Electrocardiogram data processing method and system based on large model training and storage medium

Abstract

The invention discloses an electrocardiogram data processing method and system based on large model training and a storage medium, and the method comprises the following steps: S1, periodically obtaining heart beat data of a plurality of leads; s2, the electrocardiogram large model constructs a real-time relevancy matrix and a real-time feature vector according to the heart beat data; s3, the classifier performs classification according to the real-time feature vector and the real-time relevancy matrix, and judges whether the classification result of the real-time feature vector is consistent with the classification result of the real-time relevancy matrix; and S4, if yes, prompting an abnormal type corresponding to the real-time feature vector, and outputting an electrocardiogram corresponding to the heart beat data, and if not, outputting the similarity between the real-time feature vector and each abnormal type and the electrocardiogram corresponding to the heart beat data, manually judging the abnormal type, and correcting the electrocardiogram large model according to a manual judgment result. And comprehensive analysis on electrocardiogram data of multiple leads is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of electrocardiogram data processing, and in particular to an electrocardiogram data processing method, system and storage medium based on large model training. Background Art

[0002] The electrocardiogram (ECG) is a commonly used preliminary screening tool for cardiovascular disease. When measuring an ECG, multiple electrodes are typically placed at various locations on the human body. The voltage changes between the electrodes are used to measure the heart's electrical activity. The connections formed by the electrodes are called leads. A conventional ECG includes twelve leads recording the heartbeat waveform: six are limb leads and the other six are chest leads. By observing and comparing the waveforms of these twelve leads, a preliminary diagnosis of cardiovascular disease can be made. In practical ECG applications, patients sometimes need to be monitored continuously for extended periods of time to promptly detect abnormal changes in their heartbeats. However, this generates a large amount of ECG data, and relying solely on manual screening by medical staff can lead to missed or incorrect detections due to visual fatigue or inattention.

[0003] Currently, intelligent analysis of electrocardiogram (ECG) data has garnered widespread attention. Prior art CN112528783B discloses an ECG heartbeat data clustering method that compares heartbeat data with templates in a template library to determine the type of ECG abnormality. In this prior art, only the heartbeat data recorded from a single lead is compared. However, in actual diagnosis of heartbeat data, a single cardiovascular disease often results in abnormal heartbeat data from multiple leads. Therefore, using only the heartbeat data from a single lead as a basis for diagnosis can lead to discrepancies between the results and the actual situation. Summary of the Invention

[0004] In order to solve the technical problem that lead misconnection or misplacement affects electrocardiogram data analysis, the present application provides an electrocardiogram data processing method, system, and storage medium based on large model training, wherein the electrocardiogram data processing method based on large model training includes the following steps: S1. Periodically acquiring heartbeat data of multiple leads, where the heartbeat data includes multiple complete heartbeat cycles; S2. The electrocardiogram large model constructs a real-time correlation matrix and a real-time eigenvector based on the heartbeat data, where each element of the real-time correlation matrix represents the correlation between the heartbeat data recorded by any two leads, and each component of the real-time eigenvector represents the real-time eigenvalue of the corresponding lead; S3, a classifier classifies the real-time feature vector and the real-time correlation matrix respectively, and determines whether the classification result of the real-time feature vector is consistent with the classification result of the real-time correlation matrix; S4. If yes, the abnormality type corresponding to the real-time feature vector is prompted, and the electrocardiogram corresponding to the heartbeat data is output; if no, the similarity between the real-time feature vector and each abnormality type and the electrocardiogram corresponding to the heartbeat data are output, and the abnormality type is manually determined by an expert, and the electrocardiogram large model is corrected according to the expert manual determination result.

[0005] The electrocardiogram large model generates real-time feature vectors and real-time correlation matrices based on the heartbeat data, and performs classification based on the real-time feature vectors and real-time correlation matrix, thereby taking into account the differences in the correlation of heartbeat data of different leads in different abnormal heartbeats, thereby improving the accuracy of the classification results.

[0006] Specifically, in step S3, the classifier classifies the real-time feature vector and the real-time correlation matrix respectively as follows: S31, the classifier forms a prediction feature vector and a prediction correlation matrix for each abnormality type; S32, selecting one of the predicted feature vectors one by one, and calculating the vector similarity between the real-time feature vector and the predicted feature vector; S33, determining whether the vector similarity is greater than a first threshold; if so, comparing the real-time correlation matrix with the predicted correlation matrix; if not, determining whether there are any predicted feature vectors that have not participated in the calculation; S34: If there is a predicted feature vector that is not involved in the calculation, then go to step S32; if not, the classification result of the real-time feature vector is inconsistent with the classification result of the real-time correlation matrix.

[0007] By comparing the similarity between real-time feature vectors and predicted feature vectors, classification can be performed, preliminarily determining the anomaly type. The vector similarity is also provided, facilitating subsequent expert judgment based on the analysis results. Furthermore, when the real-time feature vectors are similar to the predicted feature vectors, the real-time correlation matrix is compared with the predicted correlation matrix, effectively reducing computational complexity and improving efficiency.

[0008] Specifically, the vector similarity is the similarity between the real-time feature vector in step S4 and each abnormality type.

[0009] Specifically, in step S33, the steps of comparing the real-time correlation matrix with the predicted correlation matrix are as follows: S331, calculating the matrix similarity between the real-time correlation matrix and the predicted correlation matrix, and determining whether the matrix similarity is greater than a second threshold; S332. If yes, the real-time correlation matrix matches the predicted correlation matrix, and the classification result of the real-time feature vector is consistent with the classification result of the real-time correlation matrix. If no, the real-time correlation matrix does not match the predicted correlation matrix, and the classification result of the real-time feature vector is inconsistent with the classification result of the real-time correlation matrix.

[0010] Specifically, in step S3, the classifier is obtained through the following steps: S34, establishing the electrocardiogram large model for different user groups; S35. Obtain user information and select a classifier based on the user information.

[0011] Specifically, the electrocardiogram large model is obtained by the following steps: S21. Obtain historical electrocardiogram data of different user groups; S22. Classify all or part of the historical electrocardiogram data using manual expert classification to create multiple classification labels, each classification label corresponding to an abnormality type; S23, establishing a historical electrocardiogram data set based on the historical electrocardiogram data; S24. Using the historical electrocardiogram dataset to train the classifier, and obtain the electrocardiogram large model.

[0012] Furthermore, in step S4, after the abnormality type corresponding to the real-time feature vector is prompted, the following steps are further performed: S41. The manual expert determines whether the classification result is accurate. If so, the electrocardiogram model is trained based on the heartbeat data and the manual determination result. If not, the abnormality type is re-assigned and the electrocardiogram model is corrected based on the manual determination result.

[0013] By introducing artificial expert judgment, the output results of the electrocardiogram large model are judged so as to timely discover the errors of the electrocardiogram large model and make corrections to the electrocardiogram large model in time.

[0014] Specifically, in step S41, the step of correcting the electrocardiogram large model includes: S411, the electrocardiogram large model updates the electrocardiogram dataset according to the manual determination result and the heartbeat data, and generates a test feature vector and a test similarity matrix for each abnormality type; S412, calculating the test vector similarity between the real-time feature vector and the test feature vector, and the test matrix similarity between the real-time correlation matrix and the test correlation matrix; S413. Determine whether the test vector similarity and the test matrix similarity are both less than a third threshold. If so, complete the iteration; if not, adjust the Q value function and jump to the step of generating a test feature vector and a test correlation matrix.

[0015] When the output result of the ECG model does not match the actual situation, the ECG model is trained through reinforcement learning to improve the prediction accuracy of the ECG model for the abnormality type.

[0016] The present application also provides an electrocardiogram data processing system based on large model training, comprising: a memory, a processor, and an electrocardiogram data processing program based on large model training stored in the memory and runnable on the processor. When the electrocardiogram data processing program based on large model training is executed by the processor, the steps of the above-mentioned electrocardiogram data processing method based on large model training are implemented.

[0017] The present application also provides a computer-readable storage medium storing an electrocardiogram data processing program based on large model training. When the program is executed by a processor, the steps of the above-mentioned electrocardiogram data processing method based on large model training are implemented.

[0018] The technical effects and advantages of the present invention are as follows: 1. The ECG model generates real-time feature vectors and real-time correlation matrices based on heartbeat data, and performs classification based on these real-time feature vectors and real-time correlation matrices. This allows for the differences in the correlation of heartbeat data from different leads across different abnormality types, thereby improving the accuracy of the classification results. 2. By comparing the similarity between the real-time feature vector and the predicted feature vector, the classification can be performed, which can preliminarily determine the anomaly type and provide vector similarity, facilitating subsequent judgment by human experts based on the analysis results. Furthermore, when the real-time feature vector is similar to the predicted feature vector, the real-time correlation matrix is compared with the predicted correlation matrix, which can effectively reduce the amount of calculation and improve computational efficiency. 3. By introducing manual expert judgment, the output results of the ECG model are judged so that errors in the ECG model can be discovered in a timely manner and the ECG model can be corrected in a timely manner. The ECG model is reinforced through reinforcement learning to improve the prediction accuracy of the ECG model for this abnormality type. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is an overall flow chart of the electrocardiogram data processing method of the present invention.

[0020] Figure 2 This is a flow chart of the present invention for classifying real-time feature vectors and real-time correlation matrices in sequence.

[0021] Figure 3 This is a flow chart of the present invention for simultaneously classifying real-time feature vectors and real-time correlation matrices.

[0022] Figure 4 This is a flow chart of the present invention for correcting the electrocardiogram model based on the results of manual expert judgment. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] Based on existing diagnostic experience, different cardiovascular diseases can cause abnormalities in ECG data from different leads. Therefore, different ECG abnormality types often result in abnormalities in different leads. Therefore, when analyzing ECG data, it is important to reference ECG data from all leads and consider the correlation between the ECG data from each lead to comprehensively determine the ECG abnormality.

[0025] Based on the above analysis, this embodiment provides an electrocardiogram data processing method based on large model training. Figure 1 , including the following steps: S1. Periodically acquiring heartbeat data of multiple leads, where the heartbeat data includes multiple complete heartbeat cycles; S2. The ECG model constructs a real-time correlation matrix and a real-time eigenvector based on the heartbeat data. Each element of the real-time correlation matrix represents the correlation between the heartbeat data recorded by any two leads, and each component of the real-time eigenvector represents the real-time eigenvalue of the corresponding lead. S3, the classifier classifies the real-time feature vector and the real-time correlation matrix respectively, and determines whether the classification result of the real-time feature vector is consistent with the classification result of the real-time correlation matrix; S4. If yes, the abnormality type corresponding to the real-time feature vector is prompted, and the electrocardiogram corresponding to the heartbeat data is output. If no, the similarity between the real-time feature vector and each abnormality type and the electrocardiogram corresponding to the heartbeat data are output, and the abnormality type is manually determined by experts, and the electrocardiogram large model is corrected according to the results of the expert manual determination.

[0026] Because different cardiovascular diseases can cause abnormal heartbeat data in different leads, the correlation between the eigenvalues of each lead corresponding to different abnormality types varies. Therefore, when analyzing heartbeat data, the heartbeat data recorded by all leads should be analyzed, and the correlation between the heartbeat data recorded by each lead should also be analyzed. The large ECG model generates real-time feature vectors and a real-time correlation matrix based on the heartbeat data, and classifies them based on the real-time feature vectors and the real-time correlation matrix. This can take into account the differences in the correlation between the heartbeat data of different leads for different abnormality types, thereby improving the accuracy of the classification results.

[0027] It is particularly important to note that ECGs vary significantly between people of different ages, genders, and physical conditions (for example, athletes and ordinary people in good health have distinct ECG differences, including differences in heart rate and QRS waveform). Therefore, when building a large ECG model, different ECG models should be created based on the age, gender, and physical condition of the user group. This requires that the historical ECG data used in large-scale model training contain this classification information.

[0028] Therefore, the ECG model can be obtained by the following steps: S21. Obtain historical electrocardiogram data of different user groups; S22. Classify all or part of the historical electrocardiogram data using manual expert classification to create multiple abnormality type classification labels, each classification label corresponding to an abnormality type; S23, establishing a historical electrocardiogram data set based on the historical electrocardiogram data; S24. Use the historical electrocardiogram dataset to train the classifier and obtain a large electrocardiogram model.

[0029] Using historical ECG data with existing classification labels to train the ECG large model can effectively improve the training speed and accuracy of the ECG large model. When manual experts classify the historical ECG data, they can classify all the historical ECG data or classify part of the historical ECG data, and then the ECG large model automatically completes the classification of the remaining historical ECG data. The more historical ECG data with existing classification labels in the historical ECG data participating in the training, the higher the accuracy of the trained ECG large model. Therefore, preferably, when the scale of the historical ECG data can be realized by manual experts, it is best to classify all the historical ECG data to improve the accuracy of the ECG large model.

[0030] In step S3, the following steps may also be included: S34. Establish a large ECG model for different user groups; S35. Obtain user information and select a classifier based on the user information; It's also best if the historical ECG data includes multiple types of ECG anomalies, clearly categorized and labeled. For example, ECG anomalies caused by various cardiovascular diseases can be included. Ideally, the historical ECG data should also include abnormal ECG data caused by misconnected leads, so that abnormal ECG data caused by misconnected or misplaced leads can be screened out.

[0031] The real-time characteristic value of the lead can be a certain characteristic data of each heartbeat, such as heart rate, QRS wave peak, QRS wave duration, etc., or it can be multiple characteristic values of multiple heartbeat stages, for example, the real-time characteristic vector includes the QRS wave peak, QRS wave duration, QRS wave gap, etc. recorded by each lead.

[0032] The advantage of using eigenvectors is that the data format of vectors is easy to organize. Each component of the vector can correspond to an eigenvalue in the data set, a vector can correspond to a classification label in the data set, and a classification label corresponds to an abnormal type of electrocardiogram, which is convenient for subsequent comparison operations.

[0033] The following table shows a specific embodiment of the correlation matrix, in which the correlation between leads is calculated based on the amplitude of the QRS wave in the electrocardiogram waveform of each lead.

[0034] In specific applications, the correlation of other features can also be statistically analyzed to form a correlation matrix, such as the frequency of the QRS wave, the interval duration of the ST wave, the amplitude of the T wave, etc. Each lead can also have multiple data participating in the construction of the correlation matrix.

[0035] Furthermore, the fast Fourier transform and wavelet transform can be combined to obtain the first-order fundamental frequency in the spectrum analysis results, and then the correlation of the fundamental frequencies in the heartbeat data of each lead can be analyzed to construct a correlation matrix, or the amplitude corresponding to the fundamental frequency can be used to construct a correlation matrix. Of course, the fundamental frequency and amplitude can also be used to construct a correlation matrix at the same time.

[0036] It is worth noting that the more parameters involved in constructing the correlation matrix, the larger the matrix size, the larger the scale of the calculation, and the higher the performance requirements for the hardware device. Therefore, in actual application, the appropriate number of parameters can be selected based on the conditions of the hardware device.

[0037] Specifically, refer to Figure 2 In step S3, the classifier classifies the real-time feature vector and the real-time correlation matrix respectively as follows: S31, the classifier forms a prediction feature vector and a prediction correlation matrix for each abnormality type; S32, selecting a predicted feature vector one by one, and calculating the vector similarity between the real-time feature vector and the predicted feature vector; S33: Determine whether the vector similarity is greater than a first threshold. If so, compare the real-time correlation matrix with the predicted correlation matrix. If not, determine whether there are any predicted feature vectors that have not yet been calculated. S34: If there is a predicted feature vector that is not involved in the calculation, then go to step S32; if not, the classification result of the real-time feature vector is inconsistent with the classification result of the real-time correlation matrix.

[0038] By comparing the similarity between real-time feature vectors and predicted feature vectors, classification can be performed, preliminarily determining the anomaly type. The vector similarity is also provided, facilitating subsequent expert judgment based on the analysis results. Furthermore, when the real-time feature vectors are similar to the predicted feature vectors, the real-time correlation matrix is compared with the predicted correlation matrix, effectively reducing computational complexity and improving efficiency.

[0039] The steps of classifying the real-time feature vector and the real-time correlation matrix can also be as follows: Figure 3 As shown, the real-time feature vector and real-time correlation matrix are classified at the same time: S35, the classifier forms a prediction feature vector and a prediction correlation matrix for each abnormality type; S36, selecting one prediction feature vector at a time, calculating the vector similarity between the real-time feature vector and the prediction feature vector, and the matrix similarity between the real-time correlation matrix and the prediction correlation matrix corresponding to the prediction feature vector involved in the calculation; S37, determining whether both the vector similarity and the matrix similarity are greater than a first threshold; if so, the classification result of the real-time feature vector is consistent with the classification result of the real-time correlation matrix; if not, determining whether there are any predicted feature vectors that have not participated in the calculation; S38. If there is a predicted feature vector that is not involved in the calculation, then go to step S32. If not, the classification result of the real-time feature vector is inconsistent with the classification result of the real-time correlation matrix.

[0040] Of course, in step S37 , the vector similarity may be compared with the first threshold, and the matrix similarity may be compared with the second threshold, so as to change the strictness of the two and thus better adapt to different situations.

[0041] Specifically, the vector similarity is the similarity between the real-time feature vector in step S4 and each abnormality type.

[0042] Specifically, vector similarity Calculated by the following formula: Where, is the real-time feature vector, is the predicted feature vector.

[0043] Vector similarity can also be calculated in other ways, such as real-time feature vector The modulus and predicted feature vector However, this calculation method is not as accurate as the above formula.

[0044] Specifically, in step S33, the steps of comparing the real-time correlation matrix with the predicted correlation matrix are as follows: S331, calculating the matrix similarity between the real-time correlation matrix and the predicted correlation matrix, and determining whether the matrix similarity is greater than a second threshold; S332. If so, the real-time correlation matrix matches the predicted correlation matrix, and the classification result of the real-time feature vector is consistent with the classification result of the real-time correlation matrix. If not, the real-time correlation matrix does not match the predicted correlation matrix, and the classification result of the real-time feature vector is inconsistent with the classification result of the real-time correlation matrix.

[0045] Specifically, the matrix similarity Calculated by the following formula: Where, is the real-time correlation matrix, is the prediction correlation matrix, is the F-norm, is the F-norm.

[0046] There are many matrix norms, and the F-norm is preferred, which can well reflect the matrix The actual size of is used to determine the similarity between the real-time correlation matrix and the predicted correlation matrix.

[0047] Furthermore, considering that the output results of the large ECG model are difficult to be completely accurate, human experts are introduced to judge the output results of the large ECG model, and the large ECG model is corrected according to the judgment results, which can make the output of the large ECG model more accurate.

[0048] like Figure 2 、 Figure 3 and Figure 4 As shown, in step S4, after prompting the abnormality type corresponding to the real-time feature vector, the following steps are also performed: S41. The manual expert determines whether the classification result is accurate. If so, the ECG model is trained based on the heartbeat data and the manual judgment result. If not, the abnormality type is re-assigned and the ECG model is corrected based on the manual judgment result.

[0049] Specifically, in step S41, the steps of correcting the electrocardiogram large model include: S411. The electrocardiogram large model updates the electrocardiogram dataset according to the manual judgment result and the heartbeat data, and generates a test feature vector and a test similarity matrix for each abnormality type; S412, calculating the test vector similarity between the real-time feature vector and the test feature vector, and the test matrix similarity between the real-time correlation matrix and the test correlation matrix; S413: Determine whether the test vector similarity and the test matrix similarity are both less than a third threshold value. If so, complete the iteration; if not, adjust the Q value function and jump to the step of generating a test feature vector and a test correlation matrix.

[0050] When the output result of the ECG model does not match the actual situation, the ECG model is trained through reinforcement learning to improve the prediction accuracy of the ECG model for the abnormality type.

[0051] The present application also provides an electrocardiogram data processing system based on large model training, comprising: a memory, a processor, and an electrocardiogram data processing program based on large model training stored in the memory and runnable on the processor. When the electrocardiogram data processing program based on large model training is executed by the processor, the steps of the above-mentioned electrocardiogram data processing method based on large model training are implemented.

[0052] The present application also provides a computer-readable storage medium storing an electrocardiogram data processing program based on large model training. When the program is executed by a processor, the steps of the above-mentioned electrocardiogram data processing method based on large model training are implemented.

[0053] The specific implementations of the above-mentioned system and readable storage medium are well known to those skilled in the art and will not be described in detail here.

[0054] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for processing electrocardiogram data based on large model training, characterized in that: The following steps are involved: S1. Periodically acquiring heartbeat data of multiple leads, where the heartbeat data includes multiple complete heartbeat cycles; S2. The electrocardiogram model constructs a real-time correlation matrix and a real-time eigenvector based on the heartbeat data, where each element of the real-time correlation matrix represents the correlation between the heartbeat data recorded by any two leads, and each component of the real-time eigenvector represents the real-time eigenvalue of the corresponding lead; S3, a classifier classifies the real-time feature vector and the real-time correlation matrix respectively, and determines whether the classification result of the real-time feature vector is consistent with the classification result of the real-time correlation matrix; S4. If yes, the abnormality type corresponding to the real-time feature vector is prompted, and the electrocardiogram corresponding to the heartbeat data is output; if no, the similarity between the real-time feature vector and each abnormality type and the electrocardiogram corresponding to the heartbeat data are output, and the abnormality type is manually determined by an expert, and the electrocardiogram large model is corrected according to the expert manual determination result.

2. The electrocardiogram data processing method based on large model training according to claim 1 is characterized in that: In step S3, the classifier classifies the real-time feature vector and the real-time correlation matrix respectively as follows: S31, the classifier forms a prediction feature vector and a prediction correlation matrix for each abnormality type; S32, selecting one of the predicted feature vectors one by one, and calculating the vector similarity between the real-time feature vector and the predicted feature vector; S33, determining whether the vector similarity is greater than a first threshold; if so, comparing the real-time correlation matrix with the predicted correlation matrix; if not, determining whether there are any predicted feature vectors that have not participated in the calculation; S34: If there is a predicted feature vector that is not involved in the calculation, then go to step S32; if not, the classification result of the real-time feature vector is inconsistent with the classification result of the real-time correlation matrix.

3. The electrocardiogram data processing method based on large model training according to claim 2 is characterized in that: The vector similarity is the similarity between the real-time feature vector in step S4 and each abnormality type.

4. The electrocardiogram data processing method based on large model training according to claim 2, characterized in that: In step S33, the steps of comparing the real-time correlation matrix with the predicted correlation matrix are as follows: S331, calculating the matrix similarity between the real-time correlation matrix and the predicted correlation matrix, and determining whether the matrix similarity is greater than a second threshold; S332. If yes, the real-time correlation matrix matches the predicted correlation matrix, and the classification result of the real-time feature vector is consistent with the classification result of the real-time correlation matrix. If no, the real-time correlation matrix does not match the predicted correlation matrix, and the classification result of the real-time feature vector is inconsistent with the classification result of the real-time correlation matrix.

5. The electrocardiogram data processing method based on large model training according to claim 1 is characterized in that: In step S3, the classifier is obtained by the following steps: S34, establishing the electrocardiogram large model for different user groups; S35. Obtain user information and select a classifier based on the user information.

6. The electrocardiogram data processing method based on large model training according to claim 5, characterized in that: The electrocardiogram model is obtained by the following steps: S21. Obtain historical electrocardiogram data of different user groups; S22. Classify all or part of the historical electrocardiogram data using manual expert classification to create multiple classification labels, each classification label corresponding to an abnormality type; S23, establishing a historical electrocardiogram data set based on the historical electrocardiogram data; S24. Using the historical electrocardiogram dataset to train the classifier, and obtain the electrocardiogram large model.

7. The electrocardiogram data processing method based on large model training according to claim 1, characterized in that: In step S4, after the abnormality type corresponding to the real-time feature vector is prompted, the following steps are further performed: S41. The manual expert determines whether the classification result is accurate. If so, the electrocardiogram model is trained based on the heartbeat data and the manual determination result. If not, the abnormality type is re-assigned and the electrocardiogram model is corrected based on the manual determination result.

8. The electrocardiogram data processing method based on large model training according to claim 7 is characterized in that: In step S41, the steps of modifying the electrocardiogram large model include: S411, the electrocardiogram large model updates the electrocardiogram dataset according to the manual determination result and the heartbeat data, and generates a test feature vector and a test similarity matrix for each abnormality type; S412, calculating the test vector similarity between the real-time feature vector and the test feature vector, and the test matrix similarity between the real-time correlation matrix and the test correlation matrix; S413. Determine whether the test vector similarity and the test matrix similarity are both less than a third threshold. If so, complete the iteration; if not, adjust the Q value function and jump to the step of generating a test feature vector and a test correlation matrix.

9. An electrocardiogram data processing system based on large model training, characterized in that: The electrocardiogram data processing system based on large model training includes: a memory, a processor, and an electrocardiogram data processing program based on large model training stored in the memory and executable on the processor. When the electrocardiogram data processing program based on large model training is executed by the processor, the steps of the electrocardiogram data processing method based on large model training are implemented as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The readable storage medium stores an electrocardiogram data processing program based on large model training, which, when executed by a processor, implements the steps of the electrocardiogram data processing method based on large model training as described in any one of claims 1-8.

Images