Electrocardiogram arrhythmia diagnosis training method, system and storage medium

Abstract

The application provides an electrocardiogram arrhythmia diagnosis training system and a training method, comprising the following steps: S1, during hospitalization or perioperative period, original EEP data is collected by in-vivo collection equipment in an electrocardiogram collection layer, original ECG data is synchronously collected by in-vivo collection equipment, and the two groups of data are transmitted to a data analysis layer; S2, in the data analysis layer, the original EEP data is used to automatically label the corresponding original ECG data, and semi-structured ECG data is formed; S3, the semi-structured ECG data is cut to form structured ECG data, and is sent to a training database; S4, a to-be-trained model is selected, a training key evaluation index is set, an optimal training parameter is saved, and a trained algorithm is output. The system uses the convenience of hospitalization or perioperative period, uses the mode that in-vivo EEP data is automatically labeled into corresponding ECG data to establish a training database, trains an algorithm with massive ECG data with accurate labeling, and thus an algorithm with high accuracy can be obtained.

Description

Training methods, systems, and storage media for electrocardiogram arrhythmia diagnosis Technical Field

[0001] This invention relates to the field of medical diagnostic technology, and in particular to a method, system, and storage medium for training in the diagnosis of arrhythmias using electrocardiograms. Background Technology

[0002] In modern arrhythmia diagnosis and monitoring, the primary method involves users wearing surface electrocardiogram (ECG) devices for extended periods to obtain the ECG, especially wearable ECG monitoring devices such as wristwatches. Algorithms then automatically analyze the ECG to obtain diagnostic information for heart diseases such as arrhythmias. Traditional arrhythmia diagnosis algorithms extract feature parameters and set thresholds based on empirical knowledge to determine the type of arrhythmia. However, with sufficient data, algorithms trained through supervised machine learning (ML) or artificial intelligence (AI) demonstrate superior judgment patterns compared to traditional algorithms.

[0003] The current clinical diagnostic process involves first collecting ECG data using surface electrocardiogram (ECG) monitoring equipment, then analyzing the ECG data with software to label and indicate arrhythmias, and finally requiring review and correction by a professional physician to "accurately" label the data in order to form the final diagnostic report. Therefore, the data currently input into ML or AI models for algorithm training is ECG data after physician review and labeling (i.e., raw ECG data and physician-labeled classifications, such as premature ventricular contractions, premature atrial contractions, atrial fibrillation, atrial flutter, and other arrhythmia labels). Since training algorithms for ML or AI models requires a large amount of structured data with accurate annotations, the data input for algorithm training has the following disadvantages: (1) The human body has 80,000 to 100,000 heartbeats a day, and can reach 1 million heartbeats in 10 consecutive days. It is time-consuming and laborious for doctors to review and correct ECG annotations, and the data obtained is almost unsustainable; (2) Although ECG is an important standard for clinical diagnosis of arrhythmia, it is still an "indirect" measurement because ECG is performed outside the human body by modern sensors to measure the electrical activity inside the heart, such as the atria and ventricles. Since the ventricular electrical activity (QRS wave on ECG) is stronger than the atrial electrical activity, the atrial electrical activity (P wave on ECG) is very weak on ECG. It is very difficult for both the human eye and the algorithm to identify the P wave. Therefore, there is a certain error in judging various arrhythmias (such as atrial fibrillation, atrial flutter, premature atrial contractions, junctional escape beats, non-conduction, and conduction block) caused by the atrial fibrillation or related to the surface ECG. Research has found that the consistency of different doctors' interpretations of the same electrocardiogram is only about 72%. In the clinical diagnosis of atrial fibrillation, doctors mainly judge whether the interval between pairs of heartbeats (QRS waves) on the electrocardiogram is irregular, whether there is no atrial contraction P wave before each heartbeat QRS wave, or whether there are multiple P waves (important clues to distinguish atrial fibrillation from atrial flutter). Therefore, the "standard" data obtained by doctors after diagnosis through surface electrocardiogram has a low degree of quantification, resulting in a large error in the algorithm obtained by training ML or AI models. Summary of the Invention

[0004] This applicant addresses the shortcomings of existing arrhythmia diagnosis training methods, which use raw ECG data measured on the body surface and manually corrected by doctors, resulting in errors in arrhythmia labeling and classification. This leads to significant algorithm errors after training, and the time-consuming and labor-intensive process of doctors manually reviewing and correcting massive amounts of labeled data. The applicant provides an ECG arrhythmia diagnosis training method, system, and storage medium. This training system utilizes the convenience of hospitalization or the perioperative period to simultaneously collect EEP data (such as cardiac electrical activity collected from the intracardiac chamber or epicardiac wall) and ECG data from the patient using in vivo and surface ECG acquisition devices. A training database is established by automatically labeling arrhythmias accurately into the corresponding ECG data using EEP data or a combination of EEP and ECG data. The massive amount of accurately labeled ECG data is then fed into ML or AI models to train different algorithms, outputting highly accurate algorithms for application in daily diagnosis. The diagnostic results obtained can reach or even surpass the level of experienced ECG physicians.

[0005] The number of leads for a surface electrocardiogram (ECG) is determined by the number of different potential points that need to be collected from the body surface, such as a single lead, 12 leads, or even 18 leads. The number of leads for an intracardiac electrical activity acquisition device can also be determined by the number of different potential points that can be collected from within the heart chambers or on the outer wall of the heart. This method is applicable to data acquired by all of the above-mentioned acquisition devices.

[0006] The technical solution adopted in this invention and its beneficial effects are as follows:

[0007] A training method for diagnosing arrhythmias using electrocardiograms includes the following steps:

[0008] S1. During the patient's hospitalization or perioperative period, raw EEP data is collected in the ECG acquisition layer using an in-body acquisition device, and raw ECG data is collected simultaneously using an external acquisition device. The two sets of data are then transmitted to the data analysis layer.

[0009] S2. In the data analysis layer, the original EEP data is used to automatically annotate the corresponding original ECG data to form semi-structured ECG data.

[0010] S3. Cut the semi-structured ECG data into structured ECG data and send it into the training database;

[0011] S4. Select the model to be trained, set the key evaluation indicators for training, save the optimal post-training parameters, and output the trained algorithm.

[0012] As a further improvement to the above technical solution:

[0013] S2 includes the following steps:

[0014] S201. Automatically identify each QRS wave in the raw ECG data and determine the time interval between two adjacent heartbeats R1 and R2, using this as the unit analysis interval for EEP data.

[0015] S202. Automatically identify pulse signals in the unit analysis interval of the raw EEP data, detect pulse positions, calculate the number of pulses and their time distances from R1 and R2 respectively;

[0016] S203. Determine the heartbeat characteristics of ECG based on the calculation results of the original EEP data;

[0017] S204. Multiple unit analysis intervals with the same arrhythmia characteristics are aggregated into an event, and the start and end times of the event are determined. The automatic annotation module annotates the heartbeat and event into the raw ECG data, thereby obtaining semi-structured ECG data with standard annotations.

[0018] S3 includes the following steps:

[0019] S301, The data segmentation module segments the semi-structured ECG data and its corresponding annotations into several units of fixed duration.

[0020] S302. The training database is divided into a learning group and a testing group. The learning group is further divided into a training group and a validation group. The structured ECG data is randomly allocated into the above three groups according to specific rules.

[0021] S4 includes the following steps:

[0022] The structured ECG data in the S401 learning group undergoes iterative learning;

[0023] The data in group S402 and the testing group validate the data trained in group S401.

[0024] S403. For a new or previously trained ML or AI mathematical model, three sets of data can be used to obtain standard parameters as key evaluation indicators to evaluate the trained model.

[0025] S404: Automatically monitor whether learning has terminated using the F1 score, save the parameters that first reach the highest F1 score, and output the trained algorithm.

[0026] Taking advantage of hospitalization or the perioperative period, EEP (intracardiac endocardial endocardium) and ECG (external cardiac endocardial ...

[0027] S2 also includes the following steps:

[0028] S205. Doctors or professionals can review and correct the automatic annotations of semi-structured ECG data using the review and correction software.

[0029] Doctors can check or verify the automatic annotations by viewing the EEP and synchronized ECG data on the review and correction software 22. If there are any special cases of automatic annotation errors, the annotations can be manually modified to obtain corrected annotations, which are then fed into the model for training along with the corresponding semi-structured ECG data to achieve sustainable optimization of the algorithm training.

[0030] It also includes S5, which includes the following steps:

[0031] S501. Collect ECG data from the body surface using the terminal acquisition module for routine diagnosis and transmit it to the terminal application module. For different types of arrhythmia events, the terminal analysis module of the terminal application module uses a trained algorithm to label the collected raw ECG data to obtain semi-structured ECG data. The terminal analysis module sends the semi-structured ECG data and its corresponding labels that determine that the human-computer interaction changes of the user terminal software exceed the threshold to the terminal database, where they are cut into structured ECG data.

[0032] S502. After the number of newly added examples of structured ECG data reaches the preset value in the terminal database, it is sent to the optimization module of the model training layer and the current algorithm is retrained.

[0033] In routine arrhythmia diagnosis, users use terminal devices such as wristwatch ECG devices and electrocardiographs to collect ECGs of the body surface. The terminal device uses a trained algorithm to obtain labeled ECG data, thereby obtaining a diagnostic result. Users can also confirm or modify the labels by replaying the ECG. If the modification exceeds a threshold, the ECG data along with the modified labels can be sent to the optimization module of the model training layer to perform a new round of optimization and upgrade of the original trained algorithm, so as to achieve continuous supplementary optimization of the algorithm.

[0034] An electrocardiogram (ECG) arrhythmia diagnosis training system that performs the above-described training method, the diagnostic training system comprising an ECG acquisition layer, a data analysis layer, and a model training layer, wherein:

[0035] The electrocardiogram (ECG) acquisition layer is a hardware device used to simultaneously acquire, store, and transmit intracardiac EEP data and surface ECG data during hospitalization or perioperative period.

[0036] The data analysis layer is used to analyze the raw EEP data acquired by the ECG acquisition layer and automatically annotate the heartbeat characteristics and events of the raw EEP data into the raw ECG data;

[0037] The model training layer is used to build the training database and train the algorithm.

[0038] As a further improvement to the above technical solution:

[0039] The electrocardiogram (ECG) acquisition layer is an integrated device with an in vivo acquisition module, a body surface acquisition module, and a data storage module. It includes at least one lead interface for EEP signal acquisition and one interface for acquiring body surface ECG. The other end of the EEP signal acquisition lead enters the outer layer of the human heart or enters the heart cavity through a vein. The device has embedded software to control different acquisition and storage circuits.

[0040] The data analysis layer includes an automatic annotation module and review and correction software. The automatic annotation module is used for cloud analysis and automatic annotation, while the review and correction software has the function of displaying and correcting EEP and ECG data and annotations. The model training layer includes a training database and a model training module. The data segmentation module of the training database has the function of segmenting semi-structured ECG data and its corresponding annotations into structured ECG data. The model training module receives the structured ECG data in the training database and outputs the trained algorithm.

[0041] The arrhythmia diagnosis system based on surface ECG can execute the above training method. It can simultaneously collect in vivo EEP data and surface ECG data through the ECG acquisition layer. Then, the raw EEP data collected by ECG acquisition layer 1 is analyzed through data analysis layer, and the heartbeat characteristics and events of the raw EEP data are automatically labeled into the raw ECG data. Then, the training database and training algorithm are established through the model training layer. Finally, a highly accurate algorithm is obtained and applied to daily diagnosis. The diagnostic results can reach or even exceed the level of senior ECG physicians.

[0042] The model training layer also includes an optimization module, which receives newly added structured ECG data from the terminal database of the terminal device layer and retrains the trained algorithm.

[0043] If the trained algorithm produces incorrect arrhythmia classifications in daily surface heart rhythm monitoring, it can be manually modified and sent back to the algorithm as more accurate supplementary data for a new round of supplementation, optimization, and upgrade.

[0044] A computer storage medium storing a computer program, wherein the storage medium is hardware storage or cloud storage, and the computer program, when executed, implements the above-mentioned electrocardiogram arrhythmia diagnosis training method.

[0045] A computer storage medium storing a computer program, wherein the storage medium is hardware storage or cloud storage, the computer program uses the above-mentioned arrhythmia diagnosis training method to obtain a trained algorithm, and performs arrhythmia diagnosis based on the surface electrocardiogram collected by the daily terminal acquisition module.

[0046] The computer program in the computer storage medium contains algorithms. This storage medium can be embedded in independent hardware devices, such as electrocardiographs, wearable electrocardiogram devices, and wristwatch electrocardiogram devices, or it can be stored in the cloud for electrocardiogram data analysis uploaded to the cloud. Attached Figure Description

[0047] Figure 1 is a schematic diagram of the system architecture of the present invention;

[0048] Figure 2 is a flowchart of the arrhythmia diagnosis training method based on surface electrocardiogram of the present invention.

[0049] In the diagram: 1. Electrocardiogram (ECG) acquisition layer; 11. In vivo acquisition module; 12. Body surface acquisition module; 13. Data storage module; 2. Data analysis layer; 21. Automatic annotation module; 22. Review and correction software; 3. Model training layer; 31. Training database; 311. Data segmentation module; 32. Model training module; 33. Optimization module; 4. Terminal device layer; 41. Terminal acquisition module; 42. Terminal application module; 421. Terminal analysis module; 422. User terminal software; 43. Terminal database. Detailed Implementation

[0050] The specific embodiments of the present invention will now be described with reference to the accompanying drawings.

[0051] As shown in Figure 1, the present invention provides an arrhythmia diagnosis training system based on body surface electrocardiogram, which includes an electrocardiogram acquisition layer 1, a data analysis layer 2, and a model training layer 3. The electrocardiogram acquisition layer 1 is used to synchronously collect, store, and transmit the intracardiac electrical signals (i.e., the EEP defined previously) and body surface electrocardiogram signals (ECG) of patients during hospitalization or the perioperative period. The data analysis layer 2 is used to analyze the original EEP data collected in the electrocardiogram acquisition layer 1, and automatically annotate the heartbeat nature and events in the original EEP data to the original ECG data to obtain semi-structured ECG data. The model training layer 3 is used to preprocess the semi-structured ECG data to make it into structured ECG data available for algorithm training, and establish a training database, training algorithm, and optimization and upgrade algorithm.

[0052] The electrocardiogram acquisition layer 1 is mainly hardware equipment, which includes an intracorporeal acquisition module 11, a body surface acquisition module 12, and a data storage module 13. The above three modules can be integrated into the same device or be independent devices respectively. Some patients in cardiothoracic surgery or cardiology department need to receive clinical monitoring of intracorporeal / intracardiac or epicardial electrical activities during hospitalization or the perioperative period. The electrocardiogram acquisition layer 1 cleverly realizes the synchronous collection, storage, and transmission of a large amount of intracardiac electrical signals (EEP) and body surface electrocardiogram signals (ECG). In this embodiment, the above three modules are integrated into the same device, which at least includes a wire interface for pluggable EEP signal acquisition and an interface for collecting body surface ECG. The other end of the EEP signal acquisition wire enters the outer surface of the human heart or enters the intracardiac cavity through the venous blood vessel. This device has embedded software, which can control different acquisition, storage, and other circuits. Of course, independent intracorporeal devices and body surface devices can also be used for synchronous acquisition, or the data of different acquired devices can be synchronized.

[0053] The data analysis layer 2 includes an automatic annotation module 21 and a review and correction software 22. The automatic annotation module 21 can be cloud analysis, which analyzes the entire original EEP data (or can analyze mainly based on EEP and supplemented by ECG), and automatically annotates the annotations in the original EEP data (i.e., standard annotations) to the corresponding original ECG data, and outputs semi-structured ECG data with standard heartbeat and event annotations to be trained to the model training layer 3. The review and correction software 22 is a networked human-computer interaction software supporting the electrocardiogram acquisition layer 1, which can display the original EEP data, original ECG data, and semi-structured ECG data to be trained collected synchronously above, and is available for operators (such as clinical doctors, experts, etc.) to view and correct the annotations of the semi-structured ECG data.

[0054] The model training layer 3 includes a training database 31, a model training module 32, and an optimization module 33. The training database 31 contains unstructured data (raw EEP and ECG data) and semi-structured data (EEP and ECG data with heartbeat and event annotations). The training database 31 has a data segmentation module 311, which segments the semi-structured ECG data and its corresponding annotations into several structured ECG data of fixed duration (e.g., 30 seconds). The number of target event annotations on the structured ECG data is fixed (the target event annotation is 0 or 1; for example, if the target event is atrial fibrillation, then 0 indicates that this ECG segment is not atrial fibrillation, and 1 indicates that this ECG segment is atrial fibrillation, and the number of target event annotations is 1). This satisfies the data requirements for training in the model training module 32. Finally, the trained algorithm output by the model training module 32 can be uploaded to the cloud platform for terminal hardware devices to analyze the collected ECG data, or it can be embedded in independent terminal hardware devices in the form of software. The optimization module 33 is used to perform a new round of supplementary training and upgrade on the algorithm using the newly added structured ECG data in the training database 31.

[0055] In routine arrhythmia monitoring, users can obtain arrhythmia diagnoses at terminal device layer 4, which uses the trained algorithm. Terminal device layer 4 is for actual clinical or health monitoring applications and includes a terminal acquisition module 41, a terminal application module 42, and a terminal database 43. The terminal acquisition module 41 is used to acquire, store, and transmit ECG data from the body surface to the terminal application module 42. The terminal acquisition module 41 can be a hardware device such as an electrocardiograph, wearable ECG device, or wristwatch ECG device. The terminal application module 42 has a terminal analysis module 421 and user software 422. The user software 422 allows users to replay, view, edit, and modify all ECG annotation information and complete diagnostic conclusions and reports. The terminal analysis module 421 analyzes the ECG data acquired by the terminal acquisition module 41 according to the trained algorithm and automatically annotates it. It can also determine whether the user's human-computer interaction modifications in the user software 422 exceed a threshold, thereby deciding whether the data should be sent to the terminal database 43 as supplementary training data for the optimization module 33 of the model training layer 3.

[0056] As shown in Figure 2, the method for diagnostic training using the above system includes the following steps:

[0057] S1. During the patient's hospitalization or perioperative period, raw EEP data is collected in the ECG acquisition layer 1 using the in vivo acquisition module 11, and raw ECG data is collected synchronously using the body surface acquisition module 12. The two sets of data are then transmitted to the data analysis layer 2.

[0058] S2. In data analysis layer 2, the original EEP data is used to automatically annotate the corresponding original ECG data to form semi-structured ECG data.

[0059] S201. Automatically identify each QRS wave (i.e., ventricular systolic activity) in the raw ECG data, determine the time interval T between two adjacent heartbeats (such as QRS2-QRS1, or R2-R1), and use this as the unit analysis interval for EEP data;

[0060] S202. Automatically identify pulse signals (i.e., atrial contraction activity) within the unit analysis interval of the raw EEP data, detect pulse position A, calculate the number of pulses A and their time distances from R1 and R2 respectively. If multiple A's exist, the frequency of atrial contraction activity within the unit analysis interval (i.e., the mean of the intervals between multiple A's) and whether the contraction activity is stable (i.e., the standard deviation of the time intervals between multiple A's exceed a certain threshold) can be calculated. If only one A's exists, the time interval between A and R1 and R2 can be used as a measure of whether the conduction time from atrial contraction to ventricular contraction is normal to determine the nature of the heartbeat, such as normal or premature beat.

[0061] S203. Based on the calculation results of the original EEP data, and in accordance with the clinical medical definition of arrhythmia, the nature of the heartbeat (sinus normal, atrial premature beat) and whether the heartbeats of each pair are atrial fibrillation, atrial flutter, atrial tachycardia, or other subcategories of arrhythmia can be accurately determined, thereby determining the nature of the heartbeat on the ECG.

[0062] S204. Aggregate the arrhythmia type of a segment of heartbeats based on the arrhythmia type between pairs of heartbeats. If the same arrhythmia type appears in multiple unit analysis segments, they are aggregated into an "event". The start and end times of the event are determined. The automatic labeling module 21 then labels the standard heartbeats and events into the original ECG data, thereby obtaining semi-structured ECG data with standard labels.

[0063] S205. Professionals such as doctors can review and correct the annotations of the semi-structured ECG data in S204 using the review and correction software 22.

[0064] S3. Cut the semi-structured ECG data to form structured ECG data and establish a training database 31.

[0065] S301, the data segmentation module 311 segments the continuous semi-structured ECG data from multiple days and its corresponding annotations into several 30-second units (with or without overlap) to form structured ECG data that can be fed into the model for training.

[0066] S302, the training database 31 is divided into a learning group and a testing group. The learning group is further subdivided into a training group and a validation group. The structured ECG data is randomly allocated into the above three groups according to the division criteria such as the number of data sets or the number of patients. For example, the ratio of training group: testing group: validation group = 50%: 30%: 20% or 70%: 20%: 10%. The ratio can be flexibly adjusted according to the training effect of the algorithm.

[0067] S4. Select the model to be trained, set the training database and iteration target (key evaluation indicators), save the optimal post-training parameters, and output the trained algorithm.

[0068] The structured ECG data in the S401 learning group undergoes iterative learning;

[0069] The data in the S402 and testing groups are used to validate the training data in S401. The system can automatically assign data in the testing group that does not match the standard answer. Operators review or edit the data in the review and correction software 22. If the algorithm is insufficient, the data will be added to the next automatic training learning data group. If the labeling is insufficient, it will be corrected and the data will also be added to the next automatic training learning data group.

[0070] S403. For a new or previously trained ML or AI mathematical model, the above three groups can respectively obtain five standard parameters of algorithm performance as key evaluation indicators for assessing the trained model, such as sensitivity (Se), specificity (Sp), positive prediction rate (PP), negative prediction rate (NP), and F1 score (F1 score, out of 100%). Further, indicators such as accuracy, precision, recall, and F1 score can be derived. All of the above indicators are standard indicators used to evaluate the level of artificial intelligence training capabilities.

[0071] S404. Use the F1 score to automatically monitor whether the algorithm learning has terminated. If it is found that the F1 score has stopped improving, select the first time the highest F1 score is reached as the training termination point, save the optimal post-training parameters, and output the trained algorithm.

[0072] S5. In daily arrhythmia monitoring, ECG data exceeding the preset modification threshold will be used as supplementary data to automatically optimize and upgrade the algorithm.

[0073] S501, the terminal acquisition module 41 is used to acquire, store, and transmit ECG data from the body surface to the terminal application module 42. For different types of arrhythmia events, the terminal analysis module 421 of the terminal application module 42 uses the post-training algorithm in S404 to annotate the acquired raw ECG data, thereby obtaining semi-structured ECG data. Users can replay, view, edit, and modify the ECG annotation information on the user terminal software 422. The terminal analysis module 421 then determines whether the human-computer interaction modification of the user terminal software 422 exceeds the threshold. If it does, the example of semi-structured ECG data and its corresponding annotations are sent to the terminal database 43, where they are segmented into structured ECG data.

[0074] S502 and the terminal database 43 automatically monitor whether the number of newly added structured ECG data examples has reached a preset value. If it has, the structured ECG data is sent to the optimization module 33 of the model training layer 3, and a new round of automatic training is performed on the current algorithm. If the F1 score of the new round of training results improves, the upgraded algorithm is output. If it does not improve, the current algorithm is not upgraded, and the failed data in the testing group in this round is included in the learning group in the next round of training.

[0075] The above description is an explanation of the present invention and not a limitation thereof. The present invention can be modified in any form without departing from its spirit.

Claims

1. A training method for diagnosing cardiac arrhythmias using electrocardiograms, characterized in that, Includes the following steps: S1. During the patient's hospitalization or perioperative period, raw EEP data is collected in the ECG acquisition layer (1) using an in vivo acquisition device, and raw ECG data is collected simultaneously using an external acquisition device. The two sets of data are then transmitted to the data analysis layer (2). S2. In the data analysis layer (2), the raw EEP data is used to automatically annotate the corresponding raw ECG data to form semi-structured ECG data. S201. Each QRS wave in the raw ECG data is automatically identified, and the time interval between two adjacent heartbeats R1 and R2 is determined, which is used as the unit analysis interval of the EEP data. S202. The pulse signal in the unit analysis interval of the raw EEP data is automatically identified, and the pulse position is detected, the number of pulses is calculated, and their time distances from R1 and R2 are calculated respectively. S203. The ECG data is determined based on the calculation results of the raw EEP data. The heartbeat properties of G; S204, aggregate multiple unit analysis intervals with the same arrhythmia properties into an event, and determine the start and end times of the event. The automatic labeling module (21) labels the heartbeat and event into the original ECG data, thereby obtaining semi-structured ECG data with standard labels; S3, cut the semi-structured ECG data to form structured ECG data and send it into the training database (31); S301, the data cutting module (311) cuts the semi-structured ECG data and its corresponding labels into several units of fixed duration; S302, the training database (31) is divided into a learning group and a testing group. The learning group is further divided into a training group and a validation group. The structured ECG data is randomly allocated into the above three groups according to the rules and proportions. S4. Select the model to be trained, set the key evaluation indicators for training, save the optimal post-training parameters, and output the trained algorithm. The structured ECG data in the S401 learning group undergoes iterative learning; S402. The data in the testing group validates the data trained in S401; S403. For a new or previously trained ML or AI mathematical model, the standard parameters of the three sets of data are used as key evaluation indicators to evaluate the trained model; S404. The F1 score is used to automatically monitor whether the learning has terminated, the parameters that first reach the highest F1 score are saved, and the trained algorithm is output.

2. The electrocardiogram arrhythmia diagnosis training method according to claim 1, characterized in that, S2 also includes the following steps: S205, doctors or professionals can review and correct the automatic annotation of semi-structured ECG data in the review and correction software (22).

3. The electrocardiogram arrhythmia diagnosis training method according to claim 1, characterized in that, It also includes S5, which includes the following steps: S501, using the terminal acquisition module (41) for daily diagnosis to collect ECG data from the body surface and transmit it to the terminal application module (42). For different types of arrhythmia events, the terminal analysis module (421) of the terminal application module (42) uses the trained algorithm to label the collected raw ECG data, thereby obtaining semi-structured ECG data. The terminal analysis module (421) sends the semi-structured ECG data and its corresponding labels that determine that the human-computer interaction modification of the user terminal software (422) exceeds the threshold to the terminal database (43), and cuts them into structured ECG data in the terminal database (43); S502, after the number of new cases of structured ECG data reaches the preset value of the terminal database (43), it is sent to the optimization module (33) of the model training layer (3) and the current algorithm is retrained.

4. A training system for diagnosing arrhythmias using electrocardiograms and implementing the training method of claim 1, characterized in that, The diagnostic training system includes an electrocardiogram (ECG) acquisition layer (1), a data analysis layer (2), and a model training layer (3). The ECG acquisition layer (1) is a hardware device used to synchronously acquire, store, and transmit intracardiac EEP data and surface ECG data during hospitalization or perioperative period. The data analysis layer (2) is used to analyze the raw EEP data acquired by the ECG acquisition layer (1) and automatically label the heartbeat characteristics and events of the raw EEP data into the raw ECG data. The model training layer (3) is used to establish a training database and training algorithm.

5. The electrocardiogram arrhythmia diagnosis training system according to claim 4, characterized in that: The electrocardiogram acquisition layer (1) is an integrated device with an in vivo acquisition module (11), a body surface acquisition module (12) and a data storage module (13). It includes at least one lead interface for EEP signal acquisition and one interface for acquiring body surface ECG. The other end of the EEP signal acquisition lead enters the outer surface of the human heart or enters the heart cavity through a vein. The device has embedded software to control different acquisition, storage and other circuits.

6. The electrocardiogram arrhythmia diagnosis training system according to claim 4, characterized in that: The data analysis layer (2) includes an automatic annotation module (21) and a review and correction software (22). The automatic annotation module (21) is used for cloud analysis and automatic annotation. The review and correction software (22) has the function of displaying and correcting EEP, ECG data and annotations. The model training layer (3) includes a training database (31) and a model training module (32). The data cutting module (311) of the training database (31) has the function of cutting semi-structured ECG data and its corresponding annotations into structured ECG data. The model training module (32) receives the structured ECG data in the training database (31) and outputs the trained algorithm.

7. The electrocardiogram arrhythmia diagnosis training system according to claim 4, characterized in that: The model training layer (3) also includes an optimization module (33), which receives newly added structured ECG data from the terminal database (43) of the terminal device layer (4) and retrains the trained algorithm.

8. A computer storage medium storing a computer program thereon, characterized in that: The storage medium is hardware storage or cloud storage, and the computer program, when executed, implements the electrocardiogram arrhythmia diagnosis training method of claim 1.

9. A computer storage medium storing a computer program thereon, characterized in that: The storage medium is hardware storage or cloud storage. The computer program uses the arrhythmia diagnosis training method of claim 1 to obtain the trained algorithm and performs arrhythmia diagnosis based on the electrocardiogram of the body surface collected by the daily terminal acquisition module (41).

Images