Single-lead electrocardio arrhythmia intelligent detection method and system based on multi-view information decoupling and information bottleneck distillation

Abstract

The invention provides a single-lead electrocardio arrhythmia detection method based on multi-view information decoupling and information bottleneck distillation, which comprises the following steps of: S1, acquiring a multi-lead electrocardio signal and a single-lead electrocardio signal to obtain standardized electrocardio data; s2, constructing a multi-view teacher model, inputting the multi-lead electrocardiosignals as different views, and extracting multi-view electrocardiograph feature representation; s3, carrying out information decoupling on the multi-view electrocardio characteristics; s4, performing feature extraction on the single-lead electrocardiosignals, and migrating consistent information representation and view private information representation to a single-lead student model; s5, adopting a residual self-attention mechanism to carry out adaptive fusion on the information representation to obtain a minimum sufficient feature representation; and S6, outputting a corresponding arrhythmia detection result. On the premise of ensuring light weight of the model, the accuracy and robustness of single-lead electrocardio arrhythmia detection are effectively improved, and the method is suitable for long-term continuous monitoring and intelligent diagnosis application scenes of wearable electrocardio monitoring equipment.

Description

Technical Field

[0001] This invention relates to a method and system for efficiently transferring multi-lead electrocardiogram information to a lightweight single-lead model for arrhythmia detection, applicable to wearable electrocardiogram monitoring devices, and belonging to the field of biomedical signal processing technology. Background Technology

[0002] Cardiac arrhythmias are a common cardiovascular disease caused by abnormal cardiac electrical activity, characterized by insidious onset, strong intermittency, and serious complications. Clinically, a standard 12-lead electrocardiogram (ECG) is typically used for diagnosis, comprehensively reflecting cardiac electrophysiological activity from different spatial perspectives. However, in applications such as community screening, home monitoring, and long-term continuous monitoring, single-lead wearable ECG devices have become the mainstream choice due to limitations in device size, power consumption, and wearing comfort. Compared to 12-lead ECG signals, single-lead ECGs have the following technical drawbacks: limited spatial perspective, lack of complementary diagnostic information between multiple leads; insufficient ability to identify certain arrhythmia types, leading to decreased diagnostic accuracy; and limited computing resources in wearable devices, making it difficult to directly deploy complex, high-performance models.

[0003] Existing knowledge distillation methods typically treat the entire multi-lead ECG signal as a single input, directly transferring the output or intermediate features of the teacher model to the student model. This approach fails to adequately consider the multi-view characteristic of multiple leads, which contains both consistent diagnostic information and their own proprietary diagnostic information. This approach is prone to: insufficient transfer of multi-lead coupling information; estimation bias of consistent information by the single-lead model; and missing proprietary information, affecting diagnostic reliability.

[0004] Therefore, how to fully transfer consistent and private information from multi-lead ECGs while suppressing redundant noise information, while ensuring the model's lightweight nature, is a key technical problem that urgently needs to be solved in the intelligent detection of single-lead arrhythmias. Summary of the Invention

[0005] Technical Objective: This invention aims to overcome the problems of low efficiency in transferring multi-lead ECG information to single-lead models, severe information coupling, and high model redundancy in existing technologies. It proposes an intelligent detection method and system for single-lead ECG arrhythmias based on multi-view information decoupling and information bottleneck distillation. This achieves: effective decoupling of consistent multi-lead ECG information from private information; fully compensated transfer of multi-lead private information to single-lead models; and high-precision, low-complexity diagnosis using single-lead models in computationally limited scenarios.

[0006] Technical solution: The complete technical means and methods of this invention.

[0007] This invention proposes a single-lead electrocardiogram arrhythmia detection method based on multi-view information decoupling and information bottleneck distillation, comprising the following steps:

[0008] Step S1: Acquisition and Preprocessing of Multi-Lead ECG Data

[0009] The raw ECG signal data containing multiple ECG leads is acquired, and the signals are filtered, denoised, segmented, and normalized to construct a multi-view ECG input sample, denoted as:

[0010]

[0011] in Indicates the first A view of the electrocardiogram leads.

[0012] Step S2: Construction and Feature Extraction of the Multi-View Teacher Model

[0013] A multi-view teacher network model is constructed, in which ECG signals from different leads are used as inputs for different views. Through shared or non-shared feature extraction networks, multi-view feature representations (high-dimensional feature representations) of each view are extracted.

[0014] Step S3: Multi-view Figure 1 Decoupling of public information and private information

[0015] Based on the information bottleneck theory, a variational estimation method is used to decouple the features of each view, resulting in:

[0016] Consistent information representing shared diagnostic criteria across multiple leads ;

[0017] Private information representation of views characterizing the unique diagnostic features of each lead ;

[0018] The consistent information representation and the view-private information representation are decoupled by minimizing their mutual information, thereby achieving decoupling of multi-view information. The constraint objective function is expressed as: ,in, For mutual information, the above constraints ensure that consistent information mainly includes translead stability diagnostic information, while view-specific information mainly includes the unique diagnostic criteria of each lead. Information decoupling employs a variational information bottleneck model, achieved through variational approximation estimation of the multi-view feature distribution.

[0019] Step S4: Construction of the single-lead student model

[0020] A single-lead ECG signal is acquired and input into a single-lead student model (a lightweight neural network model suitable for deployment in wearable ECG monitoring devices), and single-view feature representations are extracted. Specifically, a single-lead student model for wearable device deployment is constructed, using only the single-lead ECG signal. As input, extract the initial single-view feature representation. .

[0021] Step S5: Distillation of multi-view private information and alignment of consistent information

[0022] Through knowledge distillation, the consistent information representation and view-private information representation learned in the multi-view teacher model are transferred to the single-lead student model to compensate for the missing multi-lead information in the single-lead ECG signal. Specifically, this is achieved by maximizing the single-lead features and multi-view... Figure 1 The joint mutual information of the information and the private information of the view enables the compensation transfer of multi-lead private information to the single-lead model, and its distillation constraint is expressed as:

[0023]

[0024] Meanwhile, by constraining the student model's estimation of the consistent information representation to maintain consistency with the consistent information representation in the teacher model, unbiased alignment of consistent information is achieved.

[0025] Step S6: Feature Fusion and Information Bottleneck Compression

[0026] A residual self-attention mechanism is employed to adaptively fuse consistent and private information. A variational information bottleneck constraint (achieved by minimizing the mutual information between the fused features and the input signal while maximizing the mutual information between the fused features and the arrhythmia category label) is used to discard redundant noise information and obtain the minimum sufficient feature representation. The objective function for this information bottleneck is expressed as:

[0027]

[0028] in, This indicates the corresponding arrhythmia category label.

[0029] This yields the minimum sufficient feature representation for arrhythmia detection.

[0030] Step S7: Arrhythmia Classification Output

[0031] The compressed features are input into the classifier, which outputs the arrhythmia detection results corresponding to the single-lead ECG signal (including the classification results of at least one type of arrhythmia and / or the corresponding occurrence probability).

[0032] This invention also proposes a single-lead ECG arrhythmia detection system based on multi-view information decoupling and information bottleneck distillation. The system is used to implement the single-lead ECG arrhythmia detection method described in the above method embodiments, and includes the following modules:

[0033] Multi-lead data acquisition module, used to acquire multi-lead electrocardiogram signals;

[0034] The multi-view teacher model module is used to extract multi-view feature representations of multi-lead ECG signals;

[0035] The information decoupling module is used to decouple the multi-view feature representation into a consistent information representation and a view-private information representation;

[0036] The single-lead student model module is used to extract feature representations of single-lead ECG signals;

[0037] The knowledge distillation module is used to migrate consistent information representations and view-private information representations to a single-lead student model;

[0038] The information bottleneck compression and classification module is used to output the arrhythmia detection results.

[0039] Furthermore, the information decoupling module includes a mutual information estimation unit based on variational information bottleneck; the knowledge distillation module includes a joint mutual information maximization unit for compensating for missing multi-lead private information in single-lead ECG signals; and the information bottleneck compression and classification module includes a residual self-attention fusion unit.

[0040] Beneficial effects: Compared with the prior art, the present invention has at least the following beneficial effects:

[0041] 1. Improve the diagnostic performance of single-lead ECG: When using only a single-lead input, it can approximate or reach the diagnostic performance of a multi-lead model.

[0042] 2. Achieve interpretable transfer of multi-lead information: Clearly distinguish between consistent information and private information to improve the interpretability and stability of the knowledge transfer process.

[0043] 3. Reduce model complexity and adapt to wearable devices: Through the information bottleneck mechanism, effectively compress redundant information in the model and reduce storage and computing overhead.

[0044] 4. Enhance model generalization ability and noise resistance: Suppress irrelevant noise interference and improve adaptability to different groups of people and different equipment conditions. Attached Figure Description

[0045] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0046] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a flowchart illustrating the first embodiment of a single-lead electrocardiogram arrhythmia detection method based on multi-view information decoupling and information bottleneck distillation according to this application.

[0048] Figure 2 This is a schematic diagram of a single-lead electrocardiogram arrhythmia detection method based on multi-view information decoupling and information bottleneck distillation according to an embodiment of this application.

[0049] Figure 3 This is a flowchart illustrating the third embodiment of the electrocardiogram signal classification method of this application.

[0050] Figure 4 This is a schematic diagram of the device structure of the hardware operating environment involved in the embodiments of this application. Detailed Implementation

[0051] It should be understood that the specific embodiments described herein are only for explaining the technical solutions of this application and are not intended to limit the scope of protection of this application. For those skilled in the art, any equivalent substitutions or modifications made to these embodiments without departing from the technical concept of this application should be covered within the scope of protection of this application.

[0052] To better understand the technical solution of this application, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] In existing technologies, to address the issue of insufficient diagnostic information in single-lead ECG signals, some studies have attempted to transfer knowledge from multi-lead ECG models to single-lead models through feature splicing, multi-lead feature fusion, or direct knowledge distillation. However, these methods typically treat multi-lead ECG signals as a whole, failing to distinguish between shared diagnostic information across leads and the unique diagnostic information specific to each lead. This leads to problems such as information redundancy, information interference, and consistency estimation bias during the transfer process. Furthermore, single-lead models, lacking auxiliary information from other leads, struggle to effectively compensate for the unique diagnostic features in multi-lead models, thus limiting their performance improvement in complex arrhythmia detection tasks.

[0054] In view of this, this application proposes a single-lead ECG arrhythmia detection method based on multi-view information decoupling and information bottleneck distillation. By constructing a multi-view teacher model and a single-lead student model, the multi-lead ECG features are decoupled and modeled from an information theory perspective. This explicitly separates the common consistent information shared by multiple leads from the private information unique to each lead. Furthermore, through joint mutual information distillation and information bottleneck compression, efficient transfer of multi-lead diagnostic knowledge to the single-lead model is achieved. This significantly improves the accuracy and robustness of single-lead arrhythmia detection while maintaining a lightweight model. This method is applicable to wearable ECG monitoring devices, remote arrhythmia monitoring systems, and intelligent health management platforms.

[0055] The first embodiment is a single-lead ECG arrhythmia detection method based on multi-view information decoupling and information bottleneck distillation, referring to... Figure 1 and 2 In this embodiment, the execution entity of the method can be a server, a personal computer, or a wearable device with computing and storage capabilities. This embodiment aims to achieve effective transfer of multi-lead ECG diagnostic knowledge to a single-lead model through multi-view information decoupling and information bottleneck distillation. Specifically, the method includes steps S10 to S70.

[0056] Step S10: Acquisition and Preprocessing of Multi-Lead ECG Data

[0057] The process involves acquiring raw ECG signal data containing multiple leads and performing preprocessing operations on the raw ECG signals. These preprocessing operations include, but are not limited to, bandpass filtering, noise reduction, signal segmentation, and amplitude normalization, thereby constructing multi-view ECG input samples for model training. A multi-lead ECG signal can be represented as:

[0058]

[0059] in Indicates the first A view of the electrocardiogram leads.

[0060] Step S20: Construction and Feature Extraction of the Multi-View Teacher Model

[0061] A multi-view teacher model is constructed, in which ECG signals from different leads are used as inputs for different views. Through shared or non-shared feature extraction networks, features are extracted from the ECG signals of each lead to obtain the corresponding high-dimensional feature representations.

[0062] In one feasible implementation, the multi-view teacher model can adopt a convolutional neural network, a residual network, or a combination thereof. The feature extraction network structure corresponding to each lead can be the same, or it can be designed differently according to the characteristics of the leads. This embodiment does not impose any restrictions on this.

[0063] Step S30: Multi-view Figure 1 Decoupling of public information and private information

[0064] Based on the information bottleneck theory, information decoupling modeling is performed on multi-view features to separate the following:

[0065] Consistent information representing shared diagnostic criteria across multiple leads ;

[0066] Private information representation of views characterizing the unique diagnostic features of each lead ;

[0067] In this context, the consistent information representation and the view-private information representation are decoupled by minimizing their mutual information, and the constraint objective function is expressed as follows: ,in, For mutual information,

[0068] Through the above constraints, the consistent information mainly includes the diagnostic features that are stable across leads, while the view-specific information mainly reflects the diagnostic details unique to each lead, thereby achieving effective decoupling of multi-lead features.

[0069] Step S40: Construction of the single-lead student model

[0070] A single-lead student model is constructed for deployment in wearable devices. This student model uses only a single-lead ECG signal as input and extracts the corresponding single-view feature representation. In one feasible implementation, the student model employs a lightweight network structure to reduce computational complexity and storage overhead, thereby meeting the real-time inference requirements of wearable devices.

[0071] Step S50: Distillation of multi-view private information and alignment of consistent information

[0072] Through knowledge distillation, the consistent information representation and view-private information representation learned in the multi-view teacher model are transferred to the single-lead student model. Specifically, this is achieved by maximizing the single-lead features and multi-view... Figure 1 The joint mutual information of the information and the private information of the view enables the compensation transfer of multi-lead private information to the single-lead model, and its distillation constraint is expressed as:

[0073]

[0074] Meanwhile, by constraining the student model's estimation of consistent information representation to be consistent with the teacher model, unbiased alignment of consistent information is achieved.

[0075] Step S60: Feature Fusion and Information Bottleneck Compression

[0076] A residual self-attention mechanism is used to adaptively fuse consistent information representation and view-private information representation, and redundant noise information is removed by variational information bottleneck constraint. The objective function is expressed as:

[0077]

[0078] in, This represents the corresponding arrhythmia category label, thus obtaining the minimum sufficient feature representation for arrhythmia detection.

[0079] Step S70: Arrhythmia Classification Output

[0080] The features, after being compressed by the information bottleneck, are input into the classifier, which outputs the arrhythmia detection results corresponding to the single-lead ECG signal.

[0081] Through the above steps, this embodiment effectively improves the accuracy and robustness of single-lead ECG arrhythmia detection while ensuring the model's lightweight nature.

[0082] Second Embodiment: Iterative Optimization Method for Multi-View Information Decoupling and Distillation Training. In this second embodiment, content identical or similar to that in the first embodiment will not be repeated. This embodiment further iteratively optimizes the training process in steps S50-S60 to enhance the stability of knowledge transfer.

[0083] In one feasible implementation, the distillation training process includes the following sub-steps:

[0084] Supervised fine-tuning of the teacher model was performed using labeled multi-lead electrocardiogram signals.

[0085] With fixed teacher model parameters, the student model undergoes multiple rounds of distillation training.

[0086] In each round of training, a total loss function is constructed based on classification loss, joint mutual information distillation loss, and information bottleneck loss, and the student model parameters are updated through backpropagation.

[0087] When the preset training termination condition is met, the final trained single-lead ECG arrhythmia detection model is output.

[0088] Through the above iterative optimization methods, the transfer process of multi-view diagnostic knowledge to single-lead models can be further stabilized, and the generalization ability of the model under different devices, different populations and different acquisition conditions can be improved.

[0089] The following is a systematic analysis of the experimental results of the single-lead arrhythmia detection method based on multi-view information decoupling and information bottleneck distillation proposed in the first and second embodiments of this application (hereinafter referred to as the method of this invention) on the publicly available ECG datasets PTB-XL and G12EC (both databases contain multi-lead ECG signals and corresponding arrhythmia annotation information, which can cover a variety of common and complex arrhythmia types). The method is also compared with a variety of advanced knowledge distillation methods to verify the effectiveness and advantages of the method of this invention in cross-lead knowledge transfer and single-lead modeling.

[0090] This experiment uses ResNet18 and ShuffleNet V2 as the backbone networks of the student model, and evaluates them on the PTB-XL and G12EC datasets, respectively. Comparison methods include representative knowledge distillation methods such as baseline (no distillation), OKD distillation, FitNet distillation, and AT (Attention Transfer) distillation. The performance metric is classification accuracy, and evaluations are performed on different single-lead views (views). Figure 1 Statistics were performed under the -12) and overall view, and the corresponding results are shown in Table 1 and Table 2, respectively.

[0091] Experimental results analysis based on the ResNet student model (Table 1): As shown in Table 1, on both the PTB-XL and G12EC datasets, regardless of whether ResNet is used as the student model or under different lead view configurations, the proposed method (Ours) consistently outperforms all comparative methods in both overall and single-lead performance. Overall view (all leads) performance analysis: On the PTB-XL dataset, the proposed method achieved accuracies of 0.914 (Table 1) and 0.907 (Table 2) in the "all" views, representing improvements of approximately 2.0%–2.4% compared to the baseline method and approximately 1.4%–1.8% compared to the AT method. On the G12EC dataset, the proposed method also achieved the best performance of 0.887 (Table 1) and 0.885 (Table 2), significantly outperforming the OKD, FitNet, and AT methods. This result demonstrates that by using multiple views… Figure 1 By decoupling the modeling of primary and secondary information and combining it with an information bottleneck distillation mechanism, the method of this invention can more fully compress redundant noise information at the overall level while retaining the most discriminative diagnostic features for arrhythmia.

[0092] Single-lead view performance analysis: in the view Figure 1In single-lead evaluations of -12, the method of this invention achieved best or near-best results in the vast majority of leads, with a highly consistent advantage. For example, on the PTB-XL dataset, the method of this invention achieved accuracies of 0.872–0.869 or higher in key leads such as views 4, 6, 10, and 12, significantly outperforming other distillation methods. On the G12EC dataset, the method of this invention achieved leading performance of 0.860–0.875 in multiple lead views, with even more significant advantages in leads with stronger noise and greater individual variability. This demonstrates that the "multi-lead private information compensation distillation" mechanism proposed in this invention effectively compensates for the inherent deficiencies of single-lead models in terms of spatial diagnostic information, enabling student models to learn near-multi-lead discrimination capabilities under single-lead input conditions.

[0093] Analysis of Experimental Results Based on the ShuffleNet V2 Student Model (Table 2): Table 2 presents the comparative experimental results under the condition of the more lightweight ShuffleNet V2 student model, used to verify the applicability of the method of the present invention in resource-constrained scenarios. Experimental results show that on the PTB-XL dataset, the method of the present invention achieves an accuracy of 0.907 in the "All" view, significantly outperforming Base (0.890) and AT (0.889); on the G12EC dataset, the method of the present invention also achieves the best performance of 0.887, maintaining a stable lead in most single-lead views. It is worth noting that in networks like ShuffleNet V2 with smaller parameter sizes and limited expressive power, the performance improvement of traditional distillation methods (such as FitNet and AT) is significantly limited, while the method of the present invention still maintains a stable advantage. This further illustrates that by introducing information bottleneck constraints and a multi-view decoupling distillation mechanism, the method of the present invention can not only improve model accuracy but also effectively improve the feature utilization efficiency of lightweight models, making it particularly suitable for application scenarios with limited computing resources, such as wearable devices.

[0094] Based on the experimental results in Tables 1 and 2, the following conclusions can be drawn: The method of this invention achieves consistent and significant performance improvements in both the overall view and single-lead view on both the PTB-XL and G12EC datasets; compared with traditional distillation methods such as OKD, FitNet and AT, the method of this invention shows stronger stability and generalization ability in cross-lead knowledge transfer.

[0095] The method of this invention maintains its advantages in student models of varying complexity, such as ResNet18 and ShuffleNet V2, verifying its model independence and engineering adaptability; the performance gap between single-lead and multi-lead models is significantly reduced, proving the effectiveness of multi-view... Figure 1 The effectiveness of decoupling and distilling of primary and secondary information in ECG diagnostic tasks is demonstrated. Therefore, the experimental results fully validate that the single-lead ECG arrhythmia detection method proposed in this invention, based on multi-view information decoupling and information bottleneck distillation, can significantly improve the detection accuracy and robustness of the single-lead model without increasing inference costs. It is particularly suitable for practical applications such as wearable ECG monitoring, remote arrhythmia screening, and personalized health management.

[0096] Table 1 Results using ResNet18 as the backbone model

[0097]

[0098] Table 2 Results using ShuffleNet V2 as the backbone model

[0099]

[0100] Third Embodiment: A Single-Lead ECG Signal Classification Method and Apparatus Based on Multi-View Information Decoupling and Information Bottleneck Distillation. Based on the first and / or second embodiments described above, a third embodiment of the ECG signal classification method of this application is proposed. In this embodiment, content that is the same as or similar to the first and second embodiments described above can be referred to the above description and will not be repeated hereafter.

[0101] In this embodiment, the executing entity of the ECG signal classification method can be a device with program execution function, specifically a general computing service device such as a server, personal computer, or mobile phone, or a wearable device with ECG signal acquisition and processing capabilities.

[0102] In this embodiment, refer to Figure 3 The electrocardiogram signal classification method includes steps A10 to A20.

[0103] Step A10: Acquire the target single-lead ECG signal to be detected.

[0104] Acquire the target single-lead electrocardiogram signal to be detected.

[0105] Among them, the target single-lead ECG signal refers to the raw ECG waveform data acquired through a single-lead method that requires automatic arrhythmia classification.

[0106] In specific implementations, in personal health management and daily monitoring application scenarios, the target single-lead ECG signal can be obtained from wearable devices such as smartwatches, smart bracelets, or disposable ECG patches that users wear for extended periods, either in real time or intermittently. In remote cardiac rehabilitation, home-based elderly care, or chronic disease management scenarios, the target single-lead ECG signal can also be obtained through portable home ECG monitoring devices and transmitted wirelessly or via wired means to a cloud server or home gateway.

[0107] In one feasible implementation, the original target single-lead ECG signal acquired can undergo a preprocessing procedure consistent with the model training phase before being input into the ECG signal classification model. For example, noise suppression can be performed using a bandpass filtering algorithm with the same parameter settings, baseline drift can be eliminated using the same baseline correction method, and amplitude normalization can be performed using the same scale. This eliminates the signal distribution differences caused by different devices and different acquisition environments, ensures the consistency between the input data and training data distribution during the model inference phase, and improves the stability and reliability of the classification results.

[0108] Step A20: Classify arrhythmias based on ECG signal arrhythmia classification model and output the classification results.

[0109] The target single-lead ECG signal obtained and preprocessed in step A10 is input into the ECG signal classification model for classification processing to obtain the arrhythmia classification result corresponding to the target single-lead ECG signal.

[0110] The electrocardiogram (ECG) signal classification model is trained using the ECG signal classification model training method based on multi-view information decoupling and information bottleneck distillation as described in the first or second embodiment above.

[0111] Specifically, the preprocessed target single-lead ECG signal is input into an ECG signal classification model that has been trained and deployed. The model extracts features from the target single-lead ECG signal through its feature encoder to obtain a compact and discriminative ECG feature representation. Subsequently, the classifier in the model outputs the corresponding predicted probability distribution based on the feature representation.

[0112] The predicted probability distribution is a vector, where each element corresponds to a preset arrhythmia category, such as normal rhythm, atrial fibrillation, premature ventricular contractions, etc. The value of each element in the vector represents the confidence level that the target ECG signal belongs to the corresponding category. The ECG signal classification result can be determined based on this probability distribution, for example:

[0113] The probability distribution is directly output as the result of the quantitative analysis.

[0114] The category with the highest confidence level was selected as the final arrhythmia classification label;

[0115] Alternatively, multiple judgment thresholds can be set based on clinical application needs to achieve graded risk alerts or warnings.

[0116] Because the ECG signal classification model explicitly decouples the consistent diagnostic information and lead-specific diagnostic information in multi-lead ECG signals through a multi-view teacher model during the training phase, and transfers the key diagnostic priors in multi-lead ECG signals to the single-lead student model through an information bottleneck distillation mechanism, the ECG signal classification model can still achieve high-precision multi-class arrhythmia identification even if only a single-lead ECG signal with limited information is input during the actual inference phase.

[0117] In this embodiment, the ECG signal classification model trained using the aforementioned multi-view information decoupling and information bottleneck distillation framework is used to classify target single-lead ECG signals collected from different wearable devices. In practical applications, this significantly improves the automatic diagnosis performance of arrhythmias in portable and wearable scenarios. Because this training framework effectively alleviates the diagnostic performance degradation caused by the lack of lead spatial information and limited labeled data in single-lead devices, the single-lead model has learned stable spatial diagnostic patterns and robust feature representations from multi-lead ECG signals without relying on a large amount of labeled single-lead data. Therefore, it can maintain good generalization ability and robustness when facing actual single-lead signals containing noise, individual differences, or device differences, providing reliable technical support for large-scale, low-cost cardiac health screening and long-term monitoring.

[0118] Electronic device examples

[0119] This application also provides an electronic device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is able to execute the ECG signal classification model training method or the ECG signal classification method in the above embodiments.

[0120] The following reference Figure 4 It shows a schematic diagram of the structure of an electronic device suitable for implementing the embodiments of this application. Figure 4 The electronic devices shown are merely examples and should not be construed as limiting the functionality and scope of use of the embodiments of this application.

[0121] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for detecting single-lead electrocardiographic arrhythmias based on multi-view information decoupling and information bottleneck distillation, characterized in that, Includes the following steps: S1: Acquire multi-lead ECG signal data, preprocess the multi-lead ECG signals, and construct multi-view ECG input samples; S2: Input the multi-view ECG input samples into the multi-view teacher model and extract the corresponding multi-view feature representations; S3: Based on the information bottleneck theory, information decoupling is performed on the multi-view feature representation to obtain a consistent information representation that characterizes the common diagnostic basis of multiple leads and a view-private information representation that characterizes the unique diagnostic features of each lead; S4: Acquire a single-lead ECG signal and input the single-lead ECG signal into a single-lead student model to extract single-view feature representation; S5: Through knowledge distillation, the consistent information representation and view private information representation in the multi-view teacher model are transferred to the single-lead student model to compensate for the missing multi-lead information in the single-lead ECG signal; S6: Perform feature fusion on the migrated consistent information representation and the view private information representation, and remove redundant information through information bottleneck constraints to obtain the minimum sufficient feature representation; S7: Output the corresponding arrhythmia detection result based on the minimum sufficient feature representation.

2. The method according to claim 1, characterized in that, In step S1, the multi-view ECG input sample is recorded as follows: in Indicates the first A view of the electrocardiogram leads.

3. The method according to claim 1, characterized in that, In step S3, the mutual information between the consistent information representation and the view private information representation is minimized, and the objective function of the constraint is expressed as: in, For mutual information, To characterize the consistent information representation of shared diagnostic criteria across multiple leads, This is a view-specific information representation that characterizes the unique diagnostic features of each lead.

4. The method according to claim 1, characterized in that, In step S3, the information decoupling adopts a variational information bottleneck model, which is achieved by performing variational approximation estimation on the multi-view feature distribution.

5. The method according to claim 1, characterized in that, In step S4, a single-lead student model for wearable device deployment is constructed, using only single-lead ECG signals. As input, extract the initial single-view feature representation. .

6. The method according to claim 1, characterized in that, In step S5, by maximizing the joint mutual information between the features of the single-lead student model and the view-private information in the multi-view teacher model, the compensation transfer of multi-lead private information to the single-lead model is achieved. The distillation constraint is expressed as follows: in, For mutual information, This represents the initial single-view feature representation. To characterize the consistent information representation of shared diagnostic criteria across multiple leads, A view-specific information representation that characterizes the unique diagnostic features of each lead; Meanwhile, by constraining the estimation of the consistent information representation in the single-lead student model, it is consistent with the consistent information representation in the multi-view teacher model, thus achieving unbiased transfer of consistent information.

7. The method according to claim 1, characterized in that, In step S6, the information bottleneck constraint is achieved by minimizing the mutual information between the fused features and the input signal, while maximizing the mutual information between the fused features and the arrhythmia category label; the information bottleneck objective function is expressed as: in, For mutual information, To characterize the consistent information representation of shared diagnostic criteria across multiple leads, This indicates the corresponding arrhythmia category label. This is a single-lead ECG signal.

8. The method according to claim 1, characterized in that, The arrhythmia detection results include classification results for at least one type of arrhythmia and / or the corresponding probability of occurrence.

9. A single-lead electrocardiogram arrhythmia detection system based on multi-view information decoupling and information bottleneck distillation, characterized in that, include: Multi-lead data acquisition module, used to acquire multi-lead electrocardiogram signals; The multi-view teacher model module is used to extract multi-view feature representations of multi-lead ECG signals; The information decoupling module is used to decouple the multi-view feature representation into a consistent information representation and a view-private information representation; The single-lead student model module is used to extract feature representations of single-lead ECG signals; The knowledge distillation module is used to migrate consistent information representations and view-private information representations to a single-lead student model; The information bottleneck compression and classification module is used to output the arrhythmia detection results.

10. The system according to claim 9, characterized in that, The information decoupling module includes a mutual information estimation unit based on variational information bottleneck; the knowledge distillation module includes a joint mutual information maximization unit, used to compensate for the missing multi-lead private information in single-lead ECG signals; the information bottleneck compression and classification module includes a residual self-attention fusion unit.

Images