[go: up one dir, main page]

WO2023004572A1 - 模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质 - Google Patents

模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质 Download PDF

Info

Publication number
WO2023004572A1
WO2023004572A1 PCT/CN2021/108605 CN2021108605W WO2023004572A1 WO 2023004572 A1 WO2023004572 A1 WO 2023004572A1 CN 2021108605 W CN2021108605 W CN 2021108605W WO 2023004572 A1 WO2023004572 A1 WO 2023004572A1
Authority
WO
WIPO (PCT)
Prior art keywords
target
model
task model
parameters
task
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/CN2021/108605
Other languages
English (en)
French (fr)
Inventor
张春会
张振中
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.)
BOE Technology Group Co Ltd
Original Assignee
BOE Technology Group 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 BOE Technology Group Co Ltd filed Critical BOE Technology Group Co Ltd
Priority to PCT/CN2021/108605 priority Critical patent/WO2023004572A1/zh
Priority to US17/772,405 priority patent/US20240188895A1/en
Priority to CN202180002006.0A priority patent/CN115885279A/zh
Publication of WO2023004572A1 publication Critical patent/WO2023004572A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A—HUMAN NECESSITIES
    • A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00—Measuring for diagnostic purposes; Identification of persons
    • A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235—Details of waveform analysis
    • A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Definitions

  • the present disclosure relates to the field of computer technology, and in particular to a model training method, a signal recognition method, a device, a computing processing device, a computer program, and a computer readable medium.
  • Electrocardiogram is one of the effective inspection methods for clinical diagnosis of cardiovascular diseases. In recent years, the classification and recognition of abnormal ECG signals has received extensive research and attention.
  • the classification recognition method based on deep learning has the advantage of automatically extracting features, but deep learning generally has multiple hidden layers, the network structure is deep, and contains a large number of parameters that need to be trained. To train the model to the optimum requires a lot of training. data. When performing multi-classification model training, each ECG abnormality requires a large amount of balanced training data in order to achieve better classification results.
  • the present disclosure provides a model training method, including:
  • the training sample set includes a sample ECG signal and an abnormal label of the sample ECG signal, and the abnormal label includes a target abnormal label and at least one related abnormal label;
  • the sample ECG signal is input into a multi-task model, and according to the output of the multi-task model and the abnormal label, the multi-task model is trained based on a multi-task learning mechanism; wherein, the multi-task model includes a target task A model and at least one related task model, the target output of the target task model is the target abnormal label of the input sample ECG signal, and the target output of the related task model is the related abnormal label of the input sample ECG signal;
  • the trained target task model is determined as a target abnormality identification model, and the target abnormality identification model is used to identify target abnormalities in ECG signals input to the target abnormality identification model.
  • the step of training the multi-task model based on a multi-task learning mechanism includes: adjusting the parameters of each of the related task models, and parameters, to adjust the parameters of the target task model.
  • the step of adjusting the parameters of the target task model according to the parameters of the at least one related task model includes:
  • a regularization loss term is determined, and the regularization loss term is used to make the parameters of the target task model similar to the parameters of the at least one related task model change;
  • a first loss value is determined according to the regularized loss item, and parameters of the target task model are adjusted with the goal of minimizing the first loss value.
  • the step of determining the regularization loss item according to the parameters of the target task model and the parameters of the at least one related task model includes:
  • the regularized loss term is determined according to the following formula:
  • the R( ⁇ 1 , ⁇ 2 ,..., ⁇ M ) represents the regular loss item
  • the M represents the total number of the target task model and the related task model in the multi-task model
  • the ⁇ 1 represents a parameter of the target task model
  • the ⁇ 2 , . . . , ⁇ M represent parameters of each of the relevant task models
  • the ⁇ represents a preset parameter.
  • the step of adjusting the parameters of each of the relevant task models includes:
  • the step of determining the first loss value according to the regular loss term it also includes:
  • the step of determining the first loss value according to the regular loss term includes:
  • both the second loss function and the empirical loss function are cross-entropy loss functions.
  • the target task model and the related task model share a common feature extraction layer, and the common feature extraction layer is used to extract common features of the target anomaly and the related anomaly,
  • the step of adjusting the parameters of the target task model according to the parameters of the at least one related task model includes:
  • the step of adjusting the parameters of each of the relevant task models includes:
  • the sample ECG signal into the second related task model input the output of the second related task model and the second related abnormal label into the preset third loss function, obtain the third loss value, and minimize
  • the third loss value is the target, and the parameters of the second related task model are adjusted, and the parameters of the second related task model include the parameters of the common feature extraction layer, wherein the second related task model is the Any one of the at least one related task model, the second related abnormal label is any one of the at least one related abnormal label;
  • the step of adjusting the parameters of the target task model after parameter sharing includes:
  • both the third loss function and the fourth loss function are cross-entropy loss functions.
  • the step of training the multi-task model based on a multi-task learning mechanism includes:
  • each round of iterative training includes: adjusting the parameters of each of the related task models, and according to the parameters of the at least one related task model, A step of adjusting the parameters of the target task model.
  • the step of adjusting the parameters of each of the relevant task models includes:
  • the step of adjusting the parameters of the target task model according to the parameters of the at least one related task model includes:
  • the present disclosure provides a signal identification method, including:
  • the target abnormality identification model Input the target ECG signal into the target abnormality identification model to obtain target abnormality identification result, the target abnormality identification result is used to indicate whether the target ECG signal has target abnormality; wherein, the target abnormality identification model adopts Obtained by the model training method described in any embodiment.
  • the present disclosure provides a model training device, including:
  • the sample acquisition module is configured to obtain a training sample set, the training sample set includes a sample ECG signal and an abnormal label of the sample ECG signal, and the abnormal label includes a target abnormal label and at least one related abnormal label;
  • the model training module is configured to input the sample ECG signal into a multi-task model, and train the multi-task model based on a multi-task learning mechanism according to the output of the multi-task model and the abnormal label; wherein, the The multi-task model includes a target task model and at least one related task model, the target output of the target task model is the target abnormal label of the input sample ECG signal, and the target output of the related task model is the input sample ECG signal related exception labels;
  • the model determination module is configured to determine the trained target task model as a target abnormality recognition model, and the target abnormality recognition model is used to identify target abnormalities in ECG signals input to the target abnormality recognition model.
  • the present disclosure provides a signal identification device, including:
  • the signal acquisition module is configured to acquire the target ECG signal
  • the abnormality identification module is configured to input the target ECG signal into the target abnormality identification model to obtain a target abnormality identification result, and the target abnormality identification result is used to indicate whether the target ECG signal has a target abnormality; wherein, the The target anomaly recognition model is trained by using the model training method described in any embodiment.
  • the present disclosure provides a computing processing device, including:
  • One or more processors when the computer readable code is executed by the one or more processors, the computing processing device executes the method described in any embodiment.
  • the present disclosure provides a computer program comprising computer readable codes which, when run on a computing processing device, cause the computing processing device to execute the method according to any one of the embodiments.
  • the present disclosure provides a computer-readable medium, in which the method described in any embodiment is stored.
  • Fig. 1 schematically shows a flow chart of a model training method
  • Fig. 2 schematically shows a flow chart of a signal identification method
  • FIG. 3 schematically shows another flow chart for training and obtaining a target anomaly recognition model
  • Fig. 4 schematically shows a schematic diagram of a soft parameter sharing multi-task model
  • Fig. 5 schematically shows a kind of two-channel neural network model
  • Fig. 6 schematically shows a schematic diagram of a hard parameter sharing multi-task model
  • Fig. 7 schematically shows a block diagram of a model training device
  • Fig. 8 schematically shows a block diagram of a signal identification device
  • Fig. 9 schematically shows a block diagram of a computing processing device for performing a method according to the present disclosure.
  • Fig. 10 schematically shows a storage unit for holding or carrying program codes for realizing the method according to the present disclosure.
  • Fig. 1 schematically shows a flow chart of a model training method. As shown in Fig. 1, the method may include the following steps.
  • Step S11 Obtain a training sample set.
  • the training sample set includes sample ECG signals and abnormal labels of the sample ECG signals.
  • the abnormal labels include target abnormal labels and at least one related abnormal label.
  • the execution subject of this embodiment may be a computer device, the computer device has a model training device, and the model training method provided in this embodiment is executed by the model training device.
  • the computer device may be, for example, a smart phone, a tablet computer, a personal computer, etc., which is not limited in this embodiment.
  • the execution subject of this embodiment can obtain the training sample set in various ways.
  • the execution subject may obtain the sample ECG signals stored therein from another server (such as a database server) for storing data through a wired connection or a wireless connection.
  • the execution subject may obtain sample ECG signals collected by a signal collection device such as an electrocardiograph, and store these sample ECG signals locally, thereby generating a training sample set.
  • Abnormalities in sample ECG signals may include: atrial premature beats, premature ventricular beats, supraventricular tachycardia, ventricular tachycardia, atrial flutter, atrial fibrillation, ventricular flutter, ventricular fibrillation, left bundle branch block, right bundle branch block, atrial Sexual escaping, ventricular escaping, tachycardia, bradycardia, atrioventricular block, ST-segment elevation, ST-segment depression, abnormal Brugada wave, giant R-wave ST-segment elevation, faux bundle branch block At least one of the lag exceptions.
  • the sample data of atrial flutter, Yabo, ST-segment elevation, ST-segment depression, abnormal Brugada wave, giant R-wave ST-segment elevation, and masquerading bundle branch block are relatively few.
  • the characteristics of the ECG signal may include waveform, peak, amplitude, frequency, amplitude, time and so on.
  • Some abnormal ECG signals have commonality or similarity in characteristics, which is reflected in the same part of the waveform characteristics or the same upper and lower limits of the frequency threshold and so on.
  • the lower limit of heart rate threshold for supraventricular tachycardia, paroxysmal tachycardia, atrial fibrillation, atrial flutter, and atrial tachycardia is 100 beats/min, while sinus tachycardia, atrioventricular conduction
  • the upper limit of abnormal heart rate thresholds such as heart block, sinoatrial block, and bundle branch block is 60 beats/min
  • Atrial flutter is associated with supraventricular tachycardia and sinus
  • the characteristics of abnormal ECG signals such as tachycardia are similar; (3) fake bundle branch block, left bundle branch block of limb lead ECG, right bundle branch block of precordial lead ECG; (4) relatively rare Brugada Wave abnormality was extracted in North America in 1991, and the abnormality showed ECG characteristics of right bundle branch block with ST-segment elevation in the right chest lead; (5) Giant R-wave ST-segment elevation first proposed by Wimalarlna in 1993 This abnormality has the waveform characteristics of the
  • the target anomaly indicated by the target anomaly label there is a correlation between the target anomaly indicated by the target anomaly label and the related anomaly indicated by the related anomaly label.
  • the target abnormality is atrial flutter
  • the related abnormality is atrial fibrillation.
  • each related abnormality is correlated with the target abnormality.
  • the number of sample electrocardiographic signals with any relevant abnormality in the training sample set may be greater than the number of sample electrocardiographic signals with a target abnormality.
  • the training sample set may include a plurality of sample ECG signals, assuming that these sample ECG signals involve M types of abnormalities, and the M types of abnormalities include the first abnormality, the second abnormality, . . . and the Mth abnormality. M is greater than or equal to 2. Assuming that the first anomaly is the target anomaly, any one of the second anomaly, the third anomaly, ... and the Mth anomaly is related to the first anomaly and is different from the first anomaly. Therefore, the second anomaly, Any one of the third anomaly, ... and the Mth anomaly may be a relevant anomaly.
  • the number of related abnormalities is one, that is, the second abnormality; when M ⁇ 3, the number of related abnormalities is multiple, and the multiple related abnormalities are respectively the second abnormality, the third abnormality, ... and the second abnormality M is abnormal.
  • the abnormal label of each sample ECG signal can be an M-dimensional vector, for example, the abnormal label of a certain sample ECG signal is [1,0,0,1,1,...,1 ], 1 in the abnormal label means that the sample ECG signal has a corresponding abnormality, 0 means that there is no corresponding abnormality, and the above abnormal label means that the sample ECG signal has the first abnormality, the fourth abnormality, the fifth abnormality, .. . and the Mth exception.
  • the sample ECG signals in the training sample set can be classified according to the abnormal type, and the sample ECG signals corresponding to the abnormal i after classification can include two groups: the positive sample with the i-th abnormality and the non-positive sample Negative samples with i-th anomaly.
  • i can be greater than or equal to 1 and less than or equal to M.
  • the number of positive samples and negative samples can be equal or relatively close.
  • the ratio of positive samples to negative samples can be adjusted according to actual needs, which is not limited in this embodiment.
  • the sample ECG signal may be preprocessed before step S12, as shown in FIG. 3, to remove noise interference.
  • a band-pass filter can be used to remove 50 Hz power frequency interference in the sample ECG signal;
  • a low-pass filter can be used to remove 10-300 Hz myoelectric interference in the sample ECG signal;
  • a high-pass filter can be used to remove the baseline drift; and so on.
  • the sample ECG signals in the training sample set may also be divided into a training set and a test set according to a certain ratio such as 4:1, as shown in FIG. 3 , which is not limited in this embodiment.
  • Step S12 Input the sample ECG signal into the multi-task model, and train the multi-task model based on the multi-task learning mechanism according to the output of the multi-task model and abnormal labels; wherein, the multi-task model includes a target task model and at least one related task model , the target output of the target task model is the target abnormal label of the input sample ECG signal, and the target output of the related task model is the relevant abnormal label of the input sample ECG signal.
  • Step S13 The trained target task model is determined as the target abnormality recognition model, and the target abnormality recognition model is used to identify the target abnormality in the ECG signal input to the target abnormality recognition model.
  • the target task model and each related task model can be a neural network model with the same network structure, for example, it can be a convolutional neural network model (Convolutional Neural Networks, CNN) or a recurrent neural network model (Recurrent Neural Network, RNN) with the same network structure. ).
  • the target task model and each related task model can use a long short-term memory network (LSTM, Long Short-Term Memory) in a recurrent neural network model.
  • LSTM Long Short-Term Memory
  • the target task model and each related task model may also be models with different network structures, which is not limited in this embodiment.
  • Multi-task learning is an important machine learning method that aims to use related tasks to improve the generalization ability of the main task.
  • the relationship between tasks is captured by constraining the relationship between the model parameters of each task, so that the knowledge learned from related tasks with more training data is transferred to the main task with less training data.
  • Multi-task learning imposes certain constraints on the main task, that is, the parameters of the main task model are constrained by the parameters of related task models during the optimization process, so that when all tasks meet the convergence conditions, the main task model is equivalent to integrating all related task models The learned knowledge can thus improve the generalization ability of the main task.
  • the target abnormality is correlated with each related abnormality, that is, the electrocardiographic signal with the target abnormality and the electrocardiographic signal with any one of the related abnormalities have characteristics in common. Therefore, the target anomaly identification model used to identify target anomalies and the related anomaly identification model used to identify related anomalies can be trained using a multi-task learning mechanism.
  • the task of training the target abnormality recognition model is The main task, such as task1 in Figure 4 and Figure 6.
  • the main task is used to train the target task model in the multi-task model.
  • the target task model is the target anomaly recognition model.
  • the target anomaly recognition model is used to identify whether the ECG signal input to the target anomaly recognition model has a target anomaly. Such as the first exception.
  • the task of training the related anomaly recognition model is the related task
  • the related task is used to train the related task model
  • the related task model after the training is the related anomaly recognition model.
  • the number of related tasks and related task models is at least one, and since the number of related abnormalities is M ⁇ 1, correspondingly, the number of related tasks and related task models can be both M ⁇ 1.
  • the M-1 related tasks are task2,..., taskM, as shown in Figure 4 and Figure 6.
  • each related task model can be determined as a different related anomaly recognition model.
  • Each associated anomaly identification model can be used to identify different associated anomalies.
  • two methods may be used to perform multi-task learning on the target task model and at least one related task model, namely, hard parameter sharing and soft parameter sharing.
  • the hard parameter sharing is to share the hidden layer of the network between multiple task models, that is, between the relevant task model and the target task model.
  • the parameters of the hidden layer in multiple task models are the same, and the network output layer of each task model different to perform different tasks.
  • Soft parameter sharing means that each task has its own model and parameters, but the parameters of the main task model, that is, the target task model, are constrained by the parameters of the related task model to encourage parameter similarity between the target task model and the related task model . Subsequent embodiments will introduce in detail the detailed process of training a multi-task model based on a multi-task learning mechanism.
  • the parameters of the target task model are constrained by the parameters of the related task model, and the target task model is obtained based on the parameter training of the related task model , so as to realize the transfer of the knowledge (ie, parameters) learned by the related task model with more training data to the target task model with less training data. Due to the large number of sample ECG signals with related abnormalities, and the correlation between the target abnormality and the related abnormality, the generalization ability and classification recognition effect of the target abnormality recognition model trained by the multi-task learning mechanism are improved.
  • the model training method provided in this embodiment trains the target task model and related task models in the multi-task model through a multi-task learning mechanism, so that the target task model with less training data combines the knowledge learned by the related task model with more training data ( parameter), the target task model after training is the target anomaly recognition model, so it can improve the generalization ability and classification performance of the target anomaly recognition model, and can effectively solve the problem of the target anomaly recognition model caused by insufficient sample data with target anomalies. Problems with poor classification recognition.
  • the step of training the multi-task model based on the multi-task learning mechanism in step S12 may specifically include: first adjusting the parameters of each related task model, and then according to the parameters of at least one related task model, Adjust the parameters of the target task model.
  • the step of adjusting the parameters of the target task model according to the parameters of at least one related task model may include: determining the regularization loss item according to the parameters of the target task model and at least one parameter of the related task model, and the regularization loss item is used To make the parameters of the target task model similar to the parameters of at least one related task model; determine the first loss value according to the regularized loss item, and adjust the parameters of the target task model with the goal of minimizing the first loss value.
  • the regularization loss term can be determined according to the following formula:
  • R( ⁇ 1 , ⁇ 2 , . . . , ⁇ M ) ⁇ (
  • R( ⁇ 1 , ⁇ 2 ,..., ⁇ M ) represents the regularized loss item
  • M represents the total number of target task models and related task models in the multi-task model, that is, the number of tasks in multi-task training
  • ⁇ 1 Represents the parameters of the target task model
  • ⁇ 2 , ..., ⁇ M represent the parameters of each related task model
  • ⁇ represents the preset parameters.
  • ⁇ is a hyperparameter, and its value can be set according to the sample distribution and experience.
  • parameter constraints are imposed on the target task model by adding a regularized loss item to the loss function of the target task model.
  • the regularization loss item is determined according to the parameters of the relevant task model, the parameters of the target task model can be constrained by the relevant task model, so that the knowledge learned by the relevant task model with a large number of sample ECG signals can be transferred to In the target task model, the classification and recognition performance of the target task model is improved.
  • step of adjusting the parameters of each related task model in step 122 may specifically include:
  • the goal is to adjust the parameters of the first related task model, wherein the first related task model is any one of at least one related task model, and the first related abnormal label is any one of at least one related abnormal label.
  • the first relevant abnormal label is any one of at least one relevant abnormal label of the sample ECG signal input to the first relevant task model.
  • the step of determining the first loss value according to the regularized loss item in step S12 may also include: inputting the sample ECG signal into the target task model, inputting the output of the target task model and the target abnormal label into the preset The experience loss function of , get the experience loss term.
  • the empirical loss function may be a cross-entropy loss function.
  • the step of determining the first loss value according to the regularized loss term may include: calculating the sum of the empirical loss term and the regularized loss term to obtain the first loss value. Then the target task model can be trained with the goal of minimizing the first loss value.
  • the convolutional neural network can be used to establish a target task model C 1 with the same network structure and M-1 related task models C 2 , . . . , C M .
  • the experience loss item E can be calculated using the cross-entropy loss function.
  • the second loss function T2 of any related task model C i may be a cross-entropy loss function.
  • N represents the number of sample ECG signals in the training set
  • the inner summation is the loss function of a single sample ECG signal
  • the outer summation is the loss function of all sample ECG signals
  • the summation result is divided by With N, the average loss function is obtained.
  • t nk is a sign function, if the true category of the nth sample ECG signal is k, the value is 1, otherwise the value is 0.
  • y nk represents the output of the network model, that is, the probability that the nth sample ECG signal belongs to abnormal type k.
  • t nk represents the target abnormal label of the nth sample ECG signal
  • the value of t nk is 1, otherwise the value is 0.
  • y nk represents the output of the target task model C 1 , that is, the probability that the nth sample ECG signal has a target abnormality.
  • t nk represents the first correlation abnormality label of the nth sample ECG signal
  • the value of t nk is 1, otherwise the value is 0.
  • y nk represents the output of the first correlation task model Ci , that is, the probability that the nth sample ECG signal has the first correlation abnormality.
  • the process of training the neural network model mainly includes: forward propagation to calculate the actual output, backpropagation to calculate the error and optimize the loss function, and use the gradient descent algorithm to update and adjust the model parameters layer by layer.
  • the minimum error and the optimal loss function are obtained through multiple iterative training, and the training of the neural network model is completed.
  • the step of training the multi-task model based on the multi-task learning mechanism may be implemented in various manners.
  • multiple rounds of iterative training can be performed on the multi-task model based on the multi-task learning mechanism; wherein, each round of iterative training includes: adjusting the parameters of each related task model, and according to at least one related task model parameters, a step of adjusting the parameters of the target task model.
  • the sample ECG signal in each iteration period, for each of at least one related task model, can be input to the related task model first, and the output of the related task model and the corresponding related abnormal label can be input To the preset second loss function, obtain the second loss value, and adjust the parameters of the relevant task model with the goal of minimizing the second loss value.
  • the sample ECG signal can be input into the target task model, and the output of the target task model and the target abnormal label can be input into the preset experience loss function to obtain the experience loss item.
  • the parameters of the task model and at least one (such as all) related task model adjusted parameters determine the regular loss term; then calculate the sum of the experience loss term and the regular loss term to obtain the first loss value; then the first loss can be minimized
  • the value is the target, and the parameters in the target task model are adjusted, and an iterative cycle is completed so far. Carry out multiple rounds of iterations in sequence according to the above process, until the iteration stop condition is met (such as the number of iterations reaches the set number, convergence, etc.), the training of the target task model and related task models can be completed, and the target task model after training is determined as
  • the target anomaly identification model is to determine each related task model after training as a different related anomaly identification model.
  • multiple rounds of iterative adjustment can be performed on the parameters of each related task model, until each related task model meets the corresponding training stop condition, and each related task model after training is determined to be different The related anomaly identification models; then the parameters of the target task model can be adjusted according to the parameters of at least one related anomaly identification model.
  • the sample ECG signal can be input to the related task model first, and the output of the first related task model and the first related abnormal label can be input to the preset second A loss function is used to obtain a second loss value, aiming at minimizing the second loss value, performing multiple rounds of iterative training on the related task model, and training to obtain a related abnormality recognition model.
  • the parameters of the target task model and at least one (such as all) parameters of the relevant abnormality recognition model determine the regular loss term; then calculate the sum of the experience loss term and the regular loss term to obtain the first loss value; then the first loss can be minimized
  • the value is the target, the target task model is trained, and the target task model after the training is determined as the target anomaly recognition model.
  • step S12 The following describes the process of training the multi-task model by adopting the hard parameter sharing method in step S12.
  • FIG. 6 it shows a schematic diagram of training a multi-task model by adopting the method of hard parameter sharing.
  • the target task model and the related task model share a common feature extraction layer, which is used to extract the common features of the target anomaly and related anomalies.
  • the target The step of adjusting the parameters of the task model may include: sharing the parameters of the common feature extraction layer in at least one related task model as the parameters of the common feature extraction layer in the target task model, and adjusting the parameters of the target task model after parameter sharing .
  • the target task model and each related task model are dual-channel deep learning models as shown in FIG. 5 .
  • Either one of the target task model and at least one related task model includes a private feature extraction layer and a shared feature extraction layer, the private feature extraction layer is used to extract private features, and the shared feature extraction layer is used to extract shared features.
  • each task model can identify different abnormalities and achieve specificity; by setting the common feature extraction layer, the knowledge learned by the relevant task model can be transferred to the target task model, thereby improving the classification and recognition of the target task model performance.
  • step S12 may include:
  • the second related task model input the sample ECG signal into the second related task model, input the output of the second related task model and the second related abnormal label into the preset third loss function, obtain the third loss value, and minimize the third loss
  • the value is the target
  • adjust the parameters of the second related task model, the parameters of the second related task model include the parameters of the common feature extraction layer, wherein, the second related task model is any one of at least one related task model, and the second related task model is abnormal label is any one of at least one associated exception label.
  • the second relevant abnormal label is any one of at least one relevant abnormal label of the sample ECG signal input into the second relevant task model.
  • the parameters of the private feature extraction layer in the target task model are ⁇ 1
  • the parameters of the private feature extraction layer in at least one related task model are ⁇ 2 ,..., ⁇ M
  • the target task model and the related task model share the feature extraction layer
  • the parameter is ⁇ 0 .
  • a convolutional neural network can be used to establish a target task model and M-1 related task models with the same network structure.
  • the third loss function and the fourth loss function may be cross-entropy loss functions. Calculated as follows:
  • N represents the number of sample ECG signals in the training set
  • the inner summation is the loss function of a single sample ECG signal
  • the outer summation is the loss function of all sample ECG signals
  • the summation result is divided by With N, the average loss function is obtained.
  • t nk is a sign function, if the true category of the nth sample ECG signal is k, the value is 1, otherwise the value is 0.
  • y nk represents the output of the network model, that is, the probability that the nth sample ECG signal belongs to abnormal type k.
  • t nk represents the second correlation abnormality label of the nth sample ECG signal
  • the value of t nk is 1, otherwise the value is 0.
  • y nk represents the output of the second correlation task model, that is, the probability that the nth sample ECG signal has a second correlation abnormality.
  • t nk represents the target abnormal label of the nth sample ECG signal
  • the value of t nk is 1, otherwise the value is 0.
  • y nk represents the output of the target task model, that is, the probability that the nth sample ECG signal has a target abnormality.
  • the process of training the neural network model mainly includes: forward propagation to calculate the actual output, backpropagation to calculate the error and optimize the loss function, and use the gradient descent algorithm to update and adjust the model parameters layer by layer.
  • the minimum error and the optimal loss function are obtained through multiple iterative training, and the training of the neural network model is completed.
  • the step of training the multi-task model based on the multi-task learning mechanism may be implemented in various manners.
  • multiple rounds of iterative training can be performed on the multi-task model based on the multi-task learning mechanism; wherein, each round of iterative training includes: adjusting the parameters of each related task model, and according to at least one related task model parameters, a step of adjusting the parameters of the target task model.
  • the sample ECG signal in each iteration period, for each of at least one related task model, can be input to the related task model first, and the output of the related task model and the corresponding related abnormal label can be input To a preset third loss function, obtain a third loss value, and adjust parameters of the relevant task model with the goal of minimizing the third loss value.
  • the parameters of the common feature extraction layer in at least one related task model can be shared as the parameters of the common feature extraction layer in the target task model; after that, the sample ECG input parameters can be shared
  • the target task model of the target task model, the output of the target task model and the target anomaly label are input to the preset fourth loss function to obtain the fourth loss value; after that, the target task model can be minimized with the goal of minimizing the fourth loss value, and the parameters in the target task model Make adjustments to complete an iteration cycle.
  • the training of the target task model and each related task model can be completed, and the target task model after the training is determined as the target anomaly
  • the recognition model is to determine each related task model after training as a different related abnormality recognition model.
  • multiple rounds of iterative adjustments can be performed on the parameters of each related task model, until the related task model meets the corresponding training stop condition, and each related task model after training is determined to be different Related anomaly identification models; parameters of the target task model can then be adjusted according to parameters of at least one related anomaly identification model.
  • the sample ECG signal can be first input to the relevant task model, and the output of the relevant task model and the corresponding relevant abnormal label can be input to the preset third loss function to obtain a third loss value, and with the goal of minimizing the third loss value, perform multiple rounds of iterative training on the related task model, and obtain a related abnormality recognition model through training.
  • the parameters of the common feature extraction layer in at least one related abnormality recognition model can then be shared as the parameters of the common feature extraction layer in the target task model; after that, the sample ECG signal can be input
  • the target task model after parameter sharing, input the output of the target task model and the target anomaly label to the preset fourth loss function to obtain the fourth loss value; after that, the target task model can be minimized with the goal of minimizing the fourth loss value Carry out training, and determine the target task model after training as the target anomaly recognition model.
  • Fig. 2 schematically shows a flow chart of a signal identification method, as shown in Fig. 2, the method may include the following steps.
  • Step S21 Acquiring target ECG signals.
  • this step may specifically include the following steps: firstly acquire the original ECG signal; then perform preprocessing on the original ECG signal to obtain the target ECG signal.
  • the execution subject of this embodiment may be a computer device, and the computer device has a signal recognition device, and the signal recognition method provided by this embodiment is executed by the signal recognition device.
  • the computer device may be, for example, a smart phone, a tablet computer, a personal computer, etc., which is not limited in this embodiment.
  • the executive subject of this embodiment can obtain the original ECG signal in various ways.
  • the execution subject may acquire the original ECG signal collected by a signal acquisition device such as an electrocardiograph, and then perform preprocessing on the acquired original ECG signal to obtain a target ECG signal.
  • the format of the target ECG signal can be made the same as that of the input sample ECG signal when training the target abnormality recognition model.
  • the step of preprocessing the original ECG signal may include at least one of the following steps: using a band-pass filter to remove power frequency interference in the original ECG signal; using a low-pass filter EMG interference in the original ECG signal is removed; and, baseline drift in the original ECG signal is removed by using a high-pass filter.
  • a band-pass filter can be used to remove 50 Hz power frequency interference; a low-pass filter can be used to remove 10-300 Hz myoelectric interference; and a high-pass filter can be used to remove baseline drift.
  • the noise interference in the original ECG signal can be removed, and the accuracy of classification and recognition can be improved.
  • Step S22 Input the target ECG signal into the target abnormality recognition model to obtain the target abnormality recognition result, which is used to indicate whether the target ECG signal has a target abnormality; wherein, the target abnormality recognition model adopts any embodiment Trained by the model training method.
  • the target electrocardiogram signal can be input into the target abnormality identification model, and the target abnormality identification result can be output. Whether the target ECG signal has target abnormality can be determined according to the output target abnormality recognition result.
  • the target abnormality identification result may include, for example: the probability of the target ECG signal having the target abnormality and the probability of not having the target abnormality, which is not limited in this embodiment.
  • the target anomaly recognition model may be pre-trained, or may be trained during a signal recognition process, which is not limited in this embodiment.
  • the target anomaly recognition model is a model trained with the related anomaly recognition model based on a multi-task learning mechanism
  • the target anomaly recognition model with less training data is fused with the related anomaly recognition model with more training data
  • the learned knowledge that is, parameters
  • Fig. 7 schematically shows a block diagram of a model training device. Referring to Figure 7, may include:
  • the sample acquisition module 71 is configured to obtain a training sample set, the training sample set includes a sample ECG signal and an abnormal label of the sample ECG signal, and the abnormal label includes a target abnormal label and at least one related abnormal label;
  • the model training module 72 is configured to input the sample ECG signal into a multi-task model, and train the multi-task model based on a multi-task learning mechanism according to the output of the multi-task model and the abnormal label; wherein,
  • the multi-task model includes a target task model and at least one related task model, the target output of the target task model is the target abnormal label of the input sample ECG signal, and the target output of the related task model is the input sample ECG The associated exception label for the signal;
  • the model determining module 73 is configured to determine the trained target task model as a target abnormality identification model, and the target abnormality identification model is used to identify target abnormalities in ECG signals input to the target abnormality identification model.
  • Fig. 8 schematically shows a block diagram of a signal identification device. Referring to Figure 8, may include:
  • the signal acquisition module 81 is configured to acquire the target ECG signal
  • the abnormality identification module 82 is configured to input the target ECG signal into the target abnormality identification model to obtain a target abnormality identification result, and the target abnormality identification result is used to indicate whether the target ECG signal has a target abnormality; wherein, The target anomaly recognition model is trained by using the model training method described in any embodiment.
  • the device embodiments described above are only illustrative, and the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network elements. Part or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. It can be understood and implemented by those skilled in the art without any creative efforts.
  • the various component embodiments of the present disclosure may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof.
  • a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all functions of some or all components in the computing processing device according to the embodiments of the present disclosure.
  • DSP digital signal processor
  • the present disclosure can also be implemented as an apparatus or apparatus program (eg, computer program and computer program product) for performing a part or all of the methods described herein.
  • Such a program realizing the present disclosure may be stored on a computer-readable medium, or may have the form of one or more signals.
  • Such a signal may be downloaded from an Internet site, or provided on a carrier signal, or provided in any other form.
  • FIG. 9 illustrates a computing processing device that may implement methods according to the present disclosure.
  • the computing processing device conventionally includes a processor 1010 and a computer program product or computer readable medium in the form of memory 1020 .
  • Memory 1020 may be electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), EPROM, hard disk, or ROM.
  • the memory 1020 has a storage space 1030 for program code 1031 for performing any method steps in the methods described above.
  • the storage space 1030 for program codes may include respective program codes 1031 for respectively implementing various steps in the above methods. These program codes can be read from or written into one or more computer program products.
  • These computer program products comprise program code carriers such as hard disks, compact disks (CDs), memory cards or floppy disks.
  • Such a computer program product is typically a portable or fixed storage unit as described with reference to FIG. 10 .
  • the storage unit may have storage segments, storage spaces, etc. arranged similarly to the memory 1020 in the computing processing device of FIG. 9 .
  • the program code can eg be compressed in a suitable form.
  • the storage unit includes computer readable code 1031', i.e. code readable by, for example, a processor such as 1010, which code, when executed by a computing processing device, causes the computing processing device to perform the above-described methods. each step.
  • references herein to "one embodiment,” “an embodiment,” or “one or more embodiments” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present disclosure. Additionally, please note that examples of the word “in one embodiment” herein do not necessarily all refer to the same embodiment.
  • any reference signs placed between parentheses shall not be construed as limiting the claim.
  • the word “comprising” does not exclude the presence of elements or steps not listed in a claim.
  • the word “a” or “an” preceding an element does not exclude the presence of a plurality of such elements.
  • the disclosure can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means can be embodied by one and the same item of hardware.
  • the use of the words first, second, and third, etc. does not indicate any order. These words can be interpreted as names.

Landscapes

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

Abstract

模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质。模型训练方法包括：获得训练样本集合，训练样本集合包括样本心电信号以及样本心电信号的异常标签，异常标签包括目标异常标签和至少一个相关异常标签；将样本心电信号输入多任务模型，根据多任务模型的输出以及异常标签，基于多任务学习机制对多任务模型进行训练；其中，多任务模型包括目标任务模型和至少一个相关任务模型，目标任务模型的目标输出为输入的样本心电信号的目标异常标签，相关任务模型的目标输出为输入的样本心电信号的相关异常标签；将训练后的目标任务模型确定为目标异常识别模型，目标异常识别模型用于识别输入目标异常识别模型的心电信号中的目标异常。

Description

模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质 技术领域
本公开涉及计算机技术领域，特别是涉及一种模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质。
背景技术
心电图是临床上诊断心血管疾病的有效检查手段之一。近年来，异常心电信号的分类识别受到广泛的研究与关注。
基于深度学习的分类识别方法具有自动提取特征的优势，但是深度学习一般具有多个隐藏层，网络结构层次较深，含有大量需要训练的参数，想要将模型训练到最优就需要大量的训练数据。在进行多分类模型训练时，每种心电异常都需要有大量且均衡的训练数据，才能取得较好的分类效果。
概述
本公开提供了一种模型训练方法，包括：
获得训练样本集合，所述训练样本集合包括样本心电信号以及所述样本心电信号的异常标签，所述异常标签包括目标异常标签和至少一个相关异常标签；
将所述样本心电信号输入多任务模型，根据所述多任务模型的输出以及所述异常标签，基于多任务学习机制对所述多任务模型进行训练；其中，所述多任务模型包括目标任务模型和至少一个相关任务模型，所述目标任务模型的目标输出为输入的样本心电信号的目标异常标签，所述相关任务模型的目标输出为输入的样本心电信号的相关异常标签；
将训练后的所述目标任务模型确定为目标异常识别模型，所述目标异常识别模型用于识别输入所述目标异常识别模型的心电信号中的目标异常。
在一种可选的实现方式中，所述基于多任务学习机制对所述多任务模型进行训练的步骤，包括：调整各所述相关任务模型的参数，并根据所述至少 一个相关任务模型的参数，对所述目标任务模型的参数进行调整。
在一种可选的实现方式中，所述根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤，包括：
根据所述目标任务模型的参数以及所述至少一个相关任务模型的参数，确定正则损失项，所述正则损失项用于使所述目标任务模型的参数与所述至少一个相关任务模型的参数相似化；
根据所述正则损失项确定第一损失值，并以最小化所述第一损失值为目标，调整所述目标任务模型的参数。
在一种可选的实现方式中，所述根据所述目标任务模型的参数以及所述至少一个相关任务模型的参数，确定正则损失项的步骤，包括：
按照以下公式确定所述正则损失项：
R(θ 1，θ 2，...，θ M)＝λ(|θ 1－θ 2| 2+...+|θ 1－θ M| 2)
其中，所述R(θ 1，θ 2，...，θ M)表示所述正则损失项，所述M表示所述多任务模型中所述目标任务模型以及所述相关任务模型的总数量，所述θ 1表示所述目标任务模型的参数，所述θ 2，...，θ M分别表示各所述相关任务模型的参数，所述λ表示预设参量。
在一种可选的实现方式中，所述调整各所述相关任务模型的参数的步骤，包括：
将所述样本心电信号输入第一相关任务模型，将所述第一相关任务模型的输出以及第一相关异常标签输入至预设的第二损失函数，获得第二损失值，以最小化所述第二损失值为目标，调整所述第一相关任务模型的参数，其中，所述第一相关任务模型为所述至少一个相关任务模型中的任意一个，所述第一相关异常标签为所述至少一个相关异常标签中的任意一个；
在所述根据所述正则损失项确定第一损失值的步骤之前，还包括：
将所述样本心电信号输入所述目标任务模型，将所述目标任务模型的输出以及所述目标异常标签输入至预设的经验损失函数，获得；
所述根据所述正则损失项确定第一损失值的步骤，包括：
计算所述经验损失项以及所述正则损失项之和，获得所述第一损失值。
在一种可选的实现方式中，所述第二损失函数与所述经验损失函数均为交叉熵损失函数。
在一种可选的实现方式中，所述目标任务模型与所述相关任务模型共享有共有特征提取层，所述共有特征提取层用于提取所述目标异常与所述相关异常的共有特征，所述根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤，包括：
将所述至少一个相关任务模型中所述共有特征提取层的参数共享为所述目标任务模型中所述共有特征提取层的参数，对参数共享后的所述目标任务模型的参数进行调整。
在一种可选的实现方式中，所述调整各所述相关任务模型的参数的步骤，包括：
将所述样本心电信号输入第二相关任务模型，将所述第二相关任务模型的输出以及第二相关异常标签输入至预设的第三损失函数，获得第三损失值，并以最小化所述第三损失值为目标，调整所述第二相关任务模型的参数，所述第二相关任务模型的参数包括所述共有特征提取层的参数，其中，所述第二相关任务模型为所述至少一个相关任务模型中的任意一个，所述第二相关异常标签为所述至少一个相关异常标签中的任意一个；
所述对参数共享后的所述目标任务模型的参数进行调整的步骤，包括：
将所述样本心电信号输入参数共享后的所述目标任务模型，将所述目标任务模型的输出以及所述目标异常标签输入至预设的第四损失函数，获得第四损失值，并以最小化所述第四损失值为目标，调整所述目标任务模型的参数。
在一种可选的实现方式中，所述第三损失函数与所述第四损失函数均为交叉熵损失函数。
在一种可选的实现方式中，所述基于多任务学习机制对所述多任务模型进行训练的步骤，包括：
基于多任务学习机制，对所述多任务模型进行多轮迭代训练；其中，各轮迭代训练包括：所述调整各所述相关任务模型的参数，并根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤。
在一种可选的实现方式中，所述调整各所述相关任务模型的参数的步骤，包括：
对各所述相关任务模型的参数分别进行多轮迭代调整，直到各所述相关 任务模型满足对应的训练停止条件，将训练后的各所述相关任务模型确定为不同的相关异常识别模型；
所述根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤，包括：
根据至少一个所述相关异常识别模型的参数，对所述目标任务模型的参数进行调整。
本公开提供了一种信号识别方法，包括：
获取目标心电信号；
输入所述目标心电信号至目标异常识别模型中，获得目标异常识别结果，所述目标异常识别结果用于指示所述目标心电信号是否具有目标异常；其中，所述目标异常识别模型是采用任一实施例所述的模型训练方法训练得到的。
本公开提供了一种模型训练装置，包括：
样本获取模块，被配置为获得训练样本集合，所述训练样本集合包括样本心电信号以及所述样本心电信号的异常标签，所述异常标签包括目标异常标签和至少一个相关异常标签；
模型训练模块，被配置为将所述样本心电信号输入多任务模型，根据所述多任务模型的输出以及所述异常标签，基于多任务学习机制对所述多任务模型进行训练；其中，所述多任务模型包括目标任务模型和至少一个相关任务模型，所述目标任务模型的目标输出为输入的样本心电信号的目标异常标签，所述相关任务模型的目标输出为输入的样本心电信号的相关异常标签；
模型确定模块，被配置为将训练后的所述目标任务模型确定为目标异常识别模型，所述目标异常识别模型用于识别输入所述目标异常识别模型的心电信号中的目标异常。
本公开提供了一种信号识别装置，包括：
信号获取模块，被配置为获取目标心电信号；
异常识别模块，被配置为输入所述目标心电信号至目标异常识别模型中，获得目标异常识别结果，所述目标异常识别结果用于指示所述目标心电信号是否具有目标异常；其中，所述目标异常识别模型是采用任一实施例所述的模型训练方法训练得到的。
本公开提供了一种计算处理设备，包括：
存储器，其中存储有计算机可读代码；
一个或多个处理器，当所述计算机可读代码被所述一个或多个处理器执行时，所述计算处理设备执行任一实施例所述的方法。
本公开提供了一种计算机程序，包括计算机可读代码，当所述计算机可读代码在计算处理设备上运行时，导致所述计算处理设备执行根据任一实施例所述的方法。
本公开提供了一种计算机可读介质，其中存储了任一实施例所述的方法。
上述说明仅是本公开技术方案的概述，为了能够更清楚了解本公开的技术手段，而可依照说明书的内容予以实施，并且为了让本公开的上述和其它目的、特征和优点能够更明显易懂，以下特举本公开的具体实施方式。
附图简述
为了更清楚地说明本公开实施例或相关技术中的技术方案，下面将对实施例或相关技术描述中所需要使用的附图作一简单地介绍，显而易见地，下面描述中的附图是本公开的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。需要说明的是，附图中的比例仅作为示意并不代表实际比例。
图1示意性地示出了一种模型训练方法的流程图；
图2示意性地示出了一种信号识别方法的流程图；
图3示意性地示出了另一种训练获得目标异常识别模型的流程图；
图4示意性地示出了一种软参数共享的多任务模型示意图；
图5示意性地示出了一种双通道神经网络模型；
图6示意性地示出了一种硬参数共享的多任务模型示意图；
图7示意性地示出了一种模型训练装置的框图；
图8示意性地示出了一种信号识别装置的框图；
图9示意性地示出了用于执行根据本公开的方法的计算处理设备的框图。
图10示意性地示出了用于保持或者携带实现根据本公开的方法的程序代码的存储单元。
详细描述
为使本公开实施例的目的、技术方案和优点更加清楚，下面将结合本公开实施例中的附图，对本公开实施例中的技术方案进行清楚、完整地描述，显然，所描述的实施例是本公开一部分实施例，而不是全部的实施例。基于本公开中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本公开保护的范围。
图1示意性地示出了一种模型训练方法的流程图，如图1所示，该方法可以包括以下步骤。
步骤S11：获得训练样本集合，训练样本集合包括样本心电信号以及样本心电信号的异常标签，异常标签包括目标异常标签和至少一个相关异常标签。
本实施例的执行主体可以为计算机设备，该计算机设备具有模型训练装置，通过该模型训练装置来执行本实施例提供的模型训练方法。其中，计算机设备例如可以为智能手机、平板电脑、个人计算机等，本实施例对此不作限定。
本实施例的执行主体可以通过多种方式来获取训练样本集合。例如，执行主体可以通过有线连接方式或无线连接的方式，从用于存储数据的另一服务器(例如数据库服务器)中获取存储于其中的样本心电信号。再例如，执行主体可以获取由如心电图仪等信号采集设备所采集的样本心电信号，并将这些样本心电信号存储在本地，从而生成训练样本集合。
样本心电信号中的异常可以包括：房性早搏、室性早搏、室上速、室速、房扑、房颤、室扑、室颤、左束支阻滞、右束支阻滞、房性逸博、室性逸博、心动过速、心动过缓、房室传导阻滞、ST段抬高、ST段压低、Brugada波异常、巨R波型ST段抬高、伪装性束支阻滞等异常中的至少一个。其中房扑、逸博、ST段抬高、ST段压低、Brugada波异常、巨R波型ST段抬高、伪装性束支阻滞等异常的样本数据比较少。
心电信号的特征可以包括波形、波峰、波幅、频率、振幅、时间等等。某些异常的心电信号在特征上具有共性或相似性，体现在部分波形特征相同或者频率阈值上下限相同等等。例如：(1)室上性心动过速、阵发性心动过速、房颤、房扑、房速等异常的心率阈值下限均为100次/分，而窦性心动过速、房室传导阻滞、窦房传导阻滞、束支传导阻滞等异常的心率阈值上限均为60次/分；(2)房扑在心室率较快时，与室上性心动过速、窦性心动过速等 异常的心电信号特征类似；(3)伪装性束支阻滞，肢体导联心电图像左束支阻滞，心前导联心电图像右束支阻滞；(4)比较少见的Brugada波异常于1991年在北美被提取，该异常呈现出右束支阻滞伴右胸导联ST段抬高的心电图特征；(5)1993年由Wimalarlna首次提出的巨R波型ST段抬高这一异常，具有QRS与抬高的ST段、直立T波融合为一体的波形特征；(6)1938年首次发现的J波，又叫“Osborn”波，该异常形态上与QRS波群的一部分和第二个R波非常相似。
这些异常的心电信号的特征具有共性，因此这些异常具有相关性，满足多任务学习的条件，可以进行任务间知识的迁移。
在具体实现中，目标异常标签所指示的目标异常与相关异常标签所指示的相关异常之间具有相关性。例如目标异常为房扑，相关异常为房颤。当相关异常的数量为多个时，各相关异常均与目标异常具有相关性。训练样本集合中具有任意一种相关异常的样本心电信号的数量可以大于具有目标异常的样本心电信号的数量。
本实施例中，训练样本集合可以包含多个样本心电信号，假设这些样本心电信号涉及M种异常，M种异常包括第一异常、第二异常、……和第M异常。M大于或等于2。假设第一异常为目标异常，第二异常、第三异常、……和第M异常中的任意一种异常均与第一异常具有相关性，且不同于第一异常，因此，第二异常、第三异常、……和第M异常中的任意一种异常均可以为相关异常。
当M＝2时，相关异常的数量为一个，即第二异常；当M≥3时，相关异常的数量为多个，多个相关异常分别为第二异常、第三异常、……和第M异常。
在具体实现中，每一个样本心电信号的异常标签可以为M维度的向量，例如，某一样本心电信号的异常标签为[1,0,0,1,1,....,1]，异常标签中的1代表该样本心电信号具有对应的异常，0代表不具有对应的异常，上述异常标签表示该样本心电信号具有第一异常、第四异常、第五异常、...和第M异常。
在具体实现中，可以将训练样本集合中的样本心电信号按照异常类型进行归类，归类后与异常i对应的样本心电信号可以包含的两组：具有第i异常的正样本和不具有第i异常的负样本。其中，i可以大于等于1且小于或等于 M。正样本与负样本的数量可以相等或比较接近。在具体实现中，正样本与负样本的比例可以根据实际需求进行调整，本实施例对此不作限定。
在具体实现中，还可以在步骤S12之前首先对样本心电信号进行预处理，如图3所示，实现去除噪声干扰。具体地，可以用带通滤波器去除样本心电信号中50Hz的工频干扰；用低通滤波器去除样本心电信号中10-300Hz的肌电干扰；用高通滤波器去除样本心电信号中的基线漂移；等等。
在具体实现中，还可以将训练样本集合中的样本心电信号按照一定的比例如4：1，分为训练集和测试集，如图3所示，本实施例对此不作限定。
步骤S12：将样本心电信号输入多任务模型，根据多任务模型的输出以及异常标签，基于多任务学习机制对多任务模型进行训练；其中，多任务模型包括目标任务模型和至少一个相关任务模型，目标任务模型的目标输出为输入的样本心电信号的目标异常标签，相关任务模型的目标输出为输入的样本心电信号的相关异常标签。
步骤S13：将训练后的目标任务模型确定为目标异常识别模型，目标异常识别模型用于识别输入目标异常识别模型的心电信号中的目标异常。
其中，目标任务模型和各相关任务模型可以为网络结构相同的神经网络模型，例如可以为网络结构相同的卷积神经网络模型(Convolutional Neural Networks,CNN)或者循环神经网络模型(Recurrent Neural Network,RNN)。具体地，目标任务模型和各相关任务模型可以采用循环神经网络模型中的长短期记忆网络(LSTM，Long Short-Term Memory)。当然，目标任务模型和各相关任务模型也可以为网络结构不同的模型，本实施例对此不作限定。
多任务学习(Multi-task learning,MTL)是一种重要的机器学习方法，旨在使用相关任务来提升主任务的泛化能力。在多任务学习的过程中，通过约束各任务模型参数之间的关系来捕捉各任务之间的关系，从而将训练数据多的相关任务学习到的知识迁移到训练数据少的主任务上。多任务学习对主任务进行了一定的约束，即主任务模型参数在优化的过程中受到相关任务模型参数的约束，这样在所有任务满足收敛条件时，主任务模型相当于融合了所有相关任务模型学习到的知识，因此可以提升主任务的泛化能力。
本实施例中，由于目标异常与各相关异常具有相关性，即具有目标异常的心电信号与具有任意一种相关异常的心电信号在特征上具有共性。因此， 用于识别目标异常的目标异常识别模型与用于识别相关异常的相关异常识别模型可以采用多任务学习机制训练得到。
由于具有目标异常的样本心电信号的数量小于具有任意一种相关异常的样本心电信号的数量，因此在采用多任务学习机制训练目标异常识别模型的过程中，训练目标异常识别模型的任务为主任务，如图4和图6中的task1。主任务用于对多任务模型中的目标任务模型进行训练，训练完成后的目标任务模型为目标异常识别模型，目标异常识别模型用于识别输入该目标异常识别模型的心电信号是否具有目标异常如第一异常。
训练相关异常识别模型的任务为相关任务，相关任务用于对相关任务模型进行训练，训练完成后的相关任务模型为相关异常识别模型。在多任务学习中，相关任务以及相关任务模型的数量为至少一个，由于相关异常的数量为M-1个，相应地，相关任务以及相关任务模型的数量可以均为M-1个。M-1个相关任务分别为task2，…，taskM，如图4和图6所示。
具体地，若相关任务的数量为一个，该相关任务用于训练一个相关任务模型；若相关任务的数量为多个，每一个相关任务用于训练一个相关任务模型，各相关任务用于训练不同的相关任务模型，训练完成后的各相关任务模型可以确定为不同的相关异常识别模型。各相关异常识别模型可以用于识别不同的相关异常。
在具体实现中，对目标任务模型和至少一个相关任务模型执行多任务学习例如可以采用两种方法，即硬参数共享和软参数共享。其中，硬参数共享是通过在多个任务模型之间，即在相关任务模型和目标任务模型之间共享网络的隐藏层，多个任务模型中隐藏层的参数相同，各任务模型的网络输出层不同，从而执行不同的任务。软参数共享指的是每个任务具有自己的模型和参数，但是主任务模型即目标任务模型的参数会受到相关任务模型参数的约束，以鼓励目标任务模型与相关任务模型之间的参数相似化。后续实施例会详细介绍基于多任务学习机制训练多任务模型的详细过程。
本实施例中，在通过多任务学习机制训练目标任务模型和相关任务模型的过程中，目标任务模型的参数受到相关任务模型的参数的约束，目标任务模型是基于相关任务模型的参数训练得到的，从而实现将训练数据多的相关任务模型学习到的知识(即参数)迁移到训练数据少的目标任务模型中。由 于具有相关异常的样本心电信号数量较多，并且目标异常与相关异常具有相关性，因此，通过多任务学习机制训练得到的目标异常识别模型的泛化能力和分类识别效果得到提升。
本实施例提供的模型训练方法，通过多任务学习机制训练多任务模型中的目标任务模型和相关任务模型，使得训练数据少的目标任务模型融合了训练数据多的相关任务模型学习到的知识(即参数)，训练后的目标任务模型为目标异常识别模型，因此可以提升目标异常识别模型的泛化能力和分类性能，可以有效地解决因具有目标异常的样本数据不充足导致目标异常识别模型的分类识别效果较差的问题。
在一种可选的实现方式中，步骤S12中基于多任务学习机制对多任务模型进行训练的步骤，具体可以包括：首先调整各相关任务模型的参数，然后根据至少一个相关任务模型的参数，对目标任务模型的参数进行调整。
下面介绍在步骤S12中采用软参数共享的方法对多任务模型进行训练的过程。
参照图4示出了采用软参数共享的方法对多任务模型进行训练的示意图。步骤S12中根据至少一个相关任务模型的参数，对目标任务模型的参数进行调整的步骤，可以包括：根据目标任务模型的参数以及至少一个相关任务模型的参数，确定正则损失项，正则损失项用于使目标任务模型的参数与至少一个相关任务模型的参数相似化；根据正则损失项确定第一损失值，并以最小化第一损失值为目标，调整目标任务模型的参数。
在具体实现中，可以按照以下公式确定正则损失项：
R(θ 1，θ 2，...，θ M)＝λ(|θ 1－θ 2| 2+...+|θ 1－θ M| 2)。
其中，R(θ 1，θ 2，...，θ M)表示正则损失项；M表示多任务模型中目标任务模型以及相关任务模型的总数量，即多任务训练中的任务数量；θ 1表示目标任务模型的参数；θ 2，...，θ M分别表示各相关任务模型的参数；λ表示预设参量。λ为超参数，其数值可以根据样本分布情况和经验进行设定。通过确定正则损失项，可以促进目标任务模型的参数与相关任务模型的参数相似化。
本实现方式中，通过在目标任务模型的损失函数中加入正则损失项来对目标任务模型进行参数约束。并且由于正则损失项是根据相关任务模型的参 数确定得到的，因此目标任务模型的参数可以受到相关任务模型的约束，从而可以将样本心电信号数量较多的相关任务模型学习到的知识迁移到目标任务模型中去，提升目标任务模型的分类识别性能。
本实现方式中，在步骤122中调整各相关任务模型的参数的步骤，具体可以包括：
将样本心电信号输入第一相关任务模型，将第一相关任务模型的输出以及第一相关异常标签输入至预设的第二损失函数，获得第二损失值，以最小化第二损失值为目标，调整第一相关任务模型的参数，其中，第一相关任务模型为至少一个相关任务模型中的任意一个，第一相关异常标签为至少一个相关异常标签中的任意一个。第一相关异常标签为输入第一相关任务模型的样本心电信号的至少一个相关异常标签中的任意一个。
本实现方式中，在步骤S12中根据正则损失项确定第一损失值的步骤之前，还可以包括：将样本心电信号输入目标任务模型，将目标任务模型的输出以及目标异常标签输入至预设的经验损失函数，获得经验损失项。其中，经验损失函数可以为交叉熵损失函数。
进一步地，根据正则损失项确定第一损失值的步骤，可以包括：计算经验损失项以及正则损失项之和，获得第一损失值。之后可以以最小化第一损失值为目标，对目标任务模型进行训练。
在具体实现中，可以利用卷积神经网络建立网络结构相同的目标任务模型C 1以及M-1个相关任务模型C 2，...，C M。
其中，目标任务模型C 1可以用以下公式表示：Y 1＝f(θ 1，X 1)，其中，θ 1表示目标任务模型C 1的参数，X 1表示目标任务模型C 1的输入，Y 1表示目标任务模型C 1的输出。
任意一个相关任务模型(即上述的第一相关异常模型)C i可以用以下公式表示：Y i＝f(θ i，X i)，其中，θ i表示该相关任务模型C i的参数，X i表示该相关任务模型C i的输入，Y i表示该相关任务模型C i的输出。其中，2≤i≤M。
目标任务模型C 1的第一损失值T1可以为经验损失项E以及正则损失项R(θ 1，θ 2，...，θ M)之和，即T1＝E+R(θ 1，θ 2，...，θ M)。其中经验损失项E可以采用交叉熵损失函数计算得到。
任一个相关任务模型C i的第二损失函数T2可以为交叉熵损失函数。
其中，交叉熵损失函数的计算公式如下：
Figure PCTCN2021108605-appb-000001
上述公式中，N表示训练集中的样本心电信号数量，内层求和是单个样本心电信号的损失函数，外层求和是所有样本心电信号的损失函数，之后再对求和结果除以N，得到平均损失函数。t nk为符号函数，如果第n个样本心电信号的真实类别为k时，取值为1，否则取值为0。y nk表示网络模型的输出，即第n个样本心电信号属于异常类型k的概率。
在计算第一损失值T1中的经验损失项时，交叉熵损失函数中的类别k只有两种，即具有目标异常以及不具有目标异常。t nk表示第n个样本心电信号的目标异常标签，当第n个样本心电信号具有目标异常时，t nk的取值为1，否则取值为0。y nk表示目标任务模型C 1的输出，即第n个样本心电信号具有目标异常的概率。
在计算第一相关异常模型的第二损失函数T2时，交叉熵损失函数中的类别k只有两种，即具有第一相关异常以及不具有第一相关异常。t nk表示第n个样本心电信号的第一相关异常标签，当第n个样本心电信号具有第一相关异常时，t nk的取值为1，否则取值为0。y nk表示第一相关任务模型C i的输出，即第n个样本心电信号具有第一相关异常的概率。
在具体实现中，对神经网络模型进行训练的过程主要包括：前向传播计算实际输出，反向传播计算误差以及优化损失函数，利用梯度下降算法逐层进行模型参数的更新调整。通过多次迭代训练得到最小误差与最优损失函数，完成神经网络模型的训练。
本实施例中，基于多任务学习机制对多任务模型进行训练的步骤可以有多种实现方式。在一种可选的实现方式中，可以基于多任务学习机制，对多任务模型进行多轮迭代训练；其中，各轮迭代训练包括：调整各相关任务模型的参数，并根据至少一个相关任务模型的参数，对目标任务模型的参数进行调整的步骤。
本实现方式中，在每轮迭代周期内，对于至少一个相关任务模型中的每一个，可以首先输入样本心电信号至该相关任务模型，将该相关任务模型的输出以及对应的相关异常标签输入至预设的第二损失函数，获得第二损失值， 以最小化第二损失值为目标，调整该相关任务模型的参数。按照上述过程调整各相关任务模型的参数之后，可以将样本心电信号输入目标任务模型，将目标任务模型的输出以及目标异常标签输入至预设的经验损失函数，获得经验损失项，同时基于目标任务模型的参数以及至少一个(如所有)相关任务模型调整后的参数，确定正则损失项；之后计算经验损失项以及正则损失项之和，获得第一损失值；之后可以以最小化第一损失值为目标，对目标任务模型中的参数进行调整，至此完成一轮迭代周期。按照上述过程依次进行多轮迭代，直到满足迭代停止条件(如迭代次数达到设定次数、收敛等)，可以完成目标任务模型以及各相关任务模型的训练，将训练完成后的目标任务模型确定为目标异常识别模型，将训练完成后的各相关任务模型确定为不同的相关异常识别模型。
在另一种可选的实现方式中，可以首先对各相关任务模型的参数分别进行多轮迭代调整，直到各相关任务模型满足对应的训练停止条件，将训练后的各相关任务模型确定为不同的相关异常识别模型；之后可以根据至少一个相关异常识别模型的参数，对目标任务模型的参数进行调整。
本实现方式中，对于至少一个相关任务模型中的每一个，可以首先输入样本心电信号至该相关任务模型，将第一相关任务模型的输出以及第一相关异常标签输入至预设的第二损失函数，获得第二损失值，以最小化第二损失值为目标，对该相关任务模型进行多轮迭代训练，训练得到一个相关异常识别模型。按照上述过程调整训练得到各相关异常识别模型之后，可以将样本心电信号输入目标任务模型，将目标任务模型的输出以及目标异常标签输入至预设的经验损失函数，获得经验损失项，同时基于目标任务模型的参数以及至少一个(如所有)相关异常识别模型的参数，确定正则损失项；之后计算经验损失项以及正则损失项之和，获得第一损失值；之后可以以最小化第一损失值为目标，对目标任务模型进行训练，将训练完成后的目标任务模型确定为目标异常识别模型。
下面介绍在步骤S12中采用硬参数共享的方法对多任务模型进行训练的过程。
参照图6示出了采用硬参数共享的方法对多任务模型进行训练的示意图。如图6所示，目标任务模型与相关任务模型共享有共有特征提取层，共有特 征提取层用于提取目标异常与相关异常的共有特征，步骤S12中根据至少一个相关任务模型的参数，对目标任务模型的参数进行调整的步骤，可以包括：将至少一个相关任务模型中共有特征提取层的参数共享为目标任务模型中共有特征提取层的参数，对参数共享后的目标任务模型的参数进行调整。
本实现方式中，目标任务模型和各相关任务模型均为如图5所示的双通道深度学习模型。目标任务模型和至少一个相关任务模型中的任一个都包括私有特征提取层和共有特征提取层，私有特征提取层用于提取私有特征，共有特征提取层用于提取共有特征。通过设置私有特征提取层使得各任务模型可以识别不同异常，实现特异性；通过设置共有特征提取层，可以将相关任务模型学到的知识迁移到目标任务模型中，从而提升目标任务模型的分类识别性能。
本实现方式中，步骤S12可以包括：
首先将样本心电信号输入第二相关任务模型，将第二相关任务模型的输出以及第二相关异常标签输入至预设的第三损失函数，获得第三损失值，并以最小化第三损失值为目标，调整第二相关任务模型的参数，第二相关任务模型的参数包括共有特征提取层的参数，其中，第二相关任务模型为至少一个相关任务模型中的任意一个，第二相关异常标签为至少一个相关异常标签中的任意一个。第二相关异常标签为输入第二相关任务模型的样本心电信号的至少一个相关异常标签中的任意一个。
然后将至少一个相关任务模型中共有特征提取层的参数共享为目标任务模型中共有特征提取层的参数。
之后将样本心电信号输入参数共享后的目标任务模型，将目标任务模型的输出以及目标异常标签输入至预设的第四损失函数，获得第四损失值，并以最小化第四损失值为目标，调整目标任务模型的参数。
目标任务模型中私有特征提取层的参数为θ 1，至少一个相关任务模型中私有特征提取层的参数分别为θ 2，...，θ M，目标任务模型与相关任务模型共有特征提取层的参数为θ 0。
在具体实现中，可以利用卷积神经网络建立网络结构相同的目标任务模型和M-1个相关任务模型。
其中，目标任务模型可以用以下公式表示：Y 1＝f((θ 0，θ 1，X 1)，其中， X 1表示目标任务模型的输入，Y 1表示目标任务模型的输出。
任一个相关任务模型(即上述的第二相关任务模型)可以用以下公式表示：Y i＝f(θ 0，θ i，X i)，其中，X i表示该相关任务模型的输入，Y i表示该相关任务模型的输出。
其中，第三损失函数和第四损失函数可以为交叉熵损失函数。计算公式如下：
Figure PCTCN2021108605-appb-000002
上述公式中，N表示训练集中的样本心电信号数量，内层求和是单个样本心电信号的损失函数，外层求和是所有样本心电信号的损失函数，之后再对求和结果除以N，得到平均损失函数。t nk为符号函数，如果第n个样本心电信号的真实类别为k时，取值为1，否则取值为0。y nk表示网络模型的输出，即第n个样本心电信号属于异常类型k的概率。
在计算第二相关任务模型的第三损失函数时，交叉熵损失函数中的类别k只有两种，即具有第二相关异常以及不具有第二相关异常。t nk表示第n个样本心电信号的第二相关异常标签，当第n个样本心电信号具有第二相关异常时，t nk的取值为1，否则取值为0。y nk表示第二相关任务模型的输出，即第n个样本心电信号具有第二相关异常的概率。
在计算第四损失函数时，交叉熵损失函数中的类别k只有两种，即具有目标异常以及不具有目标异常。t nk表示第n个样本心电信号的目标异常标签，当第n个样本心电信号具有目标异常时，t nk的取值为1，否则取值为0。y nk表示目标任务模型的输出，即第n个样本心电信号具有目标异常的概率。
在具体实现中，对神经网络模型进行训练的过程主要包括：前向传播计算实际输出，反向传播计算误差以及优化损失函数，利用梯度下降算法逐层进行模型参数的更新调整。通过多次迭代训练得到最小误差与最优损失函数，完成神经网络模型的训练。
本实施例中，基于多任务学习机制对多任务模型进行训练的步骤可以有多种实现方式。在一种可选的实现方式中，可以基于多任务学习机制，对多任务模型进行多轮迭代训练；其中，各轮迭代训练包括：调整各相关任务模型的参数，并根据至少一个相关任务模型的参数，对目标任务模型的参数进 行调整的步骤。
本实现方式中，在每轮迭代周期内，对于至少一个相关任务模型中的每一个，可以首先输入样本心电信号至该相关任务模型，将该相关任务模型的输出以及对应的相关异常标签输入至预设的第三损失函数，获得第三损失值，并以最小化所述第三损失值为目标，调整该相关任务模型的参数。按照上述过程调整各相关任务模型的参数之后，可以将至少一个相关任务模型中共有特征提取层的参数共享为目标任务模型中共有特征提取层的参数；之后可以将样本心电信号输入参数共享后的目标任务模型，将目标任务模型的输出以及目标异常标签输入至预设的第四损失函数，获得第四损失值；之后可以以最小化第四损失值为目标，对目标任务模型中的参数进行调整，至此完成一轮迭代周期。按照上述过程依次进行多轮迭代，直到满足迭代停止条件(如迭代次数达到设定次数等)，可以完成目标任务模型以及各相关任务模型的训练，将训练完成后的目标任务模型确定为目标异常识别模型，将训练完成后的各相关任务模型确定为不同的相关异常识别模型。
在另一种可选的实现方式中，可以首先对各相关任务模型的参数分别进行多轮迭代调整，直到相关任务模型满足对应的训练停止条件，将训练后的各相关任务模型确定为不同的相关异常识别模型；之后可以根据至少一个相关异常识别模型的参数，对目标任务模型的参数进行调整。
本实现方式中，对于至少一个相关任务模型中的每一个，可以首先输入样本心电信号至该相关任务模型，将该相关任务模型的输出以及对应的相关异常标签输入至预设的第三损失函数，获得第三损失值，并以最小化所述第三损失值为目标，对该相关任务模型进行多轮迭代训练，训练得到一个相关异常识别模型。按照上述过程调整训练得到各相关异常识别模型之后，可以然后将至少一个相关异常识别模型中共有特征提取层的参数共享为目标任务模型中共有特征提取层的参数；之后可以将样本心电信号输入参数共享后的目标任务模型，将目标任务模型的输出以及目标异常标签输入至预设的第四损失函数，获得第四损失值；之后可以以最小化第四损失值为目标，对目标任务模型进行训练，将训练完成后的目标任务模型确定为目标异常识别模型。
图2示意性地示出了一种信号识别方法的流程图，如图2所示，该方法 可以包括以下步骤。
步骤S21：获取目标心电信号。
本实施例中,该步骤具体可以包括以下步骤：首先获取原始心电信号；然后对原始心电信号进行预处理，获得目标心电信号。
本实施例的执行主体可以为计算机设备，该计算机设备具有信号识别装置，通过该信号识别装置来执行本实施例提供的信号识别方法。其中，计算机设备例如可以为智能手机、平板电脑、个人计算机等，本实施例对此不作限定。
本实施例的执行主体可以通过多种方式来获取原始心电信号。例如，执行主体可以获取由如心电图仪等信号采集设备所采集的原始心电信号，然后对获取的原始心电信号进行预处理，获得目标心电信号。
通过预处理，可以使目标心电信号的格式与训练目标异常识别模型时输入的样本心电信号的格式相同。在一种可选的实现方式中，对原始心电信号进行预处理的步骤，可以包括以下步骤至少之一：采用带通滤波器去除原始心电信号中的工频干扰；采用低通滤波器去除原始心电信号中的肌电干扰；以及，采用高通滤波器去除原始心电信号中的基线漂移。
具体地，可以用带通滤波器去除50Hz工频干扰；低通滤波器去除10-300Hz的肌电干扰；用高通滤波器去除基线漂移。通过对原始心电信号进行预处理，可以去除原始心电信号中的噪声干扰，提高分类识别的准确度。
步骤S22：输入目标心电信号至目标异常识别模型中，获得目标异常识别结果，目标异常识别结果用于指示目标心电信号是否具有目标异常；其中，目标异常识别模型是采用任一实施例的模型训练方法训练得到的。
在具体实现中，可以将目标心电信号输入目标异常识别模型中，输出目标异常识别结果。根据输出的目标异常识别结果可以确定目标心电信号是否具有目标异常。目标异常识别结果例如可以包括：目标心电信号具有目标异常的概率以及不具有目标异常的概率，本实施例对此不作限定。
其中，目标异常识别模型可以是预先训练好的，也可以是在信号识别的过程中训练得到的，本实施例对此不作限定。
本实施例提供的信号识别方法，由于目标异常识别模型是与相关异常识别模型基于多任务学习机制训练得到的模型，因此，训练数据少的目标异常 识别模型融合了训练数据多的相关异常识别模型学习到的知识(即参数)，从而可以提升目标异常识别模型的泛化能力和分类性能，提高目标异常识别的准确性。
图7示意性地示出了一种模型训练装置框图。参照图7，可以包括：
样本获取模块71，被配置为获得训练样本集合，所述训练样本集合包括样本心电信号以及所述样本心电信号的异常标签，所述异常标签包括目标异常标签和至少一个相关异常标签；
模型训练模块72，被配置为将所述样本心电信号输入多任务模型，根据所述多任务模型的输出以及所述异常标签，基于多任务学习机制对所述多任务模型进行训练；其中，所述多任务模型包括目标任务模型和至少一个相关任务模型，所述目标任务模型的目标输出为输入的样本心电信号的目标异常标签，所述相关任务模型的目标输出为输入的样本心电信号的相关异常标签；
模型确定模块73，被配置为将训练后的所述目标任务模型确定为目标异常识别模型，所述目标异常识别模型用于识别输入所述目标异常识别模型的心电信号中的目标异常。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关模型训练方法的实施例中进行了详细描述，例如，使用软件、硬件、固件等方式实现，此处将不做详细阐述说明。
图8示意性地示出了一种信号识别装置框图。参照图8，可以包括：
信号获取模块81，被配置为获取目标心电信号；
异常识别模块82，被配置为输入所述目标心电信号至目标异常识别模型中，获得目标异常识别结果，所述目标异常识别结果用于指示所述目标心电信号是否具有目标异常；其中，所述目标异常识别模型是采用任一实施例所述的模型训练方法训练得到的。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有关信号识别方法的实施例中进行了详细描述，例如，使用软件、硬件、固件等方式实现，此处将不做详细阐述说明。
以上所描述的装置实施例仅仅是示意性的，其中所述作为分离部件说明 的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部模块来实现本实施例方案的目的。本领域普通技术人员在不付出创造性的劳动的情况下，即可以理解并实施。
本公开的各个部件实施例可以以硬件实现，或者以在一个或者多个处理器上运行的软件模块实现，或者以它们的组合实现。本领域的技术人员应当理解，可以在实践中使用微处理器或者数字信号处理器(DSP)来实现根据本公开实施例的计算处理设备中的一些或者全部部件的一些或者全部功能。本公开还可以实现为用于执行这里所描述的方法的一部分或者全部的设备或者装置程序(例如，计算机程序和计算机程序产品)。这样的实现本公开的程序可以存储在计算机可读介质上，或者可以具有一个或者多个信号的形式。这样的信号可以从因特网网站上下载得到，或者在载体信号上提供，或者以任何其他形式提供。
例如，图9示出了可以实现根据本公开的方法的计算处理设备。该计算处理设备传统上包括处理器1010和以存储器1020形式的计算机程序产品或者计算机可读介质。存储器1020可以是诸如闪存、EEPROM(电可擦除可编程只读存储器)、EPROM、硬盘或者ROM之类的电子存储器。存储器1020具有用于执行上述方法中的任何方法步骤的程序代码1031的存储空间1030。例如，用于程序代码的存储空间1030可以包括分别用于实现上面的方法中的各种步骤的各个程序代码1031。这些程序代码可以从一个或者多个计算机程序产品中读出或者写入到这一个或者多个计算机程序产品中。这些计算机程序产品包括诸如硬盘，紧致盘(CD)、存储卡或者软盘之类的程序代码载体。这样的计算机程序产品通常为如参考图10所述的便携式或者固定存储单元。该存储单元可以具有与图9的计算处理设备中的存储器1020类似布置的存储段、存储空间等。程序代码可以例如以适当形式进行压缩。通常，存储单元包括计算机可读代码1031’，即可以由例如诸如1010之类的处理器读取的代码，这些代码当由计算处理设备运行时，导致该计算处理设备执行上面所描述的方法中的各个步骤。
本说明书中的各个实施例均采用递进的方式描述，每个实施例重点说明的都是与其他实施例的不同之处，各个实施例之间相同相似的部分互相参见 即可。
最后，还需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、商品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、商品或者设备所固有的要素。在没有更多限制的情况下，由语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、商品或者设备中还存在另外的相同要素。
以上对本公开所提供的一种模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质进行了详细介绍，本文中应用了具体个例对本公开的原理及实施方式进行了阐述，以上实施例的说明只是用于帮助理解本公开的方法及其核心思想；同时，对于本领域的一般技术人员，依据本公开的思想，在具体实施方式及应用范围上均会有改变之处，综上所述，本说明书内容不应理解为对本公开的限制。
应该理解的是，虽然附图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，其可以以其他的顺序执行。而且，附图的流程图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，其执行顺序也不必然是依次进行，而是可以与其他步骤或者其他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。
本领域技术人员在考虑说明书及实践这里公开的发明后，将容易想到本公开的其它实施方案。本公开旨在涵盖本公开的任何变型、用途或者适应性变化，这些变型、用途或者适应性变化遵循本公开的一般性原理并包括本公开未公开的本技术领域中的公知常识或惯用技术手段。说明书和实施例仅被视为示例性的，本公开的真正范围和精神由下面的权利要求指出。
应当理解的是，本公开并不局限于上面已经描述并在附图中示出的精确结构，并且可以在不脱离其范围进行各种修改和改变。本公开的范围仅由所 附的权利要求来限制。
本文中所称的“一个实施例”、“实施例”或者“一个或者多个实施例”意味着，结合实施例描述的特定特征、结构或者特性包括在本公开的至少一个实施例中。此外，请注意，这里“在一个实施例中”的词语例子不一定全指同一个实施例。
在此处所提供的说明书中，说明了大量具体细节。然而，能够理解，本公开的实施例可以在没有这些具体细节的情况下被实践。在一些实例中，并未详细示出公知的方法、结构和技术，以便不模糊对本说明书的理解。
在权利要求中，不应将位于括号之间的任何参考符号构造成对权利要求的限制。单词“包含”不排除存在未列在权利要求中的元件或步骤。位于元件之前的单词“一”或“一个”不排除存在多个这样的元件。本公开可以借助于包括有若干不同元件的硬件以及借助于适当编程的计算机来实现。在列举了若干装置的单元权利要求中，这些装置中的若干个可以是通过同一个硬件项来具体体现。单词第一、第二、以及第三等的使用不表示任何顺序。可将这些单词解释为名称。
最后应说明的是：以上实施例仅用以说明本公开的技术方案，而非对其限制；尽管参照前述实施例对本公开进行了详细的说明，本领域的普通技术人员应当理解：其依然可以对前述各实施例所记载的技术方案进行修改，或者对其中部分技术特征进行等同替换；而这些修改或者替换，并不使相应技术方案的本质脱离本公开各实施例技术方案的精神和范围。

Claims (17)

  1. 一种模型训练方法，其中，包括：
    获得训练样本集合，所述训练样本集合包括样本心电信号以及所述样本心电信号的异常标签，所述异常标签包括目标异常标签和至少一个相关异常标签；
    将所述样本心电信号输入多任务模型，根据所述多任务模型的输出以及所述异常标签，基于多任务学习机制对所述多任务模型进行训练；其中，所述多任务模型包括目标任务模型和至少一个相关任务模型，所述目标任务模型的目标输出为输入的样本心电信号的目标异常标签，所述相关任务模型的目标输出为输入的样本心电信号的相关异常标签；
    将训练后的所述目标任务模型确定为目标异常识别模型，所述目标异常识别模型用于识别输入所述目标异常识别模型的心电信号中的目标异常。
  2. 根据权利要求1所述的模型训练方法，其中，所述基于多任务学习机制对所述多任务模型进行训练的步骤，包括：调整各所述相关任务模型的参数，并根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整。
  3. 根据权利要求2所述的模型训练方法，其中，所述根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤，包括：
    根据所述目标任务模型的参数以及所述至少一个相关任务模型的参数，确定正则损失项，所述正则损失项用于使所述目标任务模型的参数与所述至少一个相关任务模型的参数相似化；
    根据所述正则损失项确定第一损失值，并以最小化所述第一损失值为目标，调整所述目标任务模型的参数。
  4. 根据权利要求3所述的模型训练方法，其中，所述根据所述目标任务模型的参数以及所述至少一个相关任务模型的参数，确定正则损失项的步骤，包括：
    按照以下公式确定所述正则损失项：
    R(θ 1，θ 2，...，θ M)＝λ(|θ 1－θ 2| 2+...+|θ 1－θ M| 2)
    其中，所述R(θ 1，θ 2，...，θ M)表示所述正则损失项，所述M表示所 述多任务模型中所述目标任务模型以及所述相关任务模型的总数量，所述θ 1表示所述目标任务模型的参数，所述θ 2，...，θ M分别表示各所述相关任务模型的参数，所述λ表示预设参量。
  5. 根据权利要求3所述的模型训练方法，其中，所述调整各所述相关任务模型的参数的步骤，包括：
    将所述样本心电信号输入第一相关任务模型，将所述第一相关任务模型的输出以及第一相关异常标签输入至预设的第二损失函数，获得第二损失值，以最小化所述第二损失值为目标，调整所述第一相关任务模型的参数，其中，所述第一相关任务模型为所述至少一个相关任务模型中的任意一个，所述第一相关异常标签为所述至少一个相关异常标签中的任意一个；
    在所述根据所述正则损失项确定第一损失值的步骤之前，还包括：
    将所述样本心电信号输入所述目标任务模型，将所述目标任务模型的输出以及所述目标异常标签输入至预设的经验损失函数，获得经验损失项；
    所述根据所述正则损失项确定第一损失值的步骤，包括：
    计算所述经验损失项以及所述正则损失项之和，获得所述第一损失值。
  6. 根据权利要求5所述的模型训练方法，其中，所述第二损失函数与所述经验损失函数均为交叉熵损失函数。
  7. 根据权利要求2所述的模型训练方法，其中，所述目标任务模型与所述相关任务模型共享有共有特征提取层，所述共有特征提取层用于提取所述目标异常与所述相关异常的共有特征，所述根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤，包括：
    将所述至少一个相关任务模型中所述共有特征提取层的参数共享为所述目标任务模型中所述共有特征提取层的参数，对参数共享后的所述目标任务模型的参数进行调整。
  8. 根据权利要求7所述的模型训练方法，其中，所述调整各所述相关任务模型的参数的步骤，包括：
    将所述样本心电信号输入第二相关任务模型，将所述第二相关任务模型的输出以及第二相关异常标签输入至预设的第三损失函数，获得第三损失值，并以最小化所述第三损失值为目标，调整所述第二相关任务模型的参数，所述第二相关任务模型的参数包括所述共有特征提取层的参数，其中，所述第 二相关任务模型为所述至少一个相关任务模型中的任意一个，所述第二相关异常标签为所述至少一个相关异常标签中的任意一个；
    所述对参数共享后的所述目标任务模型的参数进行调整的步骤，包括：
    将所述样本心电信号输入参数共享后的所述目标任务模型，将所述目标任务模型的输出以及所述目标异常标签输入至预设的第四损失函数，获得第四损失值，并以最小化所述第四损失值为目标，调整所述目标任务模型的参数。
  9. 根据权利要求8所述的模型训练方法，其中，所述第三损失函数与所述第四损失函数均为交叉熵损失函数。
  10. 根据权利要求2至9任一项所述的模型训练方法，其中，所述基于多任务学习机制对所述多任务模型进行训练的步骤，包括：
    基于多任务学习机制，对所述多任务模型进行多轮迭代训练；其中，各轮迭代训练包括：所述调整各所述相关任务模型的参数，并根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤。
  11. 根据权利要求2至9任一项所述的模型训练方法，其中，所述调整各所述相关任务模型的参数的步骤，包括：
    对各所述相关任务模型的参数分别进行多轮迭代调整，直到各所述相关任务模型满足对应的训练停止条件，将训练后的各所述相关任务模型确定为不同的相关异常识别模型；
    所述根据所述至少一个相关任务模型的参数，对所述目标任务模型的参数进行调整的步骤，包括：
    根据至少一个所述相关异常识别模型的参数，对所述目标任务模型的参数进行调整。
  12. 一种信号识别方法，其中，包括：
    获取目标心电信号；
    输入所述目标心电信号至目标异常识别模型中，获得目标异常识别结果，所述目标异常识别结果用于指示所述目标心电信号是否具有目标异常；其中，所述目标异常识别模型是采用如权利要求1至11任一项所述的模型训练方法训练得到的。
  13. 一种模型训练装置，其中，包括：
    样本获取模块，被配置为获得训练样本集合，所述训练样本集合包括样本心电信号以及所述样本心电信号的异常标签，所述异常标签包括目标异常标签和至少一个相关异常标签；
    模型训练模块，被配置为将所述样本心电信号输入多任务模型，根据所述多任务模型的输出以及所述异常标签，基于多任务学习机制对所述多任务模型进行训练；其中，所述多任务模型包括目标任务模型和至少一个相关任务模型，所述目标任务模型的目标输出为输入的样本心电信号的目标异常标签，所述相关任务模型的目标输出为输入的样本心电信号的相关异常标签；
    模型确定模块，被配置为将训练后的所述目标任务模型确定为目标异常识别模型，所述目标异常识别模型用于识别输入所述目标异常识别模型的心电信号中的目标异常。
  14. 一种信号识别装置，其中，包括：
    信号获取模块，被配置为获取目标心电信号；
    异常识别模块，被配置为输入所述目标心电信号至目标异常识别模型中，获得目标异常识别结果，所述目标异常识别结果用于指示所述目标心电信号是否具有目标异常；其中，所述目标异常识别模型是采用如权利要求1至11任一项所述的模型训练方法训练得到的。
  15. 一种计算处理设备，其中，包括：
    存储器，其中存储有计算机可读代码；
    一个或多个处理器，当所述计算机可读代码被所述一个或多个处理器执行时，所述计算处理设备执行如权利要求1至12中任一项所述的方法。
  16. 一种计算机程序，包括计算机可读代码，当所述计算机可读代码在计算处理设备上运行时，导致所述计算处理设备执行根据如权利要求1至12中任一项所述的方法。
  17. 一种计算机可读介质，其中存储了如权利要求1至12中任一项所述的方法。
PCT/CN2021/108605 2021-07-27 2021-07-27 模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质 Ceased WO2023004572A1 (zh)

Priority Applications (3)

Application Number Priority Date Filing Date Title
PCT/CN2021/108605 WO2023004572A1 (zh) 2021-07-27 2021-07-27 模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质
US17/772,405 US20240188895A1 (en) 2021-07-27 2021-07-27 Model training method, signal recognition method, apparatus, computing and processing device, computer program, and computer-readable medium
CN202180002006.0A CN115885279A (zh) 2021-07-27 2021-07-27 模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/108605 WO2023004572A1 (zh) 2021-07-27 2021-07-27 模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质

Publications (1)

Publication Number Publication Date
WO2023004572A1 true WO2023004572A1 (zh) 2023-02-02

Family

ID=85086124

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/108605 Ceased WO2023004572A1 (zh) 2021-07-27 2021-07-27 模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质

Country Status (3)

Country Link
US (1) US20240188895A1 (zh)
CN (1) CN115885279A (zh)
WO (1) WO2023004572A1 (zh)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115953822A (zh) * 2023-03-06 2023-04-11 之江实验室 一种基于rPPG生理信号的人脸视频鉴伪方法和装置
CN116226778A (zh) * 2023-05-09 2023-06-06 水利部珠江水利委员会珠江水利综合技术中心 基于三维分析平台的挡土墙结构异常分析方法及系统
CN116385825A (zh) * 2023-03-22 2023-07-04 小米汽车科技有限公司 模型联合训练方法、装置及车辆

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20240403833A1 (en) * 2023-06-01 2024-12-05 Highwater LLC Repair order creation and sycrnoizatoin systems and methods
CN116919414B (zh) * 2023-07-06 2024-02-13 齐鲁工业大学(山东省科学院) 基于多尺度卷积和密集连接网络的心电信号质量评估方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190130247A1 (en) * 2017-10-31 2019-05-02 General Electric Company Multi-task feature selection neural networks
CN111110224A (zh) * 2020-01-17 2020-05-08 武汉中旗生物医疗电子有限公司 一种基于多角度特征提取的心电图分类方法及装置
CN111134662A (zh) * 2020-02-17 2020-05-12 武汉大学 一种基于迁移学习和置信度选择的心电异常信号识别方法及装置
CN111401558A (zh) * 2020-06-05 2020-07-10 腾讯科技（深圳）有限公司 数据处理模型训练方法、数据处理方法、装置、电子设备
CN112800222A (zh) * 2021-01-26 2021-05-14 天津科技大学 利用共现信息的多任务辅助极限多标签短文本分类方法

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7174205B2 (en) * 2004-04-05 2007-02-06 Hewlett-Packard Development Company, L.P. Cardiac diagnostic system and method
US9421008B2 (en) * 2011-09-23 2016-08-23 Arthrex, Inc. Soft suture-based anchors
US9468386B2 (en) * 2014-03-11 2016-10-18 Ecole polytechnique fédérale de Lausanne (EPFL) Method for detecting abnormalities in an electrocardiogram
US10426364B2 (en) * 2015-10-27 2019-10-01 Cardiologs Technologies Sas Automatic method to delineate or categorize an electrocardiogram
US20210169417A1 (en) * 2016-01-06 2021-06-10 David Burton Mobile wearable monitoring systems
CA3031067A1 (en) * 2016-07-18 2018-01-25 Nantomics, Llc Distributed machine learning systems, apparatus, and methods
US11250314B2 (en) * 2017-10-27 2022-02-15 Cognizant Technology Solutions U.S. Corporation Beyond shared hierarchies: deep multitask learning through soft layer ordering
WO2019222401A2 (en) * 2018-05-17 2019-11-21 Magic Leap, Inc. Gradient adversarial training of neural networks

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190130247A1 (en) * 2017-10-31 2019-05-02 General Electric Company Multi-task feature selection neural networks
CN111110224A (zh) * 2020-01-17 2020-05-08 武汉中旗生物医疗电子有限公司 一种基于多角度特征提取的心电图分类方法及装置
CN111134662A (zh) * 2020-02-17 2020-05-12 武汉大学 一种基于迁移学习和置信度选择的心电异常信号识别方法及装置
CN111401558A (zh) * 2020-06-05 2020-07-10 腾讯科技（深圳）有限公司 数据处理模型训练方法、数据处理方法、装置、电子设备
CN112800222A (zh) * 2021-01-26 2021-05-14 天津科技大学 利用共现信息的多任务辅助极限多标签短文本分类方法

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115953822A (zh) * 2023-03-06 2023-04-11 之江实验室 一种基于rPPG生理信号的人脸视频鉴伪方法和装置
CN116385825A (zh) * 2023-03-22 2023-07-04 小米汽车科技有限公司 模型联合训练方法、装置及车辆
CN116385825B (zh) * 2023-03-22 2024-04-30 小米汽车科技有限公司 模型联合训练方法、装置及车辆
CN116226778A (zh) * 2023-05-09 2023-06-06 水利部珠江水利委员会珠江水利综合技术中心 基于三维分析平台的挡土墙结构异常分析方法及系统
CN116226778B (zh) * 2023-05-09 2023-07-07 水利部珠江水利委员会珠江水利综合技术中心 基于三维分析平台的挡土墙结构异常分析方法及系统

Also Published As

Publication number Publication date
CN115885279A (zh) 2023-03-31
US20240188895A1 (en) 2024-06-13

Similar Documents

Publication Publication Date Title
WO2023004572A1 (zh) 模型训练方法、信号识别方法、装置、计算处理设备、计算机程序及计算机可读介质
Celin et al. ECG signal classification using various machine learning techniques
Dhar et al. Cross-wavelet assisted convolution neural network (AlexNet) approach for phonocardiogram signals classification
Clifford et al. Recent advances in heart sound analysis
CN104970789B (zh) 心电图分类方法及系统
Sharma et al. Automated pre-screening of arrhythmia using hybrid combination of Fourier–Bessel expansion and LSTM
Kao et al. Automatic phonocardiograph signal analysis for detecting heart valve disorders
Al-Shammary et al. Efficient ECG classification based on Chi-square distance for arrhythmia detection
WO2021017313A1 (zh) 房颤的检测方法、装置、计算机设备及存储介质
CN111657925A (zh) 基于机器学习的心电信号分类方法、系统、终端以及存储介质
Zhang et al. [Retracted] An ECG Heartbeat Classification Method Based on Deep Convolutional Neural Network
CN110638430A (zh) 多任务级联神经网络ecg信号心律失常疾病分类模型和方法
Prakash et al. A system for automatic cardiac arrhythmia recognition using electrocardiogram signal
Lan et al. Arrhythmias classification using short-time Fourier transform and GAN based data augmentation
CN109077720B (zh) 信号处理方法、装置、设备和存储介质
Nejedly et al. Classification of ECG using ensemble of residual CNNs with or without attention mechanism
CN116503673A (zh) 一种基于心电图的心律失常识别检测方法及系统
CN115470832A (zh) 一种基于区块链的心电信号数据处理方法
Chen et al. Transfer learning for electrocardiogram classification under small dataset
Wu et al. Autonomous detection of myocarditis based on the fusion of improved quantum genetic algorithm and adaptive differential evolution optimization back propagation neural network
Moqurrab et al. HRIDM: Hybrid Residual/Inception-Based Deeper Model for Arrhythmia Detection from Large Sets of 12-Lead ECG Recordings
An et al. Research on a Lightweight Arrhythmia Classification Model Based on Knowledge Distillation for Wearable Single-Lead ECG Monitoring Systems
Serhani et al. Enhancing arrhythmia prediction through an adaptive deep reinforcement learning framework for ECG signal analysis
AIT BOURKHA et al. Optimized wavelet scattering network and cnn for ecg heartbeat classification from mit–bih arrhythmia database
CN114451898A (zh) 短时训练卷积神经网络的心电信号分类方法

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 17772405

Country of ref document: US

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

Ref document number: 21951189

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205 DATED 27.05.2024)

122 Ep: pct application non-entry in european phase

Ref document number: 21951189

Country of ref document: EP

Kind code of ref document: A1