A heartbeat clustering method and system based on SHAP value weighting and segmented HDBSCAN

Abstract

This invention provides a heartbeat clustering method and system based on SHAP value weighting and segmented HDBSCAN. The method uses the SHAP model to calculate the contribution values ​​of different features of all types of heartbeats in the classification task from a public and proprietary database, and uses these values ​​as weights in the clustering algorithm. Then, the invention employs a segmented clustering method, using the HDBSCAN algorithm to sequentially cluster the QRS wave, P wave, and T wave three times to obtain the final clustering result. This algorithm does not require a preset number of clusters and can perform relatively accurate and stable personalized clustering of ECGs from different populations, exhibiting strong anti-interference capabilities.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, specifically to a heartbeat clustering method and system based on SHAP value weighting and segmented HDBSCAN. Background Technology

[0002] Cardiovascular disease has long been considered a major threat to people's health and even lives. Electrocardiogram (ECG), as one of the main clinical diagnostic tools for cardiovascular disease, reflects the electrical currents generated during the periodic changes of cardiac contraction and relaxation. A typical ECG includes the P wave, QRS complex, and T wave. The P wave reflects the potential changes during atrial depolarization; the PR interval represents the time it takes for the impulse to travel from the sinoatrial node through the atrioventricular junction to the start of ventricular depolarization; the QRS complex reflects the potential changes during ventricular depolarization; and the T wave represents the potential changes during ventricular repolarization. If the heart undergoes pathological changes, the ECG signal will be distorted in both waveform period and morphology.

[0003] Therefore, analyzing heartbeats in electrocardiogram (ECG) signals plays a crucial role in intelligent ECG signal analysis and clinical auxiliary diagnosis. Abnormal heartbeat types in everyday ECG data are often unpredictable, and classification models trained with fixed categories often cannot cover all categories. Heartbeat clustering is an unsupervised heartbeat classification algorithm that exhibits better generalization performance compared to supervised heartbeat classification algorithms. Especially with the advent of mobile ECG acquisition devices, clustering algorithms for ECG signals over 24 hours or even longer periods are essential. Accurate heartbeat clustering can also be used to build a heartbeat dictionary, which can then be further analyzed using relevant natural language processing algorithms.

[0004] In existing technologies, ECG clustering algorithms typically group morphologically similar heartbeats together based on heartbeat feature parameters or template matching to obtain characteristic waveforms of the target object's ECG signal for analysis and diagnostic assistance by technicians. However, these clustering methods cannot distinguish the different contributions of different types of heartbeats in the clustering task. Treating them as having equal weight will inevitably affect the accuracy of the clustering results. Furthermore, they cannot perform personalized clustering based on each individual's ECG, are susceptible to noise interference, have poor stability, and the accuracy of the obtained characteristic waveforms is unstable. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a heartbeat clustering method and system based on SHAP value weighting and segmented HDBSCAN. The method uses the SHAP model to calculate the contribution values ​​of different features of all types of heartbeats in the classification task from a database of electrocardiograms, including both public and proprietary databases, and uses these values ​​as weights in the clustering algorithm. Then, the HDBSCAN algorithm is used to perform clustering operations based on the density of the clustering target. This method can perform relatively accurate and stable clustering of electrocardiograms from different populations, without limiting the number of clusters, and has strong anti-interference capabilities.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] On one hand, this invention provides a heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN, comprising the following steps:

[0008] Acquire ECG data from an ECG database and denoise it. The ECG database includes public databases and proprietary databases. Extract heartbeat data from the denoised ECG data.

[0009] All heartbeat data are preprocessed to obtain training data and test data;

[0010] The training data is classified, and the contribution values ​​of different features of all types of heartbeats in the classification task are calculated and used as weight values ​​in the clustering algorithm.

[0011] The test data is fused with the weight values ​​to obtain weighted ECG feature point fusion data;

[0012] The fused data was clustered three times in sequence using a segmented clustering method to generate the clustering results.

[0013] Furthermore, the SHAP model is used to calculate the contribution values ​​of different features of all types of heartbeats in the ECG database in the classification task, i.e., the SHAP values, and these values ​​are used as weight values ​​in the clustering algorithm. The SHAP model calculation method is an additive interpretation model that can calculate the marginal contribution of each feature to the output of the classification algorithm model.

[0014] Furthermore, the segmented clustering method performs three sequential clustering operations on the fused data, including using the HDBSCAN algorithm to perform three sequential clustering operations on the feature points of the QRS band, P band, and T band, with each clustering operation based on the previous clustering operation.

[0015] Furthermore, the denoising includes: removing power frequency interference, removing baseline drift, and removing motion artifacts and electromyographic interference.

[0016] Furthermore, the preprocessing involves segmenting and downsampling all heart rate data.

[0017] Furthermore, the training data is classified using a sample data balancing method and a random forest algorithm.

[0018] Furthermore, the training data uses raw heartbeat data without manual feature extraction.

[0019] Secondly, the present invention provides a heartbeat clustering system based on SHAP value weighting and segmented HDBSCAN, comprising:

[0020] An extraction module is used to acquire and denoise ECG data from an ECG database, which includes public databases and proprietary databases, and to extract heartbeat data from the denoised ECG data.

[0021] The preprocessing module is used to preprocess all heartbeat data to obtain training and test data;

[0022] The weight module is used to classify the training data, calculate the contribution value of different features of all types of heartbeats in the classification task, and use it as the weight value in the clustering algorithm.

[0023] The fusion module is used to fuse test data with weight values ​​to obtain weighted ECG feature point fusion data;

[0024] The clustering module uses a segmented clustering method to perform three sequential clustering operations on the fused data to generate clustering results.

[0025] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN as described above.

[0026] Fourthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN as described above.

[0027] Beneficial effects of this invention:

[0028] This invention uses the SHAP model to calculate the contribution values ​​of different features of all types of heartbeats in the classification task in the public ECG database and the proprietary database, i.e., SHAP values, and uses them as weight values ​​in the clustering algorithm, which can effectively improve the accuracy of the clustering results. The original heartbeat data is used in the SHAP value training process, and no manual feature extraction process is performed. Only downsampling is performed to reduce computational complexity.

[0029] This invention employs a segmented clustering method, performing sequential three-stage clustering of the QRS wave, P wave, and T wave to obtain the final clustering result. This algorithm does not require a preset number of clusters and can perform relatively accurate and stable personalized clustering of ECGs from different populations. It has strong anti-interference capabilities and can cluster heartbeats that are significantly different from other heartbeat trajectories separately. At the same time, this algorithm clusters based on the density of the clustering target rather than distance, which avoids the limitation of algorithms such as K-means that can only find spherical clusters.

[0030] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of the heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN in Embodiment 1 of the present invention.

[0033] Figure 2 The following are examples of heart rate feature point diagrams in Embodiment 1 of the present invention: (a) is a heart rate feature point diagram for atrial fibrillation (feature points 1-96), and (b) is a heart rate feature point diagram for normal heart rate (feature points 1-96).

[0034] Figure 3 This is a graph showing the average contribution of feature points to SHAP for 11 categories in Embodiment 1 of the present invention.

[0035] Figure 4 This is a clustering effect diagram of Embodiment 1 of the present invention.

[0036] Figure 5 This is a schematic diagram illustrating the anti-interference performance of the clustering effect in Embodiment 1 of the present invention. Detailed Implementation

[0037] Example 1

[0038] This embodiment 1 provides a heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN.

[0039] like Figure 1 As shown, a heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN includes the following steps:

[0040] Step 1: Obtain ECG data from the publicly available MIT-BIH arrhythmia database and remove noise. Extract AF heartbeats and 11 other types of heartbeat data from the ECG data.

[0041] Step 2: Preprocess all heartbeat data to obtain training data and test data;

[0042] Step 3: Classify the training data, calculate the contribution value of different features of all types of heartbeats in the classification task, and use them as weight values ​​in the clustering algorithm;

[0043] Step 4: Fuse the test data with the weight values ​​to obtain weighted ECG feature point fusion data;

[0044] Step 5: Perform three sequential clustering operations on the fused data using the segmented clustering method to generate clustering results.

[0045] Specifically, in step 1, since the P wave of atrial fibrillation beats changes into high-frequency f waves, this invention extracts atrial fibrillation beats from the database and separates them from the normal (N) category to form a separate category. Taking the MIT-BIH arrhythmia database as an example, the sampling frequency of the MIT-BIH arrhythmia database is 360Hz. AF beats are not separately classified but are given in the form of rhythm segments. All AF beats in the 48 data points of the MIT-BIH arrhythmia database are separated to form a separate category.

[0046] The denoising includes: removing power frequency interference, removing baseline drift, and removing motion artifacts and electromyographic interference.

[0047] In step 2, the preprocessing involves segmenting the ECG data of all heartbeats from the 48 data points in step 1 into 0.8 seconds of data, consisting of 0.3 seconds before the R-peak and 0.5 seconds after the R-peak. Then, an integer-multiple downsampling algorithm is used to downsample the data to 120Hz. In this embodiment, a 3x downsampling algorithm is used. At this point, the number of sample points in each segment is 0.8*360 / (360 / 120) = 96, forming training data with a feature value of 96. Each feature point is a heartbeat sample point, representing the original time-domain waveform characteristics of the data, such as... Figure 2As shown, (a) is a feature map of atrial fibrillation heartbeats (feature points 1-96), and (b) is a feature map of normal heartbeats (feature points 1-96). The integer downsampling algorithm consists of two steps: first, a digital low-pass filter is used to remove high-frequency components from the signal to prevent aliasing; then, sample points with an interval of 3 after removing high-frequency components are retained.

[0048] When performing data segmentation, different segmentation lengths can be selected for different clustering purposes. For example, if the task focuses on the classification and clustering of atrial fibrillation, a segmentation distance of about 0.8 seconds can be selected, because most atrial fibrillation is a rapid arrhythmia.

[0049] Step 3 involves classifying the training data. Using the MIT-BIH arrhythmia database, eleven categories of heartbeat data are generated, including normal (N), atrial fibrillation (af), premature ventricular contractions (V), premature atrial contractions (A), premature atrial contractions with aberrant conduction (a), junctional premature beats (J), ventricular escape beats (E), atrial escape beats (e), junctional escape beats (j), ventricular fusion beats (F), and pacing ( / ). Other databases can be used to add or remove categories as needed. Table 1 shows the category comparison of different types of heartbeats (MIT-BIH category names) in the classification model:

[0050] Table 1: Classification of different types of heartbeats in the classification model

[0051]

[0052] This invention employs a sample data balancing method and a random forest algorithm to classify the training data. It uses downsampling for multiple categories (such as N and af) and upsampling for fewer categories, resulting in an overall of 11 categories and an F1 score of 0.94. The downsampling algorithm uses random downsampling; for example, if there are many samples in category N, a certain number of samples are randomly removed, and the remaining samples are used as training data. The upsampling algorithm uses a residual bootstrap method to perform multiple resampling operations to increase the number of samples in the fewer-sample category.

[0053] This invention uses the SHAP model to calculate the contribution values ​​of different features of all types of heartbeats in the MIT-BIH arrhythmia database in a classification task, i.e., SHAP values, as the weights of features in the clustering algorithm.

[0054] SHAP is an additive interpretation model inspired by Shapley values. The model is trained using a random forest algorithm. For each predicted sample, the model generates a prediction value, and the SHAP value is the numerical value assigned to each feature in that sample. Let the i-th sample be x. i The j-th feature of the i-th sample is x. i,j The model predicts the value of the i-th sample as y. iThe baseline of the entire model (usually the mean of the target variable for all samples) is y. base Then the SHAP value follows the following equation:

[0055] y i =y base +f(x i,1 )+f(x i,2 )+…+f(x i,k )

[0056] Where f(x) i,j ) is x i,j The SHAP value represents the contribution of the j-th feature of the i-th sample to the final predicted classification value yi. i ranges from 1 to 35786, where 35786 is the number of training samples, and j ranges from 1 to 96, where 96 is the number of features. When f(x) i,j If f(x) > 0, it indicates that the feature improves the predicted value and has a positive effect; when f(x) > 0, it means that the feature improves the predicted value and has a positive effect. i,j If f(x) < 0, it indicates that the feature lowers the predicted value and has a negative effect; when f(x) < 0, it means that the feature lowers the predicted value and has a negative effect. i,j If ) = 0, it means that the feature has no effect on the predicted value.

[0057] Figure 3 This graph displays the mean absolute value of the SHAP contribution of each feature point to the 11 categories. In other words, it shows the average of the predicted values ​​for each feature across all sample points. The horizontal axis represents the contribution, i.e., the weight, and the vertical axis represents the 20 feature points with the largest total weight (e.g., the Xth feature point). This graph only shows the top 20 most important feature points, with the feature name corresponding to its position in the ECG feature map. For example, the first row shows the average absolute value of the predicted contribution of the 37th sample point feature to categories 0, 1, 2, 10, 4, 3, 5, 7, 9, 6, 8 (i.e., N, Af, V, / , a, A, J, e, F, E, j) across all samples.

[0058] Because the R-peak and other ECG bands have significant differences in ECG values, this invention employs a random forest algorithm, independent of feature measurement units, for classification tasks to calculate SHAP values. Similar algorithms such as XGBOOST achieve comparable classification results. During SHAP value training, the raw heartbeat data is used as training data; no manual feature extraction is performed, only downsampling is done to reduce computational complexity.

[0059] In step 5, a segmented clustering method is used to sequentially cluster the QRS wave, P wave, and T wave three times to obtain the final clustering result. The 96 features are clustered three times using the HDBSCAN algorithm in the order of QRS segment (feature values ​​31-65), P segment (first 30 feature values), and T segment (last 31 feature values). The first clustering focuses on the QRS segment, thus separating arrhythmias with QRS abnormalities, such as premature ventricular contractions (PVCs) and pacing. The second clustering focuses on the P wave, separating arrhythmias with P wave abnormalities, such as atrial fibrillation, premature atrial contractions (PACs), and premature ventricular contractions (PVCs). The final clustering focuses on the T wave, separating arrhythmias such as myocardial infarction, premature ventricular contractions (PVCs), and escape beats. Each clustering is based on the previous clustering, and the amplitude abnormality of a certain segment will not affect the overall clustering result. For example, a large R peak amplitude will have the effect of a small P peak and T peak. Figure 4 This diagram illustrates the clustering results. Figures a, b, c, and d show four clusters, each superimposed with multiple heartbeat signals. Each curve represents a sample data point from the cluster, containing 96 sample points. The horizontal axis represents the 96 feature points, and the vertical axis represents the amplitude corresponding to each feature point, in millivolts. Each curve represents a sample data point and is an interpolated graph. Each image represents a cluster category, superimposed with all samples from that category.

[0060] This invention employs the HDBSCAN algorithm for each clustering operation. This algorithm does not require a preset number of clusters and can perform personalized clustering based on each individual's electrocardiogram. Furthermore, it can separately cluster heartbeats whose trajectories are significantly different from other heartbeats, such as... Figure 5 As shown, each cluster category contains two data points, and the waveform of each data point is significantly different from that of the other samples. Figure 4 As shown, it was not classified into other categories, but rather into a separate category, demonstrating the strong anti-interference ability of this clustering algorithm. At the same time, this algorithm clusters based on the density of the clustering target rather than the distance, which can avoid the shortcomings of algorithms such as K-means that can only find spherical clusters.

[0061] This embodiment can perform targeted, accurate, and stable clustering of electrocardiograms from different populations. Unlike most algorithms (such as K-means), which require pre-setting the number of clusters, this embodiment does not require pre-setting the number of clusters during initialization and has strong anti-interference capabilities. Furthermore, the SHAP value training process uses raw data without manual feature extraction. Since the contribution of each feature to the prediction result is known, feature selection is predictable, making the importance of each feature point to the classification task and to each category interpretable.

[0062] Example 2

[0063] This embodiment 2 provides a heartbeat clustering system based on SHAP value weighting and segmented HDBSCAN.

[0064] A heartbeat clustering system based on SHAP value weighting and segmented HDBSCAN includes:

[0065] The extraction module is used to acquire ECG data from the ECG database and remove noise, and extract heartbeat data from the ECG data;

[0066] The preprocessing module is used to preprocess all heartbeat data to obtain training and test data;

[0067] The weight module is used to classify the training data, calculate the contribution value of different features of all types of heartbeats in the classification task, and use it as the weight value in the clustering algorithm.

[0068] The fusion module is used to fuse test data with weight values ​​to obtain weighted ECG feature point fusion data;

[0069] The clustering module uses a segmented clustering method to perform three sequential clustering operations on the fused data to generate clustering results.

[0070] Example 3

[0071] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium for storing computer instructions. When executed by a processor, the computer instructions implement the heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN as described above. The method includes:

[0072] Acquire ECG data from ECG databases and denoise them, including public databases and proprietary databases, and extract heartbeat data from the denoised ECG data;

[0073] All heartbeat data are preprocessed to obtain training data and test data;

[0074] The training data is classified, and the contribution values ​​of different features of all types of heartbeats in the classification task are calculated and used as weight values ​​in the clustering algorithm.

[0075] The test data is fused with the weight values ​​to obtain weighted ECG feature point fusion data;

[0076] The fused data was clustered three times in sequence using a segmented clustering method to generate the clustering results.

[0077] Example 4

[0078] Embodiment 4 of the present invention provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the heartbeat clustering method based on SHAP value weighting and segmented HDBSCAN as described above. The method includes:

[0079] Acquire ECG data from ECG databases and denoise them, including public databases and proprietary databases, and extract heartbeat data from the denoised ECG data;

[0080] All heartbeat data are preprocessed to obtain training data and test data;

[0081] The training data is classified, and the contribution values ​​of different features of all types of heartbeats in the classification task are calculated and used as weight values ​​in the clustering algorithm.

[0082] The test data is fused with the weight values ​​to obtain weighted ECG feature point fusion data;

[0083] The fused data was clustered three times in sequence using a segmented clustering method to generate the clustering results.

[0084] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A heart beat clustering method based on SHAP value weighting and segmented HDBSCAN, characterized in that, Includes the following steps: Acquire ECG data from ECG databases and denoise them, including public databases and proprietary databases, and extract heartbeat data from the denoised ECG data; All heartbeat data are preprocessed to obtain training data and test data; The training data is classified, and the contribution values ​​of different features of all types of heartbeats in the classification task are calculated and used as weight values ​​in the clustering algorithm. The SHAP model is used to calculate the contribution values ​​of different features of all types of heartbeats in the ECG database in the classification task, i.e., the SHAP values, and these values ​​are used as weight values ​​in the clustering algorithm. The SHAP model is an additive interpretive model that can calculate the marginal contribution of each feature to the output of the classification algorithm model. The test data is fused with the weight values ​​to obtain weighted ECG feature point fusion data; The fused data was clustered three times sequentially using a segmented clustering method to generate clustering results. The segmented clustering method performs three sequential clustering operations on the fused data, including using the HDBSCAN algorithm to perform three sequential clustering operations on the feature points of the QRS band, P band, and T band, with each clustering operation based on the previous clustering operation.

2. The beat clustering method of claim 1, wherein, The denoising includes: removing power frequency interference, removing baseline drift, and removing motion artifacts and electromyographic interference.

3. The beat clustering method of claim 1, wherein, The preprocessing involves segmenting and downsampling all heart rate data.

4. The beat clustering method of claim 1, wherein, The training data is classified using a sample data balancing method and a random forest algorithm.

5. The beat clustering method of claim 1, wherein, The training data uses raw heartbeat data without manual feature extraction.

6. A heart beat clustering system based on SHAP value weighting and segmented HDBSCAN, characterized in that, include: The extraction module is used to acquire ECG data from ECG databases and remove noise, including public databases and proprietary databases, and to extract heartbeat data from the ECG data. The preprocessing module is used to preprocess all heartbeat data to obtain training and test data; The weight module is used to classify the training data, calculate the contribution value of different features of all types of heartbeats in the classification task, and use it as the weight value in the clustering algorithm. The SHAP model is used to calculate the contribution values ​​of different features of all types of heartbeats in the ECG database in the classification task, i.e., the SHAP values, and these values ​​are used as weight values ​​in the clustering algorithm. The SHAP model is an additive interpretive model that can calculate the marginal contribution of each feature to the output of the classification algorithm model. The fusion module is used to fuse test data with weight values ​​to obtain weighted ECG feature point fusion data; The clustering module uses a segmented clustering method to perform three sequential clustering operations on the fused data to generate clustering results. The segmented clustering method performs three sequential clustering operations on the fused data, including using the HDBSCAN algorithm to perform three sequential clustering operations on the feature points of the QRS band, P band, and T band, with each clustering operation based on the previous clustering operation.

7. A non-transitory computer-readable storage medium, comprising: The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the heartbeat clustering method as described in any one of claims 1-5.

8. An electronic device, comprising: include: A processor, a memory and a computer program; wherein the processor is connected with the memory, the computer program is stored in the memory, when the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes the instructions for realizing the heart beat clustering method according to any one of claims 1-5.

Images