branch block, atrial fibrillation, premature beats, etc.
  • the training process may be executed on the server 104, and after completing the training to obtain the trained ECG abnormality monitoring model, the trained ECG abnormality monitoring model may be deployed from the server 104 to the terminal 102 for use.
  • the terminal 102 may monitor the PPG data of the user using its PPG detector, and execute the ECG abnormality monitoring method of the embodiment of the present application, so that when the output value of the ECG abnormality monitoring model indicates that the user's ECG data is estimated to be abnormal, the user device associated with the user issues an alarm, for example, the terminal 102 itself issues an alarm and/or other associated devices outside the terminal 102 issue an alarm.
  • the trained ECG abnormality monitoring model can also be deployed on the server 104.
  • the terminal 102 sends the collected PPG data to the server 104, and the server 104 uses the model to monitor and sends the monitoring results back to the terminal 102, etc. This application does not limit this.
  • the above alarm can be used to indicate that an ECG measurement is further performed on the user.
  • the ECG measurement can be performed by the terminal 102.
  • an ECG measurement device such as an ECG electrode or an ECG sensor, can be deployed on the terminal 102, so as to perform an ECG measurement on the user using the terminal 102.
  • the ECG measurement can also be measured by any other ECG measurement device outside the terminal 102.
  • a method for monitoring ECG abnormality is provided.
  • the method includes the following steps S210-S230:
  • PPG recording is a non-invasive detection method that uses photoelectric means to detect blood volume changes in living tissue.
  • PPG can be detected by a PPG data detector, which is any detector device that can detect the user's PPG in a non-invasive manner. It can include a light source and a photodetector. The source emits light of a predetermined wavelength to the measured tissue, the photodetector receives the light reflected or transmitted by the measured tissue, and obtains a detected PPG signal based on the intensity of the light reflected or transmitted by the measured tissue.
  • the PPG data detector can be, for example, a pulse oximeter.
  • the pulse oximeter may be built into the terminal 102, which may be a smart watch.
  • the pulse oximeter located at the bottom of the smart watch dial (close to the user's skin) detects the PPG data on the inside of the user's wrist.
  • the terminal 102 may also have other forms, and the test site of the PPG data is not limited to the inside of the user's wrist, but may be other sites such as the inside of the lower limb ankle, the chest, etc. It can be understood that the test site of the test data collected in the test phase is the same as the test site of the training data used in the training phase.
  • PPG data is data that characterizes the change in the amplitude of the user's pulse wave over time, and can be represented as one-dimensional data, that is, a PPG one-dimensional data sequence obtained by arranging the sampled values of the pulse wave in the order of sampling time.
  • the acquired PPG data of the user is the PPG data of the user monitored in real time by a portable terminal device.
  • the terminal 102 continuously samples and monitors the test part of the user under test at a predetermined sampling frequency.
  • the terminal 102 intercepts a PPG data segment of a predetermined duration from the PPG signal it continuously monitors as the PPG data to be input into the model.
  • the predetermined sampling frequency and the predetermined duration can be set according to actual needs.
  • the predetermined sampling frequency can be 125Hz
  • the predetermined duration can be 10s.
  • a PPG data segment of 10s in length can include approximately 1250 sampling values.
  • the data represented by the dark gray curve in FIG3 shows an example of a PPG raw data segment of 10s in length, wherein the horizontal axis represents time, which is represented by the sampling point sequence number, and the vertical axis represents the amplitude.
  • the horizontal and vertical axes in the ECG and PPG data segments shown in the remaining figures of the present application also have the same meaning.
  • the collected raw PPG data may have problems such as noise or low quality, which may affect the accuracy of the monitoring results. Therefore, the collected raw data can be preprocessed and then the preprocessed data can be input into the model for monitoring.
  • the above step S210 may include: obtaining the user's PPG raw data, preprocessing the PPG raw data to obtain PPG preprocessed data.
  • the PPG preprocessed data is then used as the PPG data to be input into the model. Accordingly, in the subsequent step S220, the PPG The preprocessed data is input into a pre-trained ECG abnormality monitoring model.
  • preprocessing the PPG raw data to obtain PPG preprocessed data may include one or more of the following steps:
  • this step it is detected whether the flatness of the signal reaches a predetermined standard. If so, the signal is retained; otherwise, the signal is discarded.
  • the signal is normalized to zero mean and unit variance.
  • each sampling value in the data segment retained by the above-mentioned flatness detection and signal screening steps can be normalized to zero mean value and unit variance to obtain a normalized data segment.
  • the signal is filtered.
  • a third-order bandpass Butterworth filter may be used to filter the normalized data segments in the above step to obtain filtered data segments, wherein the low-frequency band cutoff of the filter is 0.5 Hz and the high-frequency band cutoff is 8 Hz.
  • this step it is detected whether the number of effective peaks per unit length of the signal reaches a predetermined standard. If so, the signal is retained; otherwise, the signal is discarded.
  • the Python toolbox "HeartpPy” can be used to detect the number of valid peaks in the data segment after filtering in the above step. If the number of valid peaks in the 10s data segment is less than 5 (corresponding to less than 30bpm), the data segment is discarded. Otherwise, the data segment is retained.
  • this step it is detected whether the skewness of the signal reaches a predetermined standard. If so, the signal is retained, otherwise, the signal is discarded.
  • the skewness is calculated in a sliding window manner, with a window width of 250 sampling values (two seconds) and a step of 125 sampling values (one second).
  • the skewness of the sampling values in each of the nine time windows of 0-2s, 1-3s, 2-4s, 3-5s, 4-6s, 5-7s, 6-8s, 7-9s, and 8-10s can be calculated. If most of the nine skewnesses calculated (more than 50%) are negative, the data segment is considered to be a low-quality signal and is discarded. Otherwise, the data segment is retained.
  • outlier sampling values in the signal are detected, and the detected outlier sampling values are replaced with the median of the sampling values to obtain a signal after the outlier sampling values are eliminated.
  • a Hampel filter can be applied to detect outliers.
  • the median absolute deviation (MAD) is calculated for every 10 consecutive sample values.
  • the standard deviation (std) of the data segment is then estimated based on the MAD value assuming a normal distribution. If the sample value differs from the median of the data segment by 3std, the sample value is detected as an outlier.
  • the detected outliers are replaced with the median of the data segment. In this way, more outliers at the beginning and end of the data segment can be replaced.
  • Step 1) can be regarded as a quality screening step.
  • Steps 2) and 3) are cleaning and denoising, and steps 4) to 6) are quality control steps.
  • FIG3 shows a PPG data segment retained by the above-mentioned preprocessing steps 1) to 6), wherein the dark gray curve represents the original data segment, and the light gray curve represents the preprocessed data segment obtained after preprocessing.
  • FIG4 shows an example of data segments that have been screened out due to failure to pass different quality control standards in the above-mentioned preprocessing steps.
  • FIG4 represents a data segment that has been screened out due to unsatisfactory flatness
  • (b) represents a data segment that has been screened out due to unsatisfactory number of effective peaks
  • (c) represents a data segment that has been screened out due to unsatisfactory skewness.
  • the preprocessed data segment obtained after the above preprocessing steps can be used for monitoring in the subsequent step input model; if any raw data segment is discarded in any of the above steps, the raw data segment will no longer be used for monitoring. That is, only the preprocessed data segments that meet the requirements after being selected by the preprocessing steps are used for monitoring, thereby ensuring the accuracy of monitoring.
  • the preprocessing operations performed on the raw data in the testing phase can be the same as those performed in the training phase.
  • the preprocessing operations performed on the original data in each stage are the same to ensure the applicability of the model.
  • the ECG abnormality monitoring model of the present application is trained using a training data set including a data pair including PPG data and a label indicating whether the ECG data corresponding to the PPG data is abnormal.
  • the architecture and training process of the ECG abnormality monitoring model will be described later.
  • the trained ECG anomaly monitoring model can be used to output an output value indicating whether the ECG data corresponding to the input PPG data is estimated to be abnormal based on the input PPG data.
  • the ECG abnormality monitoring model can output the monitoring category probability As the output value, the probability of the monitoring category By comparing it with the set threshold T, it can be determined whether the corresponding ECG data is estimated to be abnormal. indicates that the ECG data estimate is abnormal, and when indicates that there is no abnormality in the ECG data estimation.
  • the user device associated with the user in S230 may include the above-mentioned terminal 102, such as a portable terminal device worn by the user, to inform the user that he or she has an ECG abnormality risk and should perform ECG measurement in time.
  • the user device associated with the user in S230 may also be a monitoring terminal device of a guardian associated with the user, such as a terminal device used by the user's family member or caregiver, or a terminal device used by the user's family doctor or attending physician, or a terminal device or server of the monitoring agency to which the user belongs, etc., to inform the guardian that the user is at risk of ECG abnormality and that ECG measurement should be performed in a timely manner.
  • a monitoring terminal device of a guardian associated with the user such as a terminal device used by the user's family member or caregiver, or a terminal device used by the user's family doctor or attending physician, or a terminal device or server of the monitoring agency to which the user belongs, etc.
  • causing a user device associated with the user to issue an alarm includes one or more combinations of the following: displaying visual information indicating that the user's ECG data estimate is abnormal on a display screen of the user device, causing an indicator light of the user device to light up indicating that the user's ECG data estimate is abnormal, causing a vibrator of the user device to vibrate indicating that the user's ECG data estimate is abnormal Prompt, causing the sound output device of the user device to issue a sound prompt indicating that the user's ECG data is estimated to be abnormal.
  • the content that causes the user device associated with the user to issue an alarm includes, in addition to notifying the monitoring result that the ECG data is estimated to be abnormal, also suggestion information suggesting that the user actively perform ECG measurement.
  • the method may further include: when the output value of the ECG abnormality monitoring model indicates that the ECG data is estimated to be abnormal, making an alarm call to a predetermined alarm receiving device.
  • the alarm call may be initiated by an alarm initiating device, which may be any suitable terminal or server.
  • the alarm call may be initiated by a portable terminal device worn by the user.
  • the portable terminal device needs to be a terminal device with a call-making function, such as a smart watch or a smart bracelet and a mobile phone paired with the smart bracelet; or the alarm call may be initiated by, for example, the server 104 or other servers or terminal devices.
  • the alarm call can be automatically initiated by the terminal or server or can be initiated based on the user's instructions.
  • the terminal or server can provide the user with the option of initiating an alarm call via a portable terminal device worn by the user, and then decide whether to make an alarm call to the predetermined alarm receiving device based on the user's selection input for the option.
  • the terminal or server does not receive the user's selection input within a predetermined period of time, it may be considered that the user may be in a dangerous condition such as losing consciousness, and the alarm call will be automatically made to the predetermined alarm receiving device.
  • the predetermined alarm receiving device for receiving alarm calls can be pre-set as needed.
  • the predetermined alarm receiving device can include the monitoring terminal device of the guardian associated with the user, and/or the alarm terminal of a designated medical institution (for example, a hospital emergency telephone terminal that can be dialed by dialing 112), etc.
  • the alarm call can be a real-time call that allows the user to talk to the recipient at the predetermined alarm receiving device.
  • the alarm call can also include pre-set voice broadcast content, for example, it can include voice broadcast content that informs the user that the ECG data is estimated to have abnormal monitoring results.
  • it can also include personal information such as the user's name, gender, age, residence, and the user's real-time location, so that the alarm call recipient can quickly know the user's current condition and real-time location.
  • the method may further include: when the output value of the ECG abnormality monitoring model indicates that the ECG data is estimated to be abnormal, performing ECG measurement on the user to obtain a measurement result indicating whether the user's ECG is abnormal.
  • the ECG measurement may be performed by a terminal 102 provided with an ECG measurement device, or by other ECG measurement equipment different from the terminal 102, such as a 12-lead ECG provided in a medical institution, or a home ECG measurement equipment, etc.
  • the terminal 102 may determine whether to trigger the ECG measurement of the user by the terminal 102 based on the user input.
  • the terminal 102 may automatically trigger the ECG measurement of the user when the output value indicates that the ECG data is estimated to be abnormal.
  • the user/guardian may choose to perform ECG measurement of the user using other ECG measurement equipment other than the terminal 102.
  • the user may go to a medical institution to perform a more professional and comprehensive ECG measurement using a 12-lead ECG, and obtain a measurement result of whether the ECG is abnormal.
  • terminal 102 may analyze the measured ECG data by itself or via an external device such as a server, and push the measured ECG data and the measurement result of whether the ECG is abnormal to the user through terminal 102 .
  • ECG abnormality monitoring method since PPG data can be obtained by measuring the user in a passive manner through a portable wearable device such as a smart watch, can conveniently and continuously monitor the user by acquiring the user's PPG data.
  • an ECG abnormality monitoring model can be obtained, which can output an output value indicating whether the ECG data corresponding to the input PPG data is estimated to be abnormal based on the input PPG data.
  • the model can be used to monitor whether the corresponding ECG data is likely to be abnormal based on the measured PPG data, and when an abnormality is detected, the user device associated with the user sends an alarm to suggest the user to actively perform ECG measurement to obtain an accurate result of whether the ECG is abnormal.
  • step S220 of the testing phase a pre-trained ECG abnormality monitoring model is used, which requires that the model architecture of the ECG abnormality monitoring model is pre-constructed in the training phase before the testing phase, and The constructed ECG abnormality monitoring model is trained, and the model parameters of the ECG abnormality monitoring model are optimized to obtain a trained signal quality assessment model.
  • the present application proposes a dual-convolutional attention network (DCA-Net) as the above-mentioned ECG abnormality monitoring model.
  • the dual-convolutional attention network is characterized in that it includes one or more residual-dual convolutional attention (Res-DAC) blocks.
  • Each residual-dual convolutional attention block is formed by adding a dual-convolutional-attention (DCA) block to the residual block.
  • DCA dual-convolutional-attention
  • the dual-convolutional attention block includes a first convolutional attention block for convolving data in the channel dimension and a second convolutional attention block for convolving data in the time dimension.
  • the ECG abnormality monitoring model may include at least a plurality of residual-dual convolution attention blocks 1 to N arranged in series.
  • the ECG abnormality monitoring model may also include an input convolution layer and a pooling layer located before the multiple residual-double convolution attention blocks and arranged in series with the multiple residual-double convolution attention blocks, and an average pooling layer and a fully connected layer located after the multiple residual-double convolution attention blocks and arranged in series with the multiple residual-double convolution attention blocks.
  • the input convolution layer receives the PPG data input to the ECG abnormality monitoring model
  • the output of the input convolution layer is fed into the pooling layer
  • the output of the pooling layer is fed into the first residual-double convolution attention block in multiple residual-double convolution attention blocks
  • the output of the previous one in multiple residual-double convolution attention blocks is fed into the next one
  • the output of the last residual-double convolution attention block in multiple residual-double convolution attention blocks is fed into the average pooling layer
  • the output of the average pooling layer is fed into the fully connected layer
  • the fully connected layer outputs the output value of the ECG abnormality monitoring model.
  • the last layer of the fully connected layer may have a single neuron that outputs a 1*1-dimensional output value, i.e., the probability The probability By comparing with the set threshold T, the monitoring result of whether the ECG is estimated to be abnormal can be obtained.
  • each convolution layer in the model such as the input convolution layer, can be a one-dimensional convolution layer to meet the processing requirements of the PPG one-dimensional data sequence.
  • model architecture in FIG5 is only an example. Under the basic premise that the model includes one or more residual-double convolution attention blocks as proposed in the present application, those skilled in the art can also modify the model.
  • the architecture can be modified in many ways, such as the number and serial/parallel mode of residual-double convolution attention blocks can be changed, the number and position of convolutional layers and pooling layers in the model can also be changed.
  • each of the multiple residual-dual convolution attention blocks may include a first main path and a first short-circuit branch path.
  • One or more convolution layers and dual convolution attention blocks are arranged in series on the first main path.
  • the dual convolution attention block includes a first convolution attention block and a second convolution attention block arranged in series or in parallel, the first convolution attention block is used to perform convolution in the channel dimension, and the second convolution attention block is used to perform convolution in the time dimension.
  • the first short-circuit branch path is connected in parallel to one or more convolution layers arranged in series on the first main path and both ends of the dual convolution attention block.
  • the former of one or more convolutional layers and dual convolutional attention blocks and the first short-circuit branch path receive the output of the previous path
  • the latter of one or more convolutional layers and dual convolutional attention blocks receive the output of the former
  • the output of the last of one or more convolutional layers and dual convolutional attention blocks is added to the output matrix of the first short-circuit branch path and fed into the subsequent path.
  • each convolution layer in each residual-dual convolution attention block is also provided with a batch normalization layer and an activation function layer on the output side.
  • the batch normalization layer and the activation function layer on the output side of each convolution layer except the last convolution layer are arranged before the next convolution layer, and the batch normalization layer on the output side of the last convolution layer is arranged between the convolution layer and the dual convolution attention block, the dual convolution attention block is arranged before the addition operation with the output of the first short-circuit branch path, and the batch normalization layer arranged on the output side of the last convolution layer is arranged after the addition operation with the output of the first short-circuit branch path.
  • each residual-dual convolution attention block two convolution layers are provided in each residual-dual convolution attention block, and they are one-dimensional convolution layers.
  • the activation function layer is a ReLU activation function layer.
  • the number and dimension of the convolution layers in the residual-dual convolution attention block can be changed according to the situation, and the activation function layer can also be other types of activation function layers.
  • the first convolutional attention block and the second convolutional attention block are arranged in series, and the first convolutional attention block is located before the second convolutional attention block.
  • the first convolutional attention block in terms of channel is placed before the second convolutional attention block in terms of time, so that the dual convolutional attention block Learning the channel-to-channel interaction first and then learning the temporal interaction based on the channel participation output performs better than doing both in parallel or with the first convolutional attention block in the channel in the latter order.
  • the second convolutional attention block may be located before the first convolutional attention block or the first convolutional attention block and the second convolutional attention block may be set in parallel.
  • each of the first convolutional attention block and the second convolutional attention block includes: a second main path and a second short-circuit branch path.
  • the second main path is provided with a maximum pooling layer and an average pooling layer, a convolution layer and an activation function layer, the maximum pooling layer and the average pooling layer are arranged in parallel, and the parallel maximum pooling layer and the average pooling layer are arranged in series with the convolution layer and the activation function layer in turn.
  • the second short-circuit branch path is connected in parallel to both ends of the maximum pooling layer and the average pooling layer, the convolution layer and the activation function layer arranged on the second main path.
  • the second short-circuit branch path, the maximum pooling layer and the average pooling layer respectively receive the output of the previous path, the output of the maximum pooling layer and the output of the average pooling layer are concatenated and fed into the convolution layer, the output of the convolution layer is fed into the activation function layer, and the output of the activation function layer is multiplied by the output of the second short-circuit branch path and fed into the subsequent path.
  • the first/second convolutional attention block has a convolutional layer, and it is a one-dimensional convolutional layer.
  • the activation function layer is a Sigmoid activation function layer.
  • the number and dimension of the convolutional layers in the first/second convolutional attention block can be changed according to the situation, and the activation function layer can also be other types of activation function layers.
  • the convolutional attention block relies on applying a convolution operation to a certain dimension of the data to learn the interactions between data points along that dimension.
  • the convolutional attention is more computationally efficient because no parameters are required in the convolution layer.
  • the first convolutional attention block and the second convolutional attention block have similar architectures, but they are applied in different directions.
  • the first convolutional attention block is applied to the convolution channel direction, and the second convolutional attention block is applied to the time domain direction.
  • the specific configurations of the parameters of each layer/block such as activation function, kernel size, step size, padding, etc., can be set according to actual needs.
  • the convolution layer in FIG8 is used as the convolution kernel Taking a one-dimensional convolutional layer of size 7 as an example, let the input of the first convolutional attention block in terms of channel (which is also the input of the DCA module) be Where N is the sample point/batch size, C is the number of convolution channels output from the original residual block, and D is the signal length (remembered as the time dimension).
  • First Convolutional Attention Block The output is:
  • Equation (1) represents the Hadamard product (element-wise matrix multiplication); ⁇ ( ⁇ ) represents the Sigmoid activation function, denotes a 1D convolutional layer with kernel size 7 and convolution in the channel direction.
  • ⁇ ] D in Equation (2) denotes the matrix union along the time direction, and the subscript D in Equations (3) and (4) indicates that the pooling layer is also applied in the time dimension.
  • Function Output channel-wise attention weights which are further extended to X of the same shape to scale X. Then the scaled X (denoted as ) is used as input to the temporal attention.
  • Equations (7)-(8) indicate that the maximum pooling and average pooling are first performed on the channel dimension, so and Then, the pooled matrix is combined along the channel dimension, (Equation (6)).
  • the one-dimensional convolutional layer convolves in the time dimension and outputs a single channel, so that Finally, this convolved matrix is activated by the Sigmoid activation function and used to scale the input time domain.
  • the two-dimensional convolutional layers in the known ResNet-34 model are replaced with one-dimensional convolutional layers
  • the last fully connected layer in the ResNet-34 model is replaced with a fully connected layer with a 512-dimensional input and a 1-dimensional output
  • a residual is added before each residual block of the ResNet-34 model.
  • Double convolution attention block and obtain a specific example of DCA-Net model. It can be understood that the DCA-Net model of this specific example will include a serial input convolution layer, a pooling layer, 16 residual-double convolution attention blocks, an average pooling layer and a fully connected layer.
  • the convolution kernel size of the first convolution attention block and the second convolution attention block can both be 7
  • the step size is 1
  • the average pooling layer can be a one-dimensional adaptive average pooling layer (AdaptiveAvgPool1d)
  • the parameters of other layers and blocks in the model can all refer to the parameter settings of the known ResNet-34 model.
  • the forward propagation process of the ResNet-34 model of this specific example is explained.
  • the batch size is 64
  • 64 ⁇ 64 ⁇ 625 data is output
  • 64 ⁇ 64 ⁇ 313 data is output.
  • the data is input into the first residual-double convolutional attention block.
  • batch normalization layer, ReLU activation function layer, convolution layer, batch normalization layer shown in Figure 6 in sequence it is input into the double convolutional attention layer.
  • the input data first enters the first convolutional attention block in the channel.
  • the average pooling layer performs average pooling on the input data in the time dimension and outputs a 64 ⁇ 64 ⁇ 1 matrix.
  • the maximum pooling layer performs maximum pooling on the input data in the time dimension and outputs another 64 ⁇ 64 ⁇ 1 matrix.
  • the two matrices are combined along the time dimension (similar to the splicing operation) to obtain a 64 ⁇ 64 ⁇ 2 matrix.
  • the 64 ⁇ 64 ⁇ 2 matrix is convolved along the channel dimension through the convolution layer to learn the 64 ⁇ 64 ⁇ 1 attention weight, and then replicated and expanded along the time dimension into a 64 ⁇ 64 ⁇ 313 attention score.
  • This attention score is Hadamard-multiplied with the original input data output through the second short-circuit branch path to output a 64 ⁇ 64 ⁇ 313 matrix.
  • the 64 ⁇ 64 ⁇ 313 matrix output by the first convolutional attention block is input into the second convolutional attention block in time.
  • the average pooling layer performs average pooling on the 64 ⁇ 64 ⁇ 313 matrix in the channel dimension and outputs a 64 ⁇ 1 ⁇ 313 matrix.
  • the maximum pooling layer performs maximum pooling on the input data in the channel dimension and outputs another 64 ⁇ 1 ⁇ 313 matrix.
  • the 64 ⁇ 2 ⁇ 313 matrix is convolved along the time dimension through the convolution layer to obtain a 64 ⁇ 1 ⁇ 313 attention matrix.
  • the force weight is then replicated and expanded along the channel dimension into a 64 ⁇ 64 ⁇ 313 attention score.
  • This attention score is Hadamard-multiplied with the original input 64 ⁇ 64 ⁇ 313 matrix output by the second short-circuit branch path, and a 64 ⁇ 64 ⁇ 313 matrix is output.
  • This 64 ⁇ 64 ⁇ 313 matrix is then input into the next residual-double convolution attention block as the output of the second convolution attention block.
  • the output data is then passed through the average pooling layer and the fully connected layer to output the output value of the model.
  • a trained ECG abnormality monitoring model can be determined.
  • the ECG abnormality monitoring model with the above structure can be trained using various conventional or novel training methods.
  • the training data set may include multiple sets of PPG data-data pairs of labels for marking whether the corresponding ECG data is abnormal.
  • the PPG data included in the training data set and the corresponding ECG data required for marking the labels may be collected from multiple cases corresponding to multiple subjects.
  • the number of subjects from which the data used to construct the training data set is derived may be set as needed, for example, but not limited to hundreds, thousands or tens of thousands, so as to obtain data that can statistically reflect the characteristics of most users.
  • one or more PPG data and corresponding ECG data of the subject may be collected, and then each ECG data may be marked with a label for abnormal cardiac function by, for example, a doctor manually or machine-assisted, and the label may be matched to the PPG data corresponding to the ECG data, so that one or more sets of PPG data-data pairs of labels for marking whether the corresponding ECG data is abnormal may be obtained.
  • the data pairs of multiple sets of PPG data-labels for marking whether the corresponding ECG data is abnormal collected from the above multiple subjects may be collected to obtain a training data set including multiple subject data that can be used as an example.
  • the above-mentioned multiple subjects may include at least some patients with abnormal heart function diseases, so that at least some PPG data and corresponding ECG data with abnormal heart function characteristics can be collected.
  • the above-mentioned multiple subjects may also include some subjects with normal heart function, or although only patients with abnormal heart function diseases are included, at least some PPG data and corresponding ECG data with normal heart function characteristics are collected (for example, normal data collected when the patient is not ill).
  • the PPG data and corresponding ECG data used to construct the training data set can be collected from the clinical medical records of a large number of different subjects in one or more hospitals during hospitalization.
  • the subjects from which the training data used in the training phase originates do not need to be or include the subjects tested in the testing phase.
  • a large amount of clinical data from a medical system can be used to form a training data set to train the model, and then the trained model can be deployed on the terminal 102. After that, any user can use the terminal 102 to perform the above-mentioned testing phase process to monitor their own electrocardiogram abnormalities.
  • the process of constructing a training data set may include the following steps:
  • the correspondence between the ECG signal and the PPG signal means that the ECG signal and the PPG signal are measured simultaneously on the same subject to ensure that the ECG signal and the PPG signal are time-aligned and correlated with each other.
  • ECG signals of multiple leads and their corresponding PPG signals can be obtained from a matching subset of the MIMIC-III Waveform Database (WDB).
  • WDB MIMIC-III Waveform Database
  • the MIMIC-III WDB is linked to clinical records - a large number of patients can be included in the matching subset, and all patients in the matching subset are identified and matched to the medical record system.
  • the MIMIC-III WDB includes ECG signals of multiple leads and their paired PPG signals. For each patient, waveform data of different lengths are collected at different time points during their hospitalization. The sampling frequency of these signals is 125 Hz. Since different ECG records have different numbers and types of leads, in this example, data of the three most common and representative ECG leads are selected, namely, leads II, V, and AVR.
  • ECG signals and corresponding PPG signal records using these three leads are included in this analysis.
  • more or fewer ECG leads may be selected, and the type of the selected leads may also be changed.
  • only records of signals with a duration exceeding a predetermined duration, such as two minutes, are selected. Signals shorter than two minutes will be filtered out.
  • the predetermined length of the data segment to be intercepted during the training process is the same as the predetermined length of the data segment intercepted during the test process, for example, 10 seconds.
  • the second 10-second data segment can be intercepted from each ECG/PPG signal to avoid instability of the sampling value in the initial period.
  • An example of the resulting 10-second long ECG raw data segments of leads II, V and AVR and the corresponding 10-second long PPG raw data segments are shown in FIG. 9 .
  • the same preprocessing process including steps 1) to 6) as in the test process described above can be performed to screen out the PPG preprocessed data segment that meets the requirements.
  • a 5th-order high-pass Butterworth filter of 0.5Hz can be used, followed by power line filtering with a power line frequency of 50Hz to obtain 10s of ECG pre-processed data segments of leads II, V and AVR. Then, a predetermined signal quality assessment method is used to assess the signal quality of each ECG pre-processed data segment. The assessment outputs one of three quality categories - unacceptable, almost unacceptable or excellent.
  • the ECG pre-processed data segment of a certain lead is assessed as excellent, the ECG pre-processed data segment of the lead is retained, otherwise, when the ECG pre-processed data segment of a certain lead is assessed as unacceptable or almost unacceptable, the ECG pre-processed data segment of the lead is discarded.
  • any group of 10s ECG raw data segments of multiple leads and their corresponding PPG raw data segments if any one of the ECG raw data segments of the leads cannot be retained through the above preprocessing process, or the PPG raw data segments cannot be retained through the above preprocessing process, the group of data segments is discarded as a whole. Only when each of the ECG raw data segments of multiple leads in a group and their corresponding PPG raw data segments are retained through the above preprocessing process, can the group of ECG raw data segments of multiple leads and their corresponding PPG raw data segments of the predetermined time length that meet the requirements be obtained.
  • FIG. 10 shows an example of an ECG preprocessed data segment marked as normal and its corresponding PPG preprocessed data segment
  • Figure 11 shows an example of an ECG preprocessed data segment marked as abnormal and its corresponding PPG preprocessed data segment.
  • professional technicians such as clinicians, can also manually label whether the ECG preprocessed data segment is abnormal to ensure the accuracy of the label.
  • multiple sets of data pairs such as PPG preprocessed data segments-labels of whether the corresponding preprocessed data segments ECG are abnormal can be obtained to constitute the required training data set.
  • the ECG abnormality monitoring model can be trained using the training data set to determine a trained ECG abnormality monitoring model.
  • the training process of the ECG abnormality monitoring model includes:
  • the training data set includes multiple sets of PPG data-labeled data pairs indicating whether the corresponding ECG data is abnormal;
  • the training set is used to train the ECG abnormality monitoring model to obtain the fitted model parameters.
  • the validation set is used to verify the ECG abnormality monitoring model determined by the fitted model parameters and adjust the model hyperparameters.
  • the test set is used to evaluate the performance of the trained ECG abnormality monitoring model determined by the adjusted hyperparameters and the trained model parameters.
  • the DCA-Net model of the above-mentioned specific example can be applied to the PPG data segments processed based on MIMIC-III WDB, and the corresponding ECG data segments are marked with a label for binary classification to be trained.
  • the 38,320 sets of PPG data segments-labels of whether the corresponding ECG data segments are abnormal are divided into training sets, validation sets, and test sets, with a ratio of 64%: 16%: 20%.
  • the model parameters are learned using the training set and verified in the validation set to avoid overfitting.
  • the test set is only used to report the results.
  • the model performance was evaluated using specificity, sensitivity/recall, precision, accuracy, AUROC (area under the receiver operating characteristic curve) and AUPRC (area under the precision-recall curve).
  • the model was further randomly repeated 10 times, and the average and standard deviation of the above evaluation indicators were used.
  • the DCA-Net model can be implemented using PyTorch version 1.11.0 and CUDA version 11.7.
  • the loss function is binary cross entropy. Further experiments adjust for class imbalance by increasing the loss weight.
  • the model is optimized by the Adam optimizer with default hyperparameter settings and trained using a batch size of 64.
  • a learning rate scheduler in terms of stride is applied to reduce the learning rate by a factor of 0.1 every 10 epochs. Early stopping is also applied - the training is terminated if the validation loss does not decrease compared to the current lowest loss for 5 consecutive epochs.
  • the model with the lowest loss is always saved and applied to the test set for evaluation.
  • 2 may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.
  • the DCA-Net model of the above specific example of this application and the existing neural network models ResNet-18, ResNet-34, ResNet-50, and ECA-Net are also used to monitor ECG abnormalities based on PPG data for performance testing and comparison.
  • the comparison results are shown in Table 1 below.
  • the numbers shown are the average score percentages of 10 random repetitions, and their standard deviations are shown in brackets.
  • the best average accuracy, AUROC, and AUPRC are highlighted in bold.
  • the DCA-Net proposed in this application has the best average accuracy, AUROC and AUPRC scores.
  • the performance of DCA-Net is more stable than that of ECA-Net.
  • an ECG abnormality monitoring device 1200 comprising:
  • the data acquisition module 1210 is used to acquire the PPG data of the user
  • the model monitoring module 1220 is used to input the PPG data into a pre-trained ECG abnormality monitoring model to obtain an output value of the ECG abnormality monitoring model; wherein the ECG abnormality monitoring model is trained using a training data set including a data pair of PPG data and a label indicating whether the ECG data corresponding to the PPG data is abnormal, and the ECG abnormality monitoring model is used to output an output value indicating whether the ECG data corresponding to the input PPG data is estimated to be abnormal based on the input PPG data;
  • the abnormality alarm module 1230 is configured to cause a user device associated with a user to issue an alarm when the output value of the ECG abnormality monitoring model indicates that an abnormality exists in the ECG data estimation.
  • the ECG abnormality monitoring device 1200 also includes a model training module 1240, and the model monitoring module 1220 is also used to obtain a trained ECG abnormality monitoring model from the model training module 1240.
  • the model training module 1240 is used to train the ECG abnormality monitoring model to obtain a trained ECG abnormality monitoring model.
  • the data acquisition module 1210 , the model monitoring module 1220 , and the abnormal alarm module 1230 may be arranged in the terminal 102 , while the model training module 1300 may be arranged in the server 104 .
  • the specific definition of the ECG abnormality monitoring device 1200 can refer to the definition of the ECG abnormality monitoring method above, which will not be repeated here. All or part of the above modules can be implemented by software, hardware and their combination. The above modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.
  • an ECG abnormality monitoring device 1300 including a PPG data detector 1310 , a memory 1320 , and a processor 1330 ;
  • the PPG data detector 1310 is used to monitor the PPG data of the user and transmit the measured PPG data to the processor;
  • the memory 1320 stores a computer program
  • processor 1330 executes the computer program, it is used to receive PPG data from the PPG data detector 1310 and implement the method of any of the above embodiments.
  • the PPG data detector may be a pulse oximeter.
  • the ECG abnormality monitoring device 1300 further includes an output device 1340, which may include a display screen, an indicator light, a vibrator, and/or a sound output device.
  • the display screen may be a liquid crystal display screen or an electronic ink display screen, etc.
  • the sound output device may be a device with a sound output function such as a speaker or a buzzer. In this way, when the ECG abnormality monitoring device 1300 itself is used to issue an alarm, the alarm may be issued through the output device 1340.
  • the ECG abnormality monitoring device 1300 also has a phone-making function.
  • the ECG abnormality monitoring device 1300 may also include a dialing device with a phone-making function.
  • the dialing device may be implemented, for example, by a smart watch, mobile phone or server associated with the user, so that when an alarm call needs to be made to a predetermined alarm receiving device, the alarm call can be made through the dialing device.
  • the ECG abnormality monitoring device 1300 also has an ECG measurement function.
  • the ECG abnormality monitoring device 1300 may also include an ECG measurement device to further perform ECG measurement on the user using the ECG measurement device when the ECG abnormality is estimated using PPG data.
  • the ECG abnormality monitoring device 1300 can be any device that can execute the method of the above-mentioned embodiment of the present application, and it may have various different forms.
  • it may include the aforementioned terminal 102, or it may also include the aforementioned terminal 102 and the server 104.
  • a computer readable storage medium on which a computer program is stored.
  • the computer program is characterized in that when executed by a processor, the following steps are implemented:
  • the ECG abnormality monitoring model is trained using a training data set including a data pair of PPG data and a label indicating whether the ECG data corresponding to the PPG data is abnormal, and the ECG abnormality monitoring model is used to output an output value indicating whether the ECG data corresponding to the input PPG data is estimated to be abnormal based on the input PPG data;
  • Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory.
  • Volatile memory can include random access memory (RAM) or external cache memory.
  • RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

Landscapes

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

Abstract

一种心电图异常监测方法、装置、计算机设备和存储介质。方法包括：（S210）获取用户的PPG数据；（S220）将PPG数据输入预先训练好的ECG异常监测模型中，得到ECG异常监测模型的输出值；其中，ECG异常监测模型被预先训练成用于基于输入的PPG数据，输出指示所输入的PPG数据对应的ECG数据是否估计存在心脏功能异常的输出值；（S230）当ECG异常监测模型的输出值指示用户的ECG数据估计存在心脏功能异常时，使与用户相关联的用户设备发出警报。

Description

心电图异常监测方法、装置、设备和存储介质
相关申请的交叉引用
本申请要求于2023年07月18日提交中国专利局、申请号为202310881402.6、发明名称为“心电图异常监测方法、装置、设备和存储介质”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及心电图监测技术领域，特别是涉及一种心电图异常监测方法、装置、设备和存储介质。
背景技术
心电图(Electrocardiogram，ECG)是一种广泛用于监测/分类心脏异常(如心肌梗死、心室肥厚、心力衰竭等)的医学检查手段。
12导联ECG是在基层医疗中获得心脏功能信息的标准方法，其是通过在胸部和四肢的皮肤表面放置10个电极来测量，以记录心脏的电活动。ECG波形由代表心室去极化的QRS波群、代表心房去极化的P波以及代表心室复极化的T波构成。这样的波形是反映心脏功能的信息性和可靠的测量方法，因此广泛用于心脏疾病诊断的临床实践中。12导联ECG设备能够较为完善地测量ECG数据，但是其通常体积庞大，配有电极、中央单元以及显示器和键盘等附件，普通用户难以方便地使用12导联ECG设备对自身ECG进行日常持续监测。
相比之下，智能手表和健身追踪器等便携式ECG设备体积较小，但只能测量一条ECG导联，使得测量的ECG数据不够完善。此外，它们还需要一些用户发起的动作，例如，握持探测器以闭合导电电路，以便进行测量，因此用户需要选择合适的时机来主动发起对ECG的测量。而心律失常等心脏异常具有阵发性性质，对ECG测量时机的选择即使对于临床专家而言也是一项挑战。因此， 用便携式ECG设备也难以持续地采集和监测ECG。
因此，上述现有的ECG测量方法仍然存在着改进的空间。
发明内容
根据本申请各种实施例，提供一种心电图(ECG)异常监测方法、装置、设备和存储介质。
在一方面，提供一种ECG异常监测方法，方法包括：获取用户的光电容积脉搏波(Photoplethysmography，PPG)数据；将PPG数据输入预先训练好的ECG异常监测模型中，以得到ECG异常监测模型的输出值；其中，ECG异常监测模型被预先训练成用于基于输入的PPG数据，输出指示所输入的PPG数据对应的ECG数据是否估计存在心脏功能异常的输出值；当ECG异常监测模型的输出值指示ECG数据估计存在心脏功能异常时，使与用户相关联的用户设备发出警报。
在另一方面，提供一种ECG异常监测装置，包括：数据获取模块，用于获取用户的PPG数据；模型监测模块，用于将PPG数据输入预先训练好的ECG异常监测模型中，以得到ECG异常监测模型的输出值；其中，ECG异常监测模型用于基于输入的PPG数据，输出指示所输入的PPG数据对应的ECG数据是否估计存在心脏功能异常的输出值；异常警报模块，用于当ECG异常监测模型的输出值指示ECG数据估计存在心脏功能异常时，使与用户相关联的用户设备发出警报。
在另一方面，提供一种ECG异常监测设备，包括PPG数据探测器、输出装置、存储器和处理器；PPG数据探测器用于监测用户的PPG数据，并将测得的PPG数据传输至处理器；存储器存储有计算机程序；并且处理器执行计算机程序时，用于从PPG数据探测器接收PPG数据，并实现如上的方法，以在当ECG异常监测模型的输出值指示用户的ECG数据估计存在心脏功能异常时，利用输出装置使与用户相关联的用户设备发出警报。
在另一方面，提供一种计算机可读存储介质，其上存储有计算机程序，其特征在于，计算机程序被处理器执行时实现如上的方法。
本发明的一个或多个实施例的细节在下面的附图和描述中提出。本发明的其它特征、目的和优点将从说明书、附图以及权利要求书变得明显。
附图说明
为了更好地描述和说明这里公开的那些发明的实施例和/或示例，可以参考一幅或多幅附图。用于描述附图的附加细节或示例不应当被认为是对所公开的发明、目前描述的实施例和/或示例以及目前理解的这些发明的最佳模式中的任何一者的范围的限制。
图1为一个实施例中ECG异常监测方法的应用环境图；
图2为一个实施例中ECG异常监测方法的流程示意图；
图3为一个实施例中通过预处理而被保留的PPG数据片段的示例；
图4为一个实施例中未通过预处理而被筛除的数据片段的示例；
图5为一个实施例中ECG异常监测模型的架构示意图；
图6为一个实施例中残差-双卷积注意力块的架构示意图；
图7为一个实施例中双卷积注意力块的架构示意图；
图8为一个实施例中卷积注意力块的架构示意图；
图9为一个实施例中导联II、V和AVR的ECG原始数据片段以及对应的PPG原始数据片段的示例；
图10为一个实施例中被标记为正常的ECG预处理数据片段及其对应的PPG预处理数据片段的示例；
图11为一个实施例中被标记为异常的ECG预处理数据片段及其对应的PPG预处理数据片段的示例；
图12为一个实施例中ECG异常监测装置的结构示意图；
图13为一个实施例中ECG异常监测设备的结构示意图。
具体实施方式
为了使本申请的目的、技术方案及优点更加清楚明白，以下结合附图及实 施例，对本申请进行进一步详细说明。应当理解，此处描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
由于心脏功能与ECG之间的密切关联，且ECG数据包含了较为丰富的信息内容，尤其是ECG的波形结构和蕴涵的时间信息，因此12导联ECG是临床中用来监测心脏功能异常的标准手段。然而，现有的ECG测量设备均存在着使用不便的缺陷，例如12导联ECG设备庞大，不便于用户日常监测，而便携式ECG设备通常仅能测量一条ECG导联，且通常需要由用户通过设备“主动地”采样ECG，难以把握合适的测量时机。
PPG是一种通过检测血容量变化来测量心脏周期的光学技术。PPG可以通过例如脉搏血氧计等来“被动地”测量，并且通常作为一种功能嵌入到可穿戴设备中，如健身腕带和智能手表。PPG的波形由一个收缩波和一个舒张波组成，比ECG的波形更简单、更平滑。PPG可方便地用于长期持续监测例如脉搏和呼吸率等生理参数。然而，根据PPG对心脏异常进行分类或监测非常困难，因为它是一种间接监测心脏的运作的外周测量信号，且与ECG相比，其相对平滑的形态反映的心脏信息较少。因此，现有的PPG多用于测量脉搏频率、呼吸频率、血压等生理参数。
本申请提出了一种ECG异常监测方法，其使用连续监测的、普遍存在的PPG信号，来监测ECG可能存在异常的情况，所述异常包括但不限于房室传导阻滞、窦性心律异常、束支传导阻滞、房颤、早搏等常见的心脏功能异常。以提示用户进行ECG检查以得知心脏风险。
本申请提供的ECG异常监测方法，可以应用于如图1所示的应用环境中。其中，终端102通过网络与服务器104进行通信。终端102可以是具有测量PPG数据的终端设备，例如终端102可以设置有PPG数据探测器，例如脉搏血氧计。示例地，终端102可以是例如内置有PPG数据探测器的智能手表、智能手环、智能耳机等便携式可穿戴终端设备。终端102也可以具有其他形态，例如配置有脉搏血氧计的医用监仪，又例如内置有脉搏血氧计的智能手环和与该智能手环无线通信的智能手机组成的系统等等，本领域技术人员将能想到许多其他变 形，本申请对此不作限定。服务器104可以用独立的服务器或者是多个服务器组成的服务器集群来实现。
本申请的ECG异常监测方法可以包括训练阶段和测试阶段。在训练阶段，可以利用训练数据集对ECG异常监测模型进行训练，以得到训练好的ECG异常监测模型，该异常包括但不限于房室传导阻滞、窦性心律异常、束支传导阻滞、房颤、早搏等常见的心电导联异常。该训练过程可以在服务器104上执行，并在完成训练得到训练好的ECG异常监测模型后，将训练好的ECG异常监测模型从服务器104部署到终端102上以备使用。在完成训练得到训练好的ECG异常监测模型后，在测试阶段，终端102可以利用其具有的PPG探测器监测用户的PPG数据，并执行本申请实施例的ECG异常监测方法，以当ECG异常监测模型的输出值指示用户的ECG数据估计存在异常时，使与该用户相关联的用户设备发出警报，例如使终端102自身发出警报和/或使终端102之外的其他关联设备发出警报。可以理解，上述ECG异常监测方法的部分和全部步骤在何处执行，可以在终端102和服务器104之间进行适当调整，例如，训练好的ECG异常监测模型也可以部署在服务器104上，终端102将采集的PPG数据发送至服务器104，由服务器104利用模型进行监测，并将监测结果发回给终端102，等等，本申请对此不作限定。
进一步地，上述警报可以用于指示对用户进一步进行ECG测量。该ECG测量可以由终端102进行，例如，终端102上可以部署有ECG测量装置，例如ECG电极或ECG传感器，从而利用终端102对用户进行ECG测量。该ECG测量也可以由终端102之外的其他任何ECG测量设备测量。
在一个实施例中，如图2所示，提供了一种ECG异常监测方法，在测试阶段，该方法包括以下步骤S210-S230：
S210，获取用户的PPG数据。
PPG描记法是借用光电手段在活体组织中检测血液容积变化的一种无创检测方法。PPG可以通过PPG数据探测器进行检测，PPG数据探测器是能够以无创方式检测用户的PPG的任何探测器设备，其可以包括光源和光电检测器，光 源向被测组织发射预定波长的光，光电检测器接收被测组织反射或透射的光，基于被测组织反射或透射的光的强度得到检测的PPG信号。PPG数据探测器例如可以是脉搏血氧计。
在一个示例中，脉搏血氧计例如可以内置于终端102中，该终端102例如可以是智能手表，当用户佩戴该智能手表时，位于智能手表表盘底部(靠近用户皮肤侧)的脉搏血氧计检测用户手腕内侧的PPG数据。在其他示例中，终端102也可以具有其他形态，PPG数据的测试部位也不限于用户手腕内侧，而可以是例如下肢踝内侧、胸部等其他部位。可以理解，在测试阶段采集的测试数据的测试部位与在训练阶段所使用的训练数据的测试部位相同。
PPG数据是表征用户的脉搏波幅值随着时间变化的数据，其可以表示为一维数据，即，将脉搏波的采样值按采样时间顺序排列得到的PPG一维数据序列。
在一个实施例中，获取用户的PPG数据是由便携终端设备实时监测的用户的PPG数据。
在一个示例中，终端102以预定采样频率对被测用户的测试部位进行持续采样监测。终端102从其持续监测的PPG信号中，截取预定时长的PPG数据片段，作为待输入模型的PPG数据。该预定采样频率和预定时长可以根据实际需要设定，例如预定采样频率可以是125Hz,预定时长可以是10s，则一个10s长度的PPG数据片段大约可以包括1250个采样值。图3中的深灰色曲线表示的数据示出了一个10s长度的PPG原始数据片段的示例，其中，横坐标表示时间，其用采样点序数表示，纵坐标表示幅值，本申请其余附图中显示的ECG和PPG的数据片段中的横纵坐标也具有同理的含义。
采集得到的原始PPG数据可能存在噪音或质量不高等问题，而影响监测结果的准确性。因此，可以对采集得到的原始数据进行预处理，再使用预处理后的数据输入模型进行监测。
在一个实施例中，上述步骤S210可以包括：获取用户的PPG原始数据，对PPG原始数据进行预处理，以得到PPG预处理数据。从而利用PPG预处理数据作为要输入到模型中的PPG数据。相应地，在后续步骤S220中，可以将PPG 预处理数据输入预先训练好的ECG异常监测模型中。
在一个实施例中，对PPG原始数据进行预处理，以得到PPG预处理数据，可以包括如下步骤中的一者或多者：
1)平坦度检测和信号筛选：
在此步骤中，检测信号的平坦度是否达到预定标准，若是，则保留该信号，否则，丢弃该信号。
示例地，对于截取的10s时长的PPG原始数据片段，如果其中存在任意60个连续采样点(约0.5秒)的采样值未超过阈值(1e-5)，则将其视为平坦信号，并丢弃这整个10s的数据片段。否则，保留该数据片段。
2)归一化：
在此步骤中，将信号归一化为零平均值和单位方差。
示例地，可以将前述平坦度检测和信号筛选步骤保留的数据片段中的各个采样值归一化为零平均值和单位方差，得到归一化后的数据片段。
3)滤波：
在此步骤中，对信号进行滤波处理。
示例地，可以采用三阶带通巴特沃斯滤波器，对前述步骤归一化后的数据片段进行滤波，以得到滤波后的数据片段。其中，该滤波器的低频带截止为0.5Hz，高频带截止为8Hz。
4)峰值检测和信号筛选：
在此步骤中，检测信号的单位长度内有效峰值数量是否达到预定标准，若是，则保留该信号，否则，丢弃该信号。
示例地，可以使用Python工具箱“HeartpPy”来检测前述步骤滤波后的数据片段内的有效峰值的数量。如果该10s数据片段内的有效峰值的数量低于5个(对应于低于30bpm)，则丢弃该数据片段。否则，保留该数据片段。
5)偏度SQI检测和信号筛选：
在此步骤中，检测信号的偏度是否达到预定标准，若是，则保留该信号，否则，丢弃该信号。
示例地，偏度以滑动窗口的方式来计算，窗口宽度为250个采样值(两秒)，步幅为125个采样值(一秒)。对于前述峰值检测和信号筛选步骤中保留的10s的数据片段，即可计算0～2s、1～3s、2～4s、3～5s、4～6s、5～7s、6～8s、7～9s、8～10s共9个时间窗口中每个时间窗口内采样值的偏度。如果计算得到的9个偏度中，大多数偏度(超过50％)为负，则认为该数据片段是低质量信号，丢弃该数据片段。否则，保留该数据片段。
6)离群采样值替换：
在此步骤中，检测信号中的离群采样值，将检测到的离群采样值替换为采样值中位数，以得到消除离群采样值后的信号。
示例地，可以应用Hampel滤波器来检测离群点。在前述偏度SQI检测和信号筛选步骤中保留的数据片段中，对于每连续10个采样值，计算中位绝对偏差(MAD)。然后根据假设正态分布的MAD值来估计该数据片段的标准差(std)。如果采样值与该数据片段的中位数相差3std，则该采样值被检测为离群点。将检测到的离群点替换为该数据片段的中位数。如此，能够将数据片段开头和结尾的较多离群点替换掉。
步骤1)可视为质量筛查步骤。步骤2)和3)是清洁和去噪，步骤4)至6)是质量控制步骤。参见图3所示，图3中示出了一个通过上述预处理步骤1)至6)而被保留的PPG数据片段，其中深灰色曲线代表原始数据片段，浅灰色曲线代表预处理后得到的预处理数据片段。参见图4所示，图4中示出了由于未能通过上述预处理步骤中的不同质量控制标准而被筛除的数据片段的示例。其中，图4中的(a)表示因平坦度未达标而被筛除的数据片段，(b)表示因有效峰值数量未达标而被筛除的数据片段，(c)表示因偏度未达标而被筛除的数据片段。
在本示例中，对于任意原始数据片段，经过上述预处理步骤后得到的预处理数据片段，可用于后续步骤输入模型中进行监测；而如果任意原始数据片段在上述任一步骤中被丢弃，则不再使用该原始数据片段进行监测。即，仅使用通过预处理步骤删选后的满足要求的预处理数据片段进行监测，从而保障监测的准确性。可以理解，在测试阶段对原始数据执行的预处理操作可以与在训练 阶段对原始数据执行的预处理操作相同，以保障模型的适用性。
S220，将PPG数据输入预先训练好的ECG异常监测模型中，以得到ECG异常监测模型的输出值。
其中，本申请的ECG异常监测模型是利用包括PPG数据以及标记PPG数据对应的ECG数据是否异常的标签的数据对的训练数据集训练得到的，该ECG异常监测模型的架构和训练过程将在后续描述。
训练好的ECG异常监测模型可以用于基于输入的PPG数据，输出指示所输入的PPG数据对应的ECG数据是否估计存在异常的输出值。
示例地，当将前述步骤中取得的10s的数据片段输入该ECG异常监测模型，该ECG异常监测模型可以输出监测类别概率作为输出值，将该监测类别概率与设定的阈值T进行比较，即可确定对应的ECG数据是否估计存在异常。例如，当时指示ECG数据估计存在异常，而当时指示ECG数据估计不存在异常。
S230，当ECG异常监测模型的输出值指示ECG数据估计存在异常时，使与用户相关联的用户设备发出警报。
在一个实施例中，S230中的与用户相关联的用户设备，可以包括上述的终端102，例如是用户佩戴的便携终端设备，以告知用户自身存在ECG异常风险，应及时进行ECG测量。
在另一个实施例中，S230中的与用户相关联的用户设备还可以是与用户关联的监护者的监护终端设备，例如用户的家属、看护者所用的终端设备，或者是用户的家庭医生、主治医生所用的终端设备，或者是用户所属的监护机构的终端设备或服务器等等，以告知监护者该用户存在ECG异常风险，应及时进行ECG测量。
在一个实施例中，使与用户相关联的用户设备发出警报包括以下各项中的一项或多项组合：在用户设备的显示屏上显示指示用户的ECG数据估计存在异常的视觉信息、使用户设备的指示灯发出指示用户的ECG数据估计存在异常的亮灯提示、使用户设备的振动器做出指示用户的ECG数据估计存在异常的振动 提示、使用户设备的声音输出装置发出指示用户的ECG数据估计存在异常的声音提示。
示例地，使与用户相关联的用户设备发出警报的内容除了包括告知ECG数据估计存在异常的监测结果外，还包括建议用户主动进行ECG测量的建议信息。
进一步地，在一个实施例中，上述方法还可以包括：当所述ECG异常监测模型的输出值指示所述ECG数据估计存在异常时，向预定警报接收设备拨打警报电话。
其中，该警报电话可以由警报发起设备发起，该警报发起设备可以是任何合适的终端或服务器。例如警报电话可以由上述用户所佩戴的便携终端设备发起，在此情况下，该便携终端设备需要是具有拨打电话功能的终端设备，例如是智能手表或者是智能手环和与智能手环配对的手机；或者也可以由例如上述服务器104或其他服务器或终端设备来发起。
该警报电话可以是由终端或服务器自动发起的或是基于用户的指令来决定是否发起，例如终端或服务器可以经由用户佩戴的便携终端设备向用户提供是否发起警报电话的选项，然后根据用户对选项的选择输入而决定是否向预定警报接收设备拨打警报电话，又或者终端或服务器可以在向用户提供上述选项后，如果预定时段内未收到用户的选择输入，则认为用户可能处于失去意识等危险状况中，则自动向预定警报接收设备拨打警报电话。
用于接收警报电话的预定警报接收设备可以根据需要预先设定，例如该预定警报接收设备可以包括上述与用户关联的监护者的监护终端设备，和/或指定医疗机构的警报终端(例如可通过拨号112而拨打的医院急救电话终端)等等。
取决于预先设定或根据用户的选择输入，警报电话可以是接通实时通话以使用户与预定警报接收设备处的接收人通话，警报电话也可以是包括预先设定的语音播报内容，例如可以包括告知用户的ECG数据估计存在异常的监测结果有关的语音播报内容，此外还可以包括例如用户的姓名、性别、年龄、居所等个人信息以及用户的实时位置等信息的语音播报内容，以方便警报电话接收者快速知晓用户的当前状况和实时位置。
在一个实施例中，上述方法还可以包括：当ECG异常监测模型的输出值指示ECG数据估计存在异常时，对用户进行ECG测量，以得到指示用户的ECG是否异常的测量结果。
其中，该ECG测量可由通过设置有ECG测量装置的终端102进行，也可以通过与终端102不同的其他ECG测量设备，例如设置于医疗机构中的12导联ECG，或者家用ECG测量设备等进行。例如，在步骤S230中终端102发出警报提醒用户进行ECG测量后，终端102可以基于用户输入来确定是否触发由终端102对用户的ECG测量，又例如，终端102可以在输出值指示ECG数据估计存在异常时，自动触发对用户的ECG测量，又例如，在步骤S230中终端102发出警报提醒用户/监护者进行ECG测量后，用户/监护者可以自行选择利用终端102之外的其他ECG测量设备对用户进行ECG测量，例如用户可以自行前往医疗机构利用12导联ECG进行更专业全面地ECG测量，并得到ECG是否异常的测量结果。其中，当由终端102执行上述ECG测量时，终端102可以由自身或经由服务器等外部设备对测得的ECG数据进行分析，并将测得的ECG数据和分析得到的ECG是否异常的测量结果通过终端102推送给用户。
上述ECG异常监测方法，由于PPG数据能够以被动的方式通过例如智能手表等便携式可穿戴设备对用户进行测量得到，因此通过获取用户的PPG数据，能够方便而持续地对用户进行监测，通过巧妙地使用PPG数据对应的ECG数据是否异常的标签来对PPG数据进行标记形成的数据对进行训练，能够得到可以基于输入的PPG数据，输出指示所输入的所述PPG数据对应的ECG数据是否估计存在异常的输出值的ECG异常监测模型，从而利用该模型可以基于测得PPG数据监测到对应的ECG数据是否很可能发生异常时，并在监测到发生异常时，使与所述用户相关联的用户设备发出警报，以建议用户主动进行ECG测量获知ECG是否异常的准确结果。
以上描述了在测试阶段，终端102或服务器104可能执行的方法步骤，其中在测试阶段的步骤S220中，使用了预先训练好的ECG异常监测模型，这要求在测试阶段之前的训练阶段中，预先构造ECG异常监测模型的模型架构，并 对该构造好的ECG异常监测模型进行训练，优化确定ECG异常监测模型的模型参数，以得到训练好的信号质量评估模型。
本申请提出了一种双卷积注意力网络(Dual-Convolutional Attention Network，DCA-Net)作为上述ECG异常监测模型。该双卷积注意力网络的特征是，包括有一个或多个残差-双卷积注意力(Res-DAC)块。每个残差-双卷积注意力块是在残差块中加入双卷积注意力(dual-convolutional-attention，DCA)块而形成。其中，该双卷积注意力块包括用于在通道维度上对数据进行卷积的第一卷积注意力块以及用于在时间维度上对数据进行卷积的第二卷积注意力块。
在一个实施例中，参见图5所示，该ECG异常监测模型至少可以包括串行设置的多个残差-双卷积注意力块1～N。
在一个实施例中，参见图5所示，该ECG异常监测模型还可以包括位于多个残差-双卷积注意力块之前，且与多个残差-双卷积注意力块串行设置的输入卷积层和池化层，以及位于多个残差-双卷积注意力块之后，且与多个残差-双卷积注意力块串行设置的平均池化层和全连接层。
其中，输入卷积层接收对ECG异常监测模型输入的PPG数据，输入卷积层的输出馈入池化层，池化层的输出馈入多个残差-双卷积注意力块中的首个残差-双卷积注意力块，多个残差-双卷积注意力块中的前一者的输出馈入后一者，多个残差-双卷积注意力块中的最后一个残差-双卷积注意力块的输出馈入平均池化层，平均池化层的输出馈入全连接层，全连接层输出ECG异常监测模型的输出值。
其中，全连接层的最后一层可以具有单个神经元，该单个神经元输出一个1*1维的输出值，即概率将该概率与设定的阈值T进行比较，即可得到ECG是否估计存在异常的监测结果。
示例地，模型中的各个卷积层，例如输入卷积层，可以为一维卷积层，以适应对PPG一维数据序列的处理需求。
可以理解，图5中的模型架构仅为示例，在模型中包括有一个或多个如本申请提出的残差-双卷积注意力块的基本前提下，本领域技术人员还可以对模型 的架构做出许多变形，例如残差-双卷积注意力块的数量和串行/并行方式可以改变，模型中的卷积层、池化层数量和位置等也可以改变。
在一个实施例中，参见图6和图7所示，该多个残差-双卷积注意力块中的每个残差-双卷积注意力块可以包括第一主路径以及第一短路(shortcut)分支路径。该第一主路径上串行设置有一个或多个卷积层以及双卷积注意力块。其中，双卷积注意力块包括串行或并行设置的第一卷积注意力块和第二卷积注意力块，第一卷积注意力块用于在通道维度上进行卷积，第二卷积注意力块用于在时间维度上进行卷积。该第一短路分支路径并联于第一主路径上串行设置的一个或多个卷积层以及双卷积注意力块的两端。
其中，一个或多个卷积层以及双卷积注意力块中的最前者以及第一短路分支路径接收先前路径的输出，一个或多个卷积层以及双卷积注意力块中的后一者接收前一者的输出，一个或多个卷积层以及双卷积注意力块中的最后者的输出与第一短路分支路径的输出矩阵相加后馈入后续路径。
在一个实施例中，参见图6所示，每个残差-双卷积注意力块中的各个卷积层的输出侧还分别设置有批量归一化层以及激活函数层。其中，在每个残差-双卷积注意力块中，除了最后一个卷积层之外的每个卷积层的输出侧的批量归一化层以及激活函数层设置于下一个卷积层之前，而最后一个卷积层的输出侧的批量归一化层设置于该卷积层与双卷积注意力块之间，双卷积注意力块设置于与第一短路分支路径的输出的相加运算之前，最后一个卷积层的输出侧设置的批量归一化层设置于与第一短路分支路径的输出的相加运算之后。
在一个具体示例中，参见图6所示，每个残差-双卷积注意力块中设置有两个卷积层，且它们为一维卷积层。激活函数层为ReLU激活函数层。在其他示例中，残差-双卷积注意力块中的卷积层的数量和维度可以根据情况而更改，激活函数层也可以为其他类型的激活函数层。
在一个实施例中，参见图7所示，第一卷积注意力块和第二卷积注意力块串行设置，且第一卷积注意力块位于第二卷积注意力块之前。将通道方面的第一卷积注意力块置于时间方面的第二卷积注意力块之前，使得双卷积注意力块 先学习通道间交互，然后基于通道参与输出来学习时间方面交互，这比两者并行或通道方面的第一卷积注意力块在后的顺序表现更好。在其他示例中，也可以是第二卷积注意力块位于第一卷积注意力块之前或者第一卷积注意力块和第二卷积注意力块并行设置。
在一个实施例中，参见图8所示，第一卷积注意力块和第二卷积注意力块中的每一者包括：第二主路径以及第二短路分支路径。第二主路径上设置有最大池化层和平均池化层、卷积层以及激活函数层，最大池化层和平均池化层并行设置，并行设置的最大池化层和平均池化层依次与卷积层和激活函数层串行设置。第二短路分支路径并联于第二主路径上设置的最大池化层和平均池化层、卷积层以及激活函数层的两端。
其中，第二短路分支路径、最大池化层和平均池化层分别接收先前路径的输出，最大池化层的输出与平均池化层的输出联合(concatenation)所得的输出馈入卷积层，卷积层的输出馈入激活函数层，激活函数层的输出与第二短路分支路径的输出相乘后馈入后续路径。
在一个具体示例中，参见图8所示，第一/第二卷积注意力块中具有一个卷积层，且为一维卷积层。激活函数层为Sigmoid激活函数层。在其他示例中，第一/第二卷积注意力块中的卷积层的数量和维度可以根据情况而更改，激活函数层也可以为其他类型的激活函数层。
其中，卷积注意力块依赖于将卷积运算应用到数据的某个维度，以学习沿着该维度的数据点之间的交互。与MLP注意力模块相比，卷积注意力在计算上更高效，因为卷积层中不需要参数。从图8中可见，第一卷积注意力块和第二卷积注意力块具有类似的架构，但是他们应用于不同的方向，第一卷积注意力块应用于卷积通道方向，第二卷积注意力块应用于时域方向。
在上述实施例中，各个层/块的参数，例如激活函数(activation function)、卷积核尺寸(kernel size)、步长、补丁(padding)等的具体配置可以根据实际需要设定。
具体来说，对于图7和图8所示的具体示例，以图8中的卷积层为卷积核 大小为7的一维卷积层为例，设通道方面的第一卷积注意力块的输入(也是DCA模块的输入)为其中N是采样点/批量大小，C是从原始残差块输出的卷积通道的数量，D是信号长度(记为时间维度)。第一卷积注意力块的输出为：
其中，g(X)＝[p1(X)||p2(X)]D              (2)
其中，p1(X)＝max_poolD(X)              (3)
并且，p2(X)＝mean_poolD(X)              (4)
等式(1)中，表示Hadamard乘积(元素方面矩阵乘法)；σ(·)表示Sigmoid激活函数，表示卷积核大小为7且在通道方向上卷积的1D卷积层。等式(2)中的[·||·]D表示沿时间方向的矩阵联合，等式(3)和(4)中的下标D表示池化层也在时间维度上应用。函数输出通道方面注意力权重，这些权重进一步扩展到相同形状的X以缩放X。然后将缩放后的X(表示为)用作向时间方面注意力的输入。
同样，设时间方面的第二卷积注意力块的输出为
其中，
其中，
并且，
等式(7)-(8)表示首先在通道维度上操作最大池化和平均池化，因此，和然后，将池化后的矩阵沿着通道维度联合，(等式(6))。一维卷积层在时间维度上卷积，并输出单个通道，使得最后，这个卷积后的矩阵由Sigmoid激活函数激活并用于缩放输入的时域。
在一个具体示例中，可以通过将已知的ResNet-34模型中的二维卷积层均替换为一维卷积层，将ResNet-34模型中最后的全连接层替换为输入为512维输出为1维的全连接层，并在ResNet-34模型的每个残差块中增加残差之前增加一个 双卷积注意力块，而得到一个具体示例的DCA-Net模型。可以理解，该具体示例的DCA-Net模型将包括依次串行的输入卷积层、池化层、16个残差-双卷积注意力块、平均池化层以及全连接层。该具体示例的DCA-Net模型的架构可以参考图5所示，其中N取值N＝16，并且其中每个残差-双卷积注意力块的架构可以参考图6至图8所示。其中，示例地，在每个残差-双卷积注意力块中，第一卷积注意力块和第二卷积注意力块的卷积核大小均可以为7，步长均为1，平均池化层可以为一维自适应平均池化层(AdaptiveAvgPool1d)，模型中其他层和块的参数均示例地可以参照已知ResNet-34模型的参数设置。
示例地，以通过批次训练(batch training)来执行训练为例，来说明该具体示例的ResNet-34模型的前向传播过程。当批次大小为64，向该具体示例的DCA-Net模型输入一个64x1x1250的一维PPG数据时，经过输入卷积层(步长为2)后输出64×64×625的数据，经过(最大)池化层(步长为2)后输出64×64×313的数据，然后该数据输入到第一个残差-双卷积注意力块中，在依次经过图6中所示的卷积层、批量归一化层、ReLU激活函数层、卷积层、批量归一化层后，输入到双卷积注意力层中，输入数据首先进入通道方面的第一卷积注意力块，在第一卷积注意力块内，参见图7，平均池化层对输入数据在时间维度进行平均池化而输出一个64×64×1矩阵，同时最大池化层对输入数据在时间维度进行最大池化而输出另一个64×64×1矩阵，这两个矩阵沿时间维度联合(类似于拼接操作)在一起得到64×64×2的矩阵。该64×64×2的矩阵经过卷积层而被沿通道维度进行卷积，以学习64×64×1的注意力权重，再沿时间维度被复制展开成64×64×313的注意力分数。将此注意力分数与经过第二短路分支路径输出的原输入数据进行Hadamard乘积而输出64×64×313矩阵。接着，第一卷积注意力块输出的64×64×313矩阵输入时间方面的第二卷积注意力块，在第二卷积注意力块内，参见图7，平均池化层对64×64×313矩阵在通道维度进行平均池化而输出一个64×1×313的矩阵，同时最大池化层对输入数据在通道维度进行最大池化而输出另一个64×1×313的矩阵。这两个矩阵联合在一起得到64×2×313的矩阵。该64×2×313的矩阵经过卷积层而被沿时间维度进行卷积，得到64×1×313的注意 力权重，再沿通道维度被复制展开成64×64×313的注意力分数。将此注意力分数与经过第二短路分支路径输出的原输入64×64×313矩阵进行Hadamard乘积，而输出64×64×313的矩阵。该64×64×313的矩阵作为第二卷积注意力块的输出接着被输入到下一个残差-双卷积注意力块中，在经过16个残差-双卷积注意力块后，输出数据然后经过平均池化层和全连接层后，输出模型的输出值
在训练阶段中，通过对上述任一实施例中构造的ECG异常监测模型进行训练，能够确定训练好的ECG异常监测模型。具有上述构造的ECG异常监测模型可以利用各种常规的或新型的训练方法来训练。
其中，为了执行模型训练，需要构建训练数据集，在本申请中，训练数据集可以包括多组PPG数据-标记对应的ECG数据是否异常的标签的数据对。其中，训练数据集中包括的PPG数据，及标记标签所需用到的对应的ECG数据，可以采集自多个受试者对应的多个病例。其中构建训练数据集所使用数据所来源的受试者的数量，可以根据需要而设定，例如可以但不限于是几百、几千或几万个，从而得到能够在统计学上反映大多数使用者的特性的数据。针对多个受试者中的每个受试者，可以采集该受试者的一条或多条PPG数据以及对应的ECG数据，然后通过例如医生人工或机器辅助方式标记其中每个ECG数据是否存在心脏功能异常的标签，并将该标签对应到与该ECG数据相对应的PPG数据上，从而可以获得该受试者的一组或多组PPG数据-标记对应的ECG数据是否异常的标签的数据对。将采集自上述多个受试者的多组PPG数据-标记对应的ECG数据是否异常的标签的数据对汇集起来，即可得到示例可用的包括多个受试者数据的训练数据集。其中，上述多个受试者中可以至少包括一些存在心脏功能异常疾病的患者，以能够至少采集到部分具有心脏功能异常特性的PPG数据和对应的ECG数据。此外，上述多个受试者中也可以包括一些心脏功能正常的受试者，或者虽然仅包括存在心脏功能异常疾病的患者，但是至少采集到部分具有心脏功能正常特性的PPG数据和对应的ECG数据(例如在患者未发病期间采集的正常数据)。例如，在实践中，构建训练数据集所用的PPG数据和对应的ECG数据可以采集自一个或多个医院的大量不同受试者在住院期间的临床医 学数据。值得说明的是，可以理解，在训练阶段所使用的训练数据所来源的受试者，不需要是或者包含在测试阶段进行测试的受试者。例如，可以采用来自医疗系统的大量临床数据来构成训练数据集对模型进行训练，然后训练好的模型可以部署到终端102上，此后，任何用户均可以使用该终端102来执行上述测试阶段的过程以对自身的心电图异常进行监测。
示例地，训练数据集的构建过程可以包括如下步骤：
1)获取多个导联的ECG信号及其对应的PPG信号。
其中，ECG信号与PPG信号对应是指，该ECG信号与该PPG信号是在同一受试者身上同时测量的，以确保该ECG信号与该PPG信号之间时间对准并彼此关联。
示例地，可以从MIMIC-III波形数据库(Waveform Database，WDB)匹配子集中获取多个导联的ECG信号及其对应的PPG信号。MIMIC-III WDB与临床记录相联系——匹配子集中可以包含有大量患者，并且匹配子集中的所有患者均被识别并被匹配到医疗记录系统。MIMIC-III WDB包括多个导联的ECG信号及其配对的PPG信号。对于每例患者，在其住院期间的不同时间点采集不同长度的波形数据。这些信号的采样频率为125Hz。由于不同的ECG记录具有不同的导联数量和类型，在本示例中，选择了3种最常见和最有代表性的ECG导联的数据，即导联II、V和AVR。因此，将使用这三根导联的ECG信号以及对应的PPG信号记录纳入到本分析中。然而在其他示例中，也可能选择更多或更少的ECG导联，所选导联的类型也可能变更。此外，为了降低噪声水平，只选取时长超过预定时长，例如两分钟的信号的记录。时长不到两分钟的信号将被筛除。
2)从获取的多个导联的ECG信号及其对应的PPG信号中截取预定时长的多个导联的ECG原始数据片段及其对应的PPG原始数据片段。
训练过程中训练数据要截取的数据片段的预定时长与测试过程中截取的数据片段的预定时长相同，例如也为10秒。如此，可以从每个ECG/PPG信号中截取出第二个10秒的数据片段，以避免初始时段采样值的不稳定。如此，截取 得到的10秒时长的导联II、V和AVR的ECG原始数据片段以及对应的10秒时长的PPG原始数据片段的示例如图9中所示。
3)对所截取的预定时长的多个导联的ECG原始数据片段及其对应的PPG原始数据片段分别进行预处理，以预处理并筛选出满足要求的预定时长的多个导联的ECG预处理数据片段及其对应的PPG预处理数据片段。
对于10s的PPG原始数据片段，可以执行如上所述测试过程中相同的包括步骤1)至6)的预处理过程，以筛选出满足要求的PPG预处理数据片段。
对于10s的导联II、V和AVR的ECG原始数据片段，可以使用0.5Hz的5阶高通巴特沃兹滤波器，随后使用电力线频率为50Hz的电力线滤波，以得到10s的导联II、V和AVR的ECG预处理数据片段。然后，使用预定信号质量评估方法来评估每个ECG预处理数据片段的信号质量。该评估输出三类质量中的一类——不可接受、几乎不可接受或优秀。当某一导联的ECG预处理数据片段被评估为优秀时，该导联的ECG预处理数据片段被保留，反之，当某一导联的ECG预处理数据片段被评估为不可接受、几乎不可接受时，该导联的ECG预处理数据片段被丢弃。
如此，对于任意一组10s的多个导联的ECG原始数据片段及其对应的PPG原始数据片段，如果任意一个导联的ECG原始数据片段不能通过上述预处理过程被保留，或者PPG原始数据片段不能通过上述预处理过程被保留，则该组数据片段被整体丢弃。只有一组中的多个导联的ECG原始数据片段及其对应的PPG原始数据片段中的每一者均通过上述预处理过程被保留，才能得到该组的满足要求的预定时长的多个导联的ECG原始数据片段及其对应的PPG原始数据片段。
4)针对每组预定时长的多个导联的ECG预处理数据片段及其对应的PPG预处理数据片段，确定标记该多个导联的ECG预处理数据片段是否异常的标签，将该标签与该PPG预处理数据片段组成数据对，从而得到包括多组PPG数据以及标记所述PPG数据对应的ECG数据是否异常的标签的数据对的训练数据集。
由于MIMIC-III WDB中没有给出ECG标签，示例地，可以使用已知的具 有较高准确性的ECG分类模型，例如AutoNet来高效地标记ECG预处理数据片段是否异常的标签。参见图10和图11所示，图10中示出了被标记为正常的ECG预处理数据片段及其对应的PPG预处理数据片段的示例，图11中示出了被标记为异常的ECG预处理数据片段及其对应的PPG预处理数据片段的示例。在其他替代示例中，也可以由专业技术人员，例如临床医生来手动标记ECG预处理数据片段是否异常的标签，以确保标签的准确性。如此，可以得到多组PPG预处理数据片段-标记对应的预处理数据片段ECG是否异常的标签这样的数据对，以构成所需的训练数据集。
在构建好上述训练数据集后，即可利用该训练数据集对上述ECG异常监测模型进行训练以确定训练好的ECG异常监测模型。
在一个实施例中，ECG异常监测模型的训练过程包括：
获取训练数据集；训练数据集包括多组PPG数据-标记对应的ECG数据是否异常的标签的数据对；
将训练数据集划分为训练集、验证集和测试集；
利用训练集对ECG异常监测模型进行训练以得到拟合的模型参数，利用验证集对拟合的模型参数确定的ECG异常监测模型进行验证并调节模型超参数，利用测试集对调节后的超参数和训练得到模型参数确定的训练好的ECG异常监测模型进行性能评估。
示例地，可以将上述一个具体示例的DCA-Net模型应用基于MIMIC-III WDB的处理的PPG数据片段，使用标记对应的ECG数据片段是否异常的标签进行二进制分类，来进行训练。将38320组PPG数据片段-标记对应的ECG数据片段是否异常的标签的数据对分成训练集、验证集和测试集，比例为64％：16％：20％。使用训练集学习模型参数，并在验证集中进行验证，以避免过度拟合。测试集仅用于报告结果。使用特异性、灵敏度/召回率、精确率、准确率、AUROC(接收器操作特征曲线下面积)和AUPRC(精确率-召回率曲线下面积)评价了模型性能。进一步对模型进行了10次随机重复，并采用了上述评估指标的平均值和标准差。
DCA-Net模型可以使用PyTorch 1.11.0版和CUDA 11.7版实施。损失函数为二元交叉熵。进一步实验通过增加损失权重来调整类失衡。模型由Adam优化器使用默认超参数设置进行优化，并使用64的批次大小进行训练。应用了步幅方面的学习率调度器，以将学习率降低到每10历元0.1倍。还应用了提前停止——如果验证损失与5个连续历元的当前最低损失相比没有减少，则终止训练。始终保存损失最低的模型，并将其应用于测试集以进行评估。
应该理解的是，虽然图2的流程图中的各个步骤按照箭头的指示依次显示，但是这些步骤并不是必然按照箭头指示的顺序依次执行。除非本文中有明确的说明，这些步骤的执行并没有严格的顺序限制，这些步骤可以以其它的顺序执行。而且，图2中的至少一部分步骤可以包括多个子步骤或者多个阶段，这些子步骤或者阶段并不必然是在同一时刻执行完成，而是可以在不同的时刻执行，这些子步骤或者阶段的执行顺序也不必然是依次进行，而是可以与其它步骤或者其它步骤的子步骤或者阶段的至少一部分轮流或者交替地执行。
性能测试
为了验证本申请提出的DCA-Net模型在架构上的优越性能，将本申请上述具体示例的DCA-Net模型与现有的神经网络模型ResNet-18、ResNet-34、ResNet-50、ECA-Net同样用于基于PPG数据监测ECG异常，以进行性能测试与比较，比较结果如下表1所示。表1中，所显示的数字是随机重复10次的平均分数百分比，其标准差在括号中显示。最佳平均准确率、AUROC和AUPRC用粗体突出显示。
表1本申请一个具体示例的DCA-Net模型与现有神经网络模型的性能比较结果。
从上表可见，本申请提出的DCA-Net具有最佳的平均准确率、AUROC和AUPRC评分。在单侧T检验中，DCA-Net的AUPRC显著优于ECA-Net，其中p值＝0.01。此外，由于所有评估测量的标准偏差较小，DCA-Net的性能比ECA-Net更稳定。
在一个实施例中，如图12所示，提供一种ECG异常监测装置1200，包括：
数据获取模块1210，用于获取用户的PPG数据；
模型监测模块1220，用于将PPG数据输入预先训练好的ECG异常监测模型中，以得到ECG异常监测模型的输出值；其中，ECG异常监测模型是利用包括PPG数据以及标记PPG数据对应的ECG数据是否异常的标签的数据对的训练数据集训练得到的，ECG异常监测模型用于基于输入的PPG数据，输出指示所输入的PPG数据对应的ECG数据是否估计存在异常的输出值；
异常警报模块1230，用于当ECG异常监测模型的输出值指示ECG数据估计存在异常时，使与用户相关联的用户设备发出警报。
在一个实施例中，ECG异常监测装置1200还包括模型训练模块1240，模型监测模块1220还用于从模型训练模块1240获取训练好的ECG异常监测模型，该模型训练模块1240用于对ECG异常监测模型进行训练，以得到训练好的ECG异常监测模型。
示例地，上述数据获取模块1210、模型监测模块1220、异常警报模块1230可以设置于终端102中，而上述模型训练模块1300则可以设置于服务器104中。
关于ECG异常监测装置1200的具体限定可以参见上文中对于ECG异常监测方法的限定，在此不再赘述。上述ECG异常监测装置1200中的各个模块可 全部或部分通过软件、硬件及其组合来实现。上述各模块可以硬件形式内嵌于或独立于计算机设备中的处理器中，也可以以软件形式存储于计算机设备中的存储器中，以便于处理器调用执行以上各个模块对应的操作。
在一个实施例中，如图13所示，提供一种ECG异常监测设备1300，包括PPG数据探测器1310、存储器1320和处理器1330；
PPG数据探测器1310用于监测用户的PPG数据，并将测得的PPG数据传输至处理器；
存储器1320存储有计算机程序；并且
处理器1330执行计算机程序时，用于从PPG数据探测器1310接收PPG数据，并实现如上任一实施例的方法。
在一个实施例中，PPG数据探测器可以是脉搏血氧计。
在一个实施例中，ECG异常监测设备1300还包括输出装置1340，输出装置1340可以包括显示屏、指示灯、振动器和/或声音输出装置。显示屏可以是液晶显示屏或者电子墨水显示屏等。声音输出装置可以是例如扬声器、蜂鸣器等具有声音输出功能的装置。如此，当利用ECG异常监测设备1300自身来发出警报时，可以通过该输出装置1340来发出警报。
在一个实施例中，ECG异常监测设备1300还具有拨打电话功能，例如ECG异常监测设备1300还可以包括具有拨打电话功能的拨号装置，该拨号装置例如可以通过与用户关联的智能手表、手机或服务器等来实现，以在需要向预定警报接收设备拨打警报电话时，通过该拨号装置来拨打该警报电话。
在一个实施例中，ECG异常监测设备1300还具有ECG测量功能，例如ECG异常监测设备1300还可以包括ECG测量装置，以在利用PPG数据估计存在ECG异常时，进一步利用ECG测量装置对用户进行ECG测量。
可以理解，ECG异常监测设备1300可以是能够执行本申请上述实施例方法的任意设备，其可能具有各种不同的形态，例如其可以包括前述的终端102，或者其也可以包括前述的终端102和服务器104。
在一个实施例中，提供一种计算机可读存储介质，其上存储有计算机程序， 其特征在于，计算机程序被处理器执行时实现以下步骤：
获取用户的PPG数据；
将PPG数据输入预先训练好的ECG异常监测模型中，以得到ECG异常监测模型的输出值；其中，ECG异常监测模型是利用包括PPG数据以及标记PPG数据对应的ECG数据是否异常的标签的数据对的训练数据集训练得到的，ECG异常监测模型用于基于输入的PPG数据，输出指示所输入的PPG数据对应的ECG数据是否估计存在异常的输出值；
当ECG异常监测模型的输出值指示ECG数据估计存在异常时，使与用户相关联的用户设备发出警报。
在其他实施例中，计算机程序被处理器执行时还实现如上任一实施例的方法的步骤。
本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程，是可以通过计算机程序来指令相关的硬件来完成，所述的计算机程序可存储于一非易失性计算机可读取存储介质中，该计算机程序在执行时，可包括如上述各方法的实施例的流程。其中，本申请所提供的各实施例中所使用的对存储器、存储、数据库或其它介质的任何引用，均可包括非易失性和/或易失性存储器。非易失性存储器可包括只读存储器(ROM)、可编程ROM(PROM)、电可编程ROM(EPROM)、电可擦除可编程ROM(EEPROM)或闪存。易失性存储器可包括随机存取存储器(RAM)或者外部高速缓冲存储器。作为说明而非局限，RAM以多种形式可得，诸如静态RAM(SRAM)、动态RAM(DRAM)、同步DRAM(SDRAM)、双数据率SDRAM(DDRSDRAM)、增强型SDRAM(ESDRAM)、同步链路(Synchlink)DRAM(SLDRAM)、存储器总线(Rambus)直接RAM(RDRAM)、直接存储器总线动态RAM(DRDRAM)、以及存储器总线动态RAM(RDRAM)等。
以上实施例的各技术特征可以进行任意的组合，为使描述简洁，未对上述实施例中的各个技术特征所有可能的组合都进行描述，然而，只要这些技术特征的组合不存在矛盾，都应当认为是本说明书记载的范围。
以上所述实施例仅表达了本申请的几种实施方式，其描述较为具体和详细，但并不能因此而理解为对发明专利范围的限制。应当指出的是，对于本领域的普通技术人员来说，在不脱离本申请构思的前提下，还可以做出若干变形和改进，这些都属于本申请的保护范围。因此，本申请专利的保护范围应以所附权利要求为准。

Claims (20)

  1. 一种心电图异常监测方法，其特征在于，所述方法包括：
    获取用户的PPG数据；
    将所述PPG数据输入预先训练好的ECG异常监测模型中，以得到所述ECG异常监测模型的输出值；其中，所述ECG异常监测模型被预先训练成用于基于输入的PPG数据，输出指示所输入的所述PPG数据对应的ECG数据是否估计存在心脏功能异常的输出值；
    当所述ECG异常监测模型的输出值指示所述ECG数据估计存在心脏功能异常时，使与所述用户相关联的用户设备发出警报。
  2. 根据权利要求1所述的异常监测方法，其特征在于，所述ECG异常监测模型是利用训练数据集训练得到的，所述训练数据集包括PPG数据，以及包括所述PPG数据对应的ECG数据指示是否存在心脏功能异常的标签。
  3. 根据权利要求2所述的ECG异常监测方法，其特征在于，所述训练数据集的构建过程包括：
    获取多个导联的ECG信号及其对应的PPG信号；
    从获取的所述多个导联的ECG信号及其对应的PPG信号中，截取预定时长的多个导联的ECG原始数据片段及其对应的PPG原始数据片段；
    对所截取的所述预定时长的多个导联的ECG原始数据片段及其对应的PPG原始数据片段分别进行预处理，以得到多组预定时长的多个导联的ECG预处理数据片段及其对应的PPG预处理数据片段；
    针对每组所述预定时长的多个导联的ECG预处理数据片段及其对应的PPG预处理数据片段，确定指示所述多个导联的ECG预处理数据片段是否存在心脏功能异常的标签；
    将所述标签与对应PPG预处理数据片段组成数据对，从而得到包括多组所述数据对的所述训练数据集。
  4. 根据权利要求1所述的心电图异常监测方法，其特征在于，所述获取用户的PPG数据，将所述PPG数据输入预先训练好的ECG异常监测模型中， 包括：
    获取用户的PPG原始数据；
    对所述PPG原始数据进行预处理，以得到PPG预处理数据；
    将所述PPG预处理数据输入预先训练好的ECG异常监测模型中。
  5. 根据权利要求4所述的心电图异常监测方法，其特征在于，所述预处理包括以下中的一者或多者：
    平坦度检测和信号筛选；
    归一化；
    滤波；
    峰值检测和信号筛选；
    偏度SQI检测和信号筛选离群采样值替换。
  6. 根据权利要求1所述的心电图异常监测方法，其特征在于，所述ECG异常监测模型包括以下中的一者：
    DCA-Net模型；
    ResNet-18模型；
    ResNet-34模型；
    ResNet-50模型；
    ECA-Net模型。
  7. 根据权利要求1所述的心电图异常监测方法，其特征在于，所述心电图异常监测模型包括一个或多个残差-双卷积注意力块，所述残差-双卷积注意力块是在残差块中加入双卷积注意力块而形成，所述双卷积注意力块包括用于在通道维度上进行卷积的第一卷积注意力块以及用于在时间维度上进行卷积的第二卷积注意力块。
  8. 根据权利要求7所述的心电图异常监测方法，其特征在于，所述心电图异常监测模型包括串行设置的多个所述残差-双卷积注意力块，所述多个所述残差-双卷积注意力块中的每个残差-双卷积注意力块包括：
    第一主路径，所述第一主路径上串行设置有一个或多个卷积层以及所述 双卷积注意力块；其中，所述双卷积注意力块包括串行或并行设置的所述第一卷积注意力块和所述第二卷积注意力块；以及
    第一短路分支路径，所述第一短路分支路径并联于所述第一主路径上串行设置的所述一个或多个卷积层以及双卷积注意力块的两端；
    其中，所述一个或多个卷积层以及双卷积注意力块中的最前者以及所述第一短路分支路径接收先前路径的输出，所述一个或多个卷积层以及双卷积注意力块中的后一者接收前一者的输出，所述一个或多个卷积层以及双卷积注意力块中的最后者的输出与所述第一短路分支路径的输出相加后馈入后续路径。
  9. 根据权利要求8所述的心电图异常监测方法，其特征在于，所述第一卷积注意力块和所述第二卷积注意力块串行设置，且所述第一卷积注意力块位于所述第二卷积注意力块之前。
  10. 根据权利要求8所述的心电图异常监测方法，其特征在于，所述第一卷积注意力块和所述第二卷积注意力块中的每一者包括：
    第二主路径，所述第二主路径上设置有最大池化层和平均池化层、卷积层以及激活函数层，所述最大池化层和平均池化层并行设置，所述并行设置的最大池化层和平均池化层依次与卷积层和激活函数层串行设置；以及
    第二短路分支路径，所述第二短路分支路径并联于所述第二主路径上设置的所述最大池化层和平均池化层、卷积层以及激活函数层的两端；
    其中，所述第二短路分支路径、所述最大池化层和所述平均池化层分别接收先前路径的输出，所述最大池化层的输出与所述平均池化层的输出联合所得的输出馈入卷积层，所述卷积层的输出馈入激活函数层，所述激活函数层的输出与所述第二短路分支路径的输出相乘后馈入后续路径。
  11. 根据权利要求8所述的心电图异常监测方法，其特征在于，每个残差-双卷积注意力块中的各个卷积层的输出侧还分别设置有批量归一化层以及激活函数层。
  12. 根据权利要求8所述的心电图异常监测方法，其特征在于，所述ECG 异常监测模型还包括：
    输入卷积层和池化层，位于所述多个残差-双卷积注意力块之前，且与所述多个残差-双卷积注意力块串行设置，以及
    平均池化层和全连接层，位于所述多个残差-双卷积注意力块之后，且与所述多个残差-双卷积注意力块串行设置；
    其中，所述输入卷积层接收对所述ECG异常监测模型输入的PPG数据，所述输入卷积层的输出馈入池化层，所述池化层的输出馈入所述多个残差-双卷积注意力块中的首个残差-双卷积注意力块，所述多个残差-双卷积注意力块中的最后一个残差-双卷积注意力块的输出馈入所述平均池化层，所述平均池化层的输出馈入所述全连接层，所述全连接层输出所述ECG异常监测模型的输出值。
  13. 根据权利要求1至12中任一项所述的心电图异常监测方法，其特征在于，所述ECG异常监测模型的训练过程包括：
    获取训练数据集；所述训练数据集包括多组PPG数据-标记对应的ECG数据指示是否存在心脏功能异常的标签的数据对；
    将所述训练数据集划分为训练集、验证集和测试集；
    利用所述训练集对所述ECG异常监测模型进行训练以得到拟合的模型参数，利用所述验证集对拟合的所述模型参数确定的ECG异常监测模型进行验证并调节模型超参数，利用所述测试集对调节后的超参数和训练得到的模型参数确定的训练好的ECG异常监测模型进行性能评估。
  14. 根据权利要求1至12中任一项所述的心电图异常监测方法，其特征在于，与所述用户相关联的用户设备包括所述用户佩戴的便携终端设备和/或与所述用户关联的监护者的监护终端设备，所述获取用户的PPG数据包括由所述便携终端设备实时监测所述用户的PPG数据；
    所述使与所述用户相关联的用户设备发出警报包括以下各项中的一项或多项：在所述用户设备的显示屏上显示指示所述用户的ECG数据估计存在心脏功能异常的视觉信息、使所述用户设备的指示灯发出指示所述用户的ECG 数据估计存在心脏功能异常的亮灯提示、使所述用户设备的振动器做出指示所述用户的ECG数据估计存在心脏功能异常的振动提示、使所述用户设备的声音输出装置发出指示所述用户的ECG数据估计存在心脏功能异常的声音提示。
  15. 根据权利要求1至12中任一项所述的心电图异常监测方法，其特征在于，所述方法还包括：
    当所述ECG异常监测模型的输出值指示所述ECG数据估计存在心脏功能异常时，向预定警报接收设备拨打警报电话。
  16. 根据权利要求1至12中任一项所述的心电图异常监测方法，其特征在于，所述方法还包括：
    当所述ECG异常监测模型的输出值指示所述ECG数据估计存在心脏功能异常时，对所述用户进行ECG测量，以得到指示所述用户的ECG是否指示存在心脏功能异常的测量结果。
  17. 一种心电图异常监测装置，其特征在于，所述装置包括：
    数据获取模块，用于获取用户的PPG数据；
    模型监测模块，用于将所述PPG数据输入预先训练好的ECG异常监测模型中，以得到所述ECG异常监测模型的输出值；其中，所述ECG异常监测模型被预先训练成用于基于输入的PPG数据，输出指示所输入的所述PPG数据对应的ECG数据是否估计存在心脏功能异常的输出值；
    异常警报模块，用于当所述ECG异常监测模型的输出值指示所述ECG数据估计存在心脏功能异常时，使与所述用户相关联的用户设备发出警报。
  18. 一种心电图异常监测设备，包括PPG数据探测器、存储器和处理器；
    所述PPG数据探测器用于监测用户的PPG数据，并将监测到的所述PPG数据传输至所述处理器；
    所述存储器存储有计算机程序；并且
    所述处理器执行所述计算机程序时，用于从所述PPG数据探测器接收所述PPG数据，并实现权利要求1至16中任一项所述的ECG异常监测方法。
  19. 根据权利要求18所述的心电图异常监测设备，其特征在于，还包括输出装置，所述输出装置包括显示屏、指示灯、振动器和/或声音输出装置。
  20. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，所述计算机程序被处理器执行时实现权利要求1至16中任一项所述的心电图异常监测方法。
PCT/CN2024/119209 2023-07-18 2024-09-14 心电图异常监测方法、装置、设备和存储介质 Pending WO2025016490A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202310881402.6 2023-07-18
CN202310881402.6A CN116584952B (zh) 2023-07-18 2023-07-18 心电图异常监测方法、装置、设备和存储介质

Publications (1)

Publication Number Publication Date
WO2025016490A1 true WO2025016490A1 (zh) 2025-01-23

Family

ID=87612065

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2024/119209 Pending WO2025016490A1 (zh) 2023-07-18 2024-09-14 心电图异常监测方法、装置、设备和存储介质

Country Status (2)

Country Link
CN (2) CN119326418A (zh)
WO (1) WO2025016490A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120823985A (zh) * 2025-09-16 2025-10-21 南京理工大学 基于深度学习的植入心室辅助装置运行监测方法

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN119326418A (zh) * 2023-07-18 2025-01-21 牛津大学(苏州)科技有限公司 心电图异常监测方法、装置、设备和存储介质

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013036718A1 (en) * 2011-09-08 2013-03-14 Isis Innovation Ltd. Determining acceptability of physiological signals
US20200100693A1 (en) * 2017-10-03 2020-04-02 Salutron, Inc. Arrhythmia monitoring using photoplethysmography
US20200196897A1 (en) * 2018-12-20 2020-06-25 Imec Vzw Method for generating a model for generating a synthetic ecg and a method and system for analysis of heart activity
CN113168908A (zh) * 2018-10-05 2021-07-23 阿利弗克公司 使用移动装置持续监测用户健康
EP3918980A1 (en) * 2020-06-02 2021-12-08 CSEM Centre Suisse d'Electronique et de Microtechnique SA - Recherche et Développement Method for classifying photoplethysmography pulses and monitoring of cardiac arrhythmias
CN114118386A (zh) * 2021-12-31 2022-03-01 浙江工业大学 一种基于注意力机制和时域卷积神经网络的血压估计方法
CN114521900A (zh) * 2022-02-21 2022-05-24 贵州省人民医院 一种基于迁移学习心律失常分类识别的方法
CN116584952A (zh) * 2023-07-18 2023-08-15 牛津大学(苏州)科技有限公司 心电图异常监测方法、装置、设备和存储介质

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111839495B (zh) * 2020-07-30 2023-04-07 深圳前海微众银行股份有限公司 检测方法、设备和存储介质
KR102373778B1 (ko) * 2021-07-29 2022-03-15 서울대학교산학협력단 심전도 데이터로부터 광용적맥파 데이터를 생성하는 장치 및 방법
CN114391824A (zh) * 2022-01-18 2022-04-26 安徽大学 一种基于ppg信号的心率估计方法、电子设备及存储介质
CN115270848B (zh) * 2022-06-17 2023-09-29 合肥心之声健康科技有限公司 一种ppg与ecg自动转换智能算法、存储介质和计算机系统

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2013036718A1 (en) * 2011-09-08 2013-03-14 Isis Innovation Ltd. Determining acceptability of physiological signals
US20200100693A1 (en) * 2017-10-03 2020-04-02 Salutron, Inc. Arrhythmia monitoring using photoplethysmography
CN113168908A (zh) * 2018-10-05 2021-07-23 阿利弗克公司 使用移动装置持续监测用户健康
US20200196897A1 (en) * 2018-12-20 2020-06-25 Imec Vzw Method for generating a model for generating a synthetic ecg and a method and system for analysis of heart activity
EP3918980A1 (en) * 2020-06-02 2021-12-08 CSEM Centre Suisse d'Electronique et de Microtechnique SA - Recherche et Développement Method for classifying photoplethysmography pulses and monitoring of cardiac arrhythmias
CN114118386A (zh) * 2021-12-31 2022-03-01 浙江工业大学 一种基于注意力机制和时域卷积神经网络的血压估计方法
CN114521900A (zh) * 2022-02-21 2022-05-24 贵州省人民医院 一种基于迁移学习心律失常分类识别的方法
CN116584952A (zh) * 2023-07-18 2023-08-15 牛津大学(苏州)科技有限公司 心电图异常监测方法、装置、设备和存储介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN120823985A (zh) * 2025-09-16 2025-10-21 南京理工大学 基于深度学习的植入心室辅助装置运行监测方法
CN120823985B (zh) * 2025-09-16 2025-12-05 南京理工大学 基于深度学习的植入心室辅助装置运行监测方法

Also Published As

Publication number Publication date
CN116584952A (zh) 2023-08-15
CN119326418A (zh) 2025-01-21
CN116584952B (zh) 2023-12-08

Similar Documents

Publication Publication Date Title
US12471786B2 (en) Methods and systems for arrhythmia tracking and scoring
US20240164685A1 (en) Health monitoring device
CN108577830B (zh) 一种面向用户的体征信息动态监护方法和动态监护系统
WO2025016490A1 (zh) 心电图异常监测方法、装置、设备和存储介质
US10925495B2 (en) Apparatus and method for determining a blood pressure of a subject
CN108309263A (zh) 多参数监护数据分析方法和多参数监护系统
JP2007508076A (ja) 生理学的状態を診断する装置及び方法
EP3716846A1 (en) Continuous detection and monitoring of heart arrhythmia using both wearable sensors and cloud-resident analyses
US20230165506A1 (en) Apparatus for non-invasive acquisition of maternal and/or fetal physiological signals
Skrivanos et al. Home healthcare technologies and services: heart-rate fetus monitoring system using an MCU ESP8266 Node
US20230200746A1 (en) System for acquisition and analysis of maternal and/or fetal physiological signals
Randazzo et al. PulsECG-A Cuffless Non-Invasive Blood Pressure Monitoring Device through Neural Network Analysis of ECG and PPG signals
Dasarapalli et al. Three-Lead ECG and Heartrate Variability Using AD8232
Jain Smart wearable Cardio-health monitoring system using Deep Learning Technologies.
KR102905592B1 (ko) 딥러닝 기반 심전도를 이용한 심정지 위험도 예측 시스템
Sood et al. Heartbeat on the Go: portable ECG machine
Nguyen¹ et al. Developing a Mobile Vital Sign Monitoring System with Deep Learning Algorithms
CN110236522A (zh) 基于单导联心电图的健康筛查方法、系统及医疗设备

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: 24842475

Country of ref document: EP

Kind code of ref document: A1