Multi-module collaborative rapid electrocardiogram automatic diagnosis method and system

Abstract

The present application discloses a multi-module collaborative rapid electrocardiogram automatic diagnosis and system, which belongs to the field of automatic diagnosis technology. The method includes: preprocessing the original ECG signal; taking the preprocessed ECG signal as input, and outputting a multi-level feature containing local details and global temporal dependencies after sequential processing by a lightweight convolution submodule, a deformable convolution submodule, and a non-local attention submodule; fusing the multi-level features and sending them to a classifier to complete the prediction of the heart rhythm type and obtain the prediction probability of each category; using a sliding window method, sending a fixed-length ECG signal segment to the diagnostic model for inference, so that the diagnostic model continuously outputs the diagnostic results, and outputs an alarm or diagnostic prompt information in a timely manner according to the diagnostic results. This method effectively reduces the number of diagnostic model parameters and computational complexity by designing a lightweight network structure, and realizes real-time online ECG signal automatic diagnosis while ensuring high diagnostic accuracy.

Description

Technical Field

[0001] The present application proposes a multi-module collaborative rapid electrocardiogram automatic diagnosis method and system, which belongs to the field of automatic diagnosis technology. Background Art

[0002] ECG, a noninvasive method for recording cardiac electrophysiological activity, has long been an important basis for the clinical diagnosis of cardiovascular disease. Traditional ECG analysis methods rely primarily on physicians' manual observation of waveform morphology and extensive clinical experience. This approach not only relies on specialized knowledge, but is also inefficient and susceptible to subjective influences. To address this issue, researchers have recently begun exploring automated ECG signal processing methods, striving to achieve objective, rapid, and highly accurate diagnosis. Commonly used methods in the signal preprocessing stage include high-pass filtering, morphological filtering, wavelet transform, bandpass filtering, and empirical mode decomposition. These methods effectively eliminate slow baseline drift, suppress power frequency interference, myoelectric noise, and other digital superimposed noise, while highlighting diagnostically critical signal components such as the QRS complex, P wave, and T wave. In the feature extraction and classification stages, researchers examine metrics such as R peak position, RR interval, and PT interval to reflect heart rate variability and basic waveform characteristics. Furthermore, they employ machine learning methods such as support vector machines (SVMs) and random forests to classify and identify manually extracted features. In recent years, deep neural networks have been introduced into ECG signal analysis, enabling end-to-end automatic feature extraction and classification, significantly improving diagnostic accuracy and robustness. However, with the increasing demand for applications such as telemedicine and real-time monitoring, the requirements for system real-time performance and lightweightness are also becoming increasingly stringent. While existing Holter monitoring devices can achieve long-term continuous monitoring, they often face backend latency issues when processing large amounts of data. Furthermore, due to computing resource and power consumption limitations, the online processing capabilities of smart wearable devices remain insufficient.

[0003] Currently, ECG signal analysis approaches can be broadly categorized into two main groups: those based on traditional signal processing and manual feature extraction, and automated approaches based on machine learning and deep neural networks. Traditional approaches typically employ techniques such as high-pass and morphological filters to remove baseline drift, locate QRS complexes through methods such as time-domain threshold detection, and incorporate expert-defined rules for preliminary arrhythmia identification. While these approaches have a clear theoretical basis and low implementation costs, they lack adaptability to complex noise interference and diverse waveforms, making them inadequate for clinical diagnosis with high diagnostic accuracy. Another approach builds on preprocessing by further reducing noise using techniques such as wavelet transforms or empirical mode decomposition. Features such as R-peak position, PR interval, QT interval, and heart rate variability are then extracted, and classification is performed using support vector machines (SVMs) or random forests. These approaches have achieved promising results on some publicly available datasets, but their reliance on tedious manual feature engineering limits their robustness to noise and individual variability. In recent years, with the continuous development of deep learning technology, researchers have applied diagnostic models such as convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) to ECG signal analysis, and further explored the use of Transformer diagnostic models for global modeling of signal segments. Although these methods have made significant progress in improving diagnostic accuracy and anti-interference capabilities, they are often accompanied by large parameter sizes and high computational complexity, making real-time deployment on wearable devices or edge computing platforms difficult.

[0004] In clinical practice, electrocardiogram (ECG) is a non-invasive means of recording cardiac electrophysiological activity and is widely used in the diagnosis of cardiovascular diseases. However, during the actual acquisition process, ECG signals often suffer from noise congestion and waveform distortion due to factors such as baseline drift, myoelectric interference, power frequency noise, and poor electrode contact. This makes it difficult to accurately identify key features (such as P waves, QRS complexes, and T waves), which may lead to misdiagnosis or missed diagnosis of conditions such as arrhythmias and myocardial ischemia. In addition, existing ECG diagnostic systems often rely on complex deep network structures or tedious manual feature engineering. The diagnostic model has large parameters, large computational complexity, and poor real-time performance, making it unsuitable for deployment on wearable devices or edge computing platforms.

[0005] Traditional filtering methods have the problem of insufficient noise adaptability when facing low-frequency baseline drift, mixed types of noise or extreme interference, and are prone to filtering or under-filtering, resulting in the loss of key ECG features or failure to effectively suppress noise. In addition, many machine learning-based methods rely too much on manually setting thresholds or extracting time domain, frequency domain, and morphological features, resulting in poor adaptability to different types of arrhythmias and low versatility. The reliance on manual feature engineering also increases the difficulty of algorithm updates and maintenance. At the same time, although solutions based on deep convolutional networks or Transformers have greatly improved accuracy, their network structure is complex and the number of parameters is large, making them unable to run in real time in resource-constrained hardware environments. The huge data requirements and training costs also limit the large-scale promotion of these diagnostic models in actual clinical practice. In addition, although some methods perform well in offline mode, they lack the ability to be quickly deployed and updated online when performing real-time inference and immediate alarms in remote medical monitoring or emergency situations. Summary of the Invention

[0006] To address these issues, this application proposes a lightweight, rapid electrocardiogram (ECG) diagnostic method that combines multiple noise interference correction with automatic feature extraction. By designing a lightweight network structure, this method effectively reduces the number of diagnostic model parameters and computational complexity, enabling real-time, online, automated ECG signal diagnosis while ensuring high diagnostic accuracy.

[0007] In order to solve the above technical problems, the technical solutions adopted in this application are:

[0008] In a first aspect, the present application provides a multi-module collaborative rapid electrocardiogram automatic diagnosis system, comprising:

[0009] A signal preprocessing module, used for preprocessing the original ECG signal;

[0010] A multi-level feature extraction module, including a lightweight convolution submodule, a deformable convolution submodule, and a non-local attention submodule. It takes the preprocessed ECG signal as input and, after sequential processing by the lightweight convolution submodule, the deformable convolution submodule, and the non-local attention submodule, outputs multi-level features containing local details and global temporal dependencies.

[0011] The feature fusion and classification module is used to fuse multi-level features and feed them into the classifier to complete the prediction of heart rhythm type and obtain the prediction probability of each category;

[0012] The online reasoning and real-time monitoring module is used to send ECG signal segments of fixed length to the diagnostic model for reasoning using a sliding window method, so that the diagnostic model can continuously output diagnostic results and output alarms or diagnostic prompts in a timely manner based on the diagnostic results.

[0013] As a further improvement of the present invention, the signal preprocessing module is specifically used to: perform noise reduction, filtering and amplitude normalization on the original ECG signal,

[0014] The filtering adopts an adaptive filter to filter out low-frequency baseline drift, a band-stop filter to suppress power frequency interference, and an empirical mode decomposition method to eliminate myoelectric interference;

[0015] The amplitude normalization process obtains a signal including a P wave, a QRS complex, and a T wave.

[0016] As a further improvement of the present invention, the lightweight convolution submodule uses depthwise separable convolution to achieve feature extraction, and the specific method is as follows:

[0017]

[0018] Where x(t) represents the preprocessed ECG signal input, W is the weight parameter of the lightweight convolution kernel, b is the bias term, BN represents the batch normalization operation, ReLU is the activation function, and y(t) represents the final output feature sequence of the lightweight convolution submodule, which can be subsequently input into the deformable convolution submodule.

[0019] As a further improvement of the present invention, the deformable convolution submodule adopts a method of adaptively adjusting the convolution sampling points, specifically:

[0020]

[0021] Where t represents the time index of the current convolution center point, w i is the weight parameter of the convolution kernel at index i, p i is a fixed offset, Δp i is the learnable deformation offset.

[0022] As a further improvement of the present invention, the non-local attention submodule is used to capture the long-range dependencies across cycles or abnormal bands in the ECG signal, construct a global correlation matrix for each sampling point in the sequence, and obtain the features of the diagnostic time series segment through reweighting operations. The specific method is as follows:

[0023]

[0024] Where x i 、x j denote the feature vectors of the ECG input at index i and j, respectively, f(x i ,x j ) is a similarity function used to measure x i with x j The degree of correlation between g(x j ) is the feature transformation function, xj Linear mapping is performed to extract fusible information, and C(x) is the normalization factor.

[0025] As a further improvement of the present invention, the feature fusion and classification module is specifically used to use point convolution to merge and compress multi-level features in the channel dimension, retain low-frequency and high-frequency features, and fuse them in a unified dimension; then introduce a residual skip connection mechanism to reuse early shallow features for feature fusion of multi-level features of unified dimension; the final fused features enter the classification layer through a lightweight classification head, and the output scores are normalized using the Softmax function to obtain the prediction probability of each category.

[0026] As a further improvement of the present invention, the diagnostic model is compressed into a half-precision floating point number or an integer number for real-time online diagnosis on a CPU or an edge device.

[0027] In a second aspect, the present application provides a multi-module collaborative rapid electrocardiogram automatic diagnosis method, comprising:

[0028] Preprocessing the raw ECG signal;

[0029] The preprocessed ECG signal is used as input and processed sequentially by the lightweight convolution submodule, the deformable convolution submodule, and the non-local attention submodule to output multi-level features containing local details and global temporal dependencies.

[0030] The multi-level features are fused and fed into the classifier to complete the prediction of the heart rhythm type and obtain the prediction probability of each category;

[0031] Using a sliding window approach, ECG signal segments of fixed length are sent to the diagnostic model for inference, so that the diagnostic model can continuously output diagnostic results and output alarms or diagnostic prompts in a timely manner based on the diagnostic results.

[0032] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the multi-module collaborative rapid electrocardiogram automatic diagnosis method when executing the computer program.

[0033] In a fourth aspect, the present application provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-module collaborative rapid electrocardiogram automatic diagnosis method is implemented.

[0034] In a fifth aspect, the present application provides a computer program product, which includes computer instructions, and the computer instructions instruct a computer to execute the multi-module collaborative rapid electrocardiogram automatic diagnosis method.

[0035] The beneficial effects of this application compared to the prior art are:

[0036] The system of this application combines high robustness and real-time online diagnostic accuracy in complex noisy environments. First, by designing a preprocessing module that combines adaptive filtering with a deep network, it effectively suppresses low-frequency baseline drift, myoelectric interference, and power-frequency noise in the ECG signal, while ensuring the complete preservation of key features such as the P wave, QRS complex, and T wave. Secondly, a multi-level feature extraction architecture is constructed. This architecture uses an organic fusion of lightweight convolutional networks, deformable convolutions, and non-local attention mechanisms to capture local details of the ECG signal while learning global temporal dependencies, achieving end-to-end automatic feature extraction and avoiding the limitations of manual feature engineering. Thirdly, through module pruning, weight sharing, and the combination of attention mechanisms and convolution operations, the diagnostic model parameters and computational complexity are significantly reduced, thereby meeting the real-time inference requirements of wearable devices and edge platforms. Compared with traditional methods, this technical solution demonstrates higher sensitivity and specificity in arrhythmia detection while ensuring high diagnostic accuracy, effectively reducing the risk of missed diagnosis and misdiagnosis, and is suitable for multiple scenarios such as telemedicine and bedside monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following introduction is made to the drawings of the embodiments of the present application or the related technical solutions in the prior art. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative work.

[0038] Figure 1 A block diagram of a multi-module collaborative rapid electrocardiogram automatic diagnosis system provided in this application;

[0039] Figure 2 Schematic diagram of the multi-level feature extraction module. DETAILED DESCRIPTION

[0040] The embodiments of the present application are described in detail below, and examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and are not to be construed as limiting the present application. For the step numbers in the following embodiments, they are provided only for the convenience of explanation and are not intended to limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0041] In the description of this application, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in this application based on the specific content of the technical solution.

[0042] This application combines adaptive filtering with lightweight deep networks, such as Figure 1 As shown in the figure, a multi-module collaborative rapid electrocardiogram automatic diagnosis system is proposed, which includes a signal preprocessing module, a feature extraction module, a feature fusion and classification module, and an online reasoning and real-time monitoring module.

[0043] A signal preprocessing module, used for preprocessing the original ECG signal;

[0044] A multi-level feature extraction module, including a lightweight convolution submodule, a deformable convolution submodule, and a non-local attention submodule. It takes the preprocessed ECG signal as input and, after sequential processing by the lightweight convolution submodule, the deformable convolution submodule, and the non-local attention submodule, outputs multi-level features containing local details and global temporal dependencies.

[0045] The feature fusion and classification module is used to fuse multi-level features and feed them into the classifier to complete the prediction of heart rhythm type and obtain the prediction probability of each category;

[0046] The online reasoning and real-time monitoring module is used to send ECG signal segments of fixed length to the diagnostic model for reasoning using a sliding window method, so that the diagnostic model can continuously output diagnostic results and output alarms or diagnostic prompts in a timely manner based on the diagnostic results.

[0047] The system can achieve multi-module collaboration. Under the premise of low parameter volume and computing consumption, it can still realize high-precision automatic ECG diagnosis and achieve real-time online reasoning. It can run efficiently on ordinary computers, wearable devices or edge computing platforms, and meet the application needs of multiple scenarios such as telemedicine and bedside monitoring.

[0048] The lightweight, rapid ECG automatic diagnosis method designed in this application is based on four major modules: signal preprocessing module, multi-level feature extraction module, feature fusion and classification module, and online reasoning and real-time monitoring module. The specific process of the automatic diagnosis method is as follows:

[0049] First, the signal preprocessing module performs noise reduction, filtering, and amplitude normalization on the original ECG signal, effectively eliminating baseline drift, myoelectric interference, and power frequency noise, ensuring that key signal components such as the P wave, QRS complex, and T wave are fully preserved.

[0050] Secondly, the preprocessed ECG signal is fed into a multi-level feature extraction module. This module uses a lightweight convolutional neural network combined with deformable convolution and non-local attention mechanism to construct local detail features and global temporal dependencies, respectively, to automatically extract multi-level features of the signal.

[0051] Next, the feature fusion and classification module effectively integrates the features extracted at each level, performs dimensionality reduction and fusion on the features through 1×1 convolution, and uses an attention weighting mechanism to highlight key bands, ultimately completing the accurate classification of arrhythmia types.

[0052] Finally, the online reasoning and real-time monitoring module designs a dedicated inference pipeline for continuous streaming ECG signals, adopts a sliding window strategy and diagnostic model compression and quantization technology to realize real-time online diagnosis of the system on a general-purpose CPU or edge device, and outputs alarms or diagnostic prompts in a timely manner.

[0053] Each module is described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] In the signal preprocessing module, this application first addresses the common low-frequency baseline drift problem in ECG signals and uses an adaptive filter to filter out the problem, ensuring that low-frequency components such as P waves and T waves are fully preserved. To suppress power frequency interference, a band-stop filter is set. At the same time, for electromyographic interference, empirical mode decomposition (EMD) technology is combined to further reduce high-frequency noise. After the above filtering processing, the signal is amplitude normalized to balance the voltage range differences between different measurement devices and different patients, providing a unified scale for the subsequent neural network input.

[0055] like Figure 2 Figure 2 shows the structure of the multi-level feature extraction module. Combining deformable convolution with the attention mechanism, based on lightweight convolution, it takes into account both local and global information extraction.

[0056] Figure 2 In the proposed method, the preprocessed ECG sequence is first input into the lightweight convolution module, and the depth convolution and point convolution are used to extract efficient features from each channel. Then, the deformable convolution module is used to adaptively sample the key bands on the time axis using the learnable offset. Then, the non-local attention module is used to capture the long-distance dependencies across cycles and abnormal bands based on global similarity calculation and weighted summation. Finally, the output is a multi-level feature representation, which provides refined feature support for subsequent classification.

[0057] This module receives the preprocessed ECG sequence x(t) as input. After sequential processing by the lightweight convolution submodule, the deformable convolution submodule, and the non-local attention submodule, it ultimately outputs a multi-level feature representation containing local details and global temporal dependencies, providing effective support for subsequent feature fusion and classification. First, in the lightweight convolution submodule, in order to reduce the number of parameters and computational complexity, this application draws on the design of lightweight networks such as MobileNet and uses depthwise separable convolution to achieve feature extraction. This module can be described by the following formula:

[0058]

[0059] Where x(t) represents the preprocessed ECG signal input, W is the weight parameter of the lightweight convolution kernel, b is the bias term, BN represents the batch normalization operation, ReLU is the activation function, and y(t) represents the final output feature sequence of the lightweight convolution submodule, which can be subsequently input into the deformable convolution submodule.

[0060] Next, the deformable convolution submodule proposes a method for adaptively adjusting convolution sampling points to account for morphological differences in ECG signals across different patients or physiological states. This module learns a set of offsets to fine-tune the conventional convolution kernel in the time domain, allowing for more flexible capture of key location features. Its computational formula can be expressed as:

[0061]

[0062] Where t represents the time index of the current convolution center point, w i is the weight parameter of the convolution kernel at index i, p i is a fixed offset, Δp i It is a learnable deformation offset, which is used to fine-tune the original fixed offset pi, thereby adapting to different waveform characteristics and individual differences.

[0063] Finally, the non-local attention submodule is used to capture long-range dependencies across cycles or abnormal bands in the ECG signal. This module constructs a global correlation matrix for each sampling point in the sequence and highlights the features of the most critical time segments for diagnosis through a reweighting operation. The formula is described as follows:

[0064]

[0065] Where x i 、x j denote the feature vectors of the ECG input at index i and j, respectively, f(x i ,x j ) is a similarity function used to measure x i with x jThe degree of correlation between g(x j ) is the feature transformation function, x j Linear mapping is performed to extract fusible information, and C(x) is the normalization factor.

[0066] Through this non-local operation, global information can be effectively integrated, enhancing the diagnostic model's ability to capture ECG periodicity and abnormal states.

[0067] In the feature fusion and classification module, after multi-level features are extracted, they need to be fused and fed into the classifier to complete the prediction of the heart rhythm type. First, point convolution is used to merge and compress the features output by each submodule in the channel dimension, thereby retaining both low-frequency and high-frequency features and unifying the dimensions of the output of each module to ensure full information fusion. In addition, to prevent information loss due to the deepening of the network layers, a residual skip connection mechanism is introduced to reuse early shallow features to enhance the overall feature representation capability. After feature fusion is completed, the final fused features will enter the classification layer through a lightweight classification head, and the output scores will be normalized using the Softmax function to obtain the prediction probability of each category.

[0068] In the online reasoning and real-time monitoring module, first of all, for clinical or remote monitoring application scenarios, since ECG signals are often transmitted in the form of streaming data, a sliding window method is used to input ECG segments of fixed length into the diagnostic model for reasoning and continuously output diagnostic results.

[0069] To meet actual deployment requirements, the diagnostic model can be further compressed or quantized to FP16 or even INT8 precision, enabling fast and efficient execution on CPUs, GPUs, or dedicated AI accelerator chips. When abnormalities such as atrial fibrillation and premature ventricular contractions are detected, the system quickly sends an alarm signal to the monitor or smart terminal, while simultaneously uploading detailed diagnostic results to the cloud or a doctor's workstation for further decision support.

[0070] As an example, this application proposes to significantly improve the running speed and energy efficiency of the model on various hardware platforms by compressing the model from its original high precision (such as FP32) to half-precision floating point numbers (FP16) or even integer numbers (INT8) without significantly sacrificing the model's accuracy. The diagnostic model of this application effectively reduces the model's storage requirements and computational complexity by flexibly selecting FP16 or INT8 precision, and improves the model's running efficiency and compatibility on various hardware platforms. At the same time, through sophisticated quantization-aware training and hardware acceleration optimization strategies, it ensures that the quantized model achieves significant performance improvements while maintaining high accuracy, providing strong technical support for the actual deployment of the diagnostic model.

[0071] In summary, this application proposes a multi-level feature representation module that combines deformable convolution and a non-local attention mechanism to perform multi-level feature extraction on ECG signals. Specifically, deformable convolution adaptively captures the offset of key waveform locations to construct a local detail representation; while the non-local attention mechanism calculates the long-range dependencies between locations in the signal to form global timing information. This multi-level representation ensures precise identification of various arrhythmias and improves diagnostic accuracy.

[0072] This application utilizes a lightweight convolutional backbone network structure, which forms the core of multi-level feature extraction, using depthwise separable convolution technology and a lightweight network design. This significantly reduces the number of diagnostic model parameters and computational complexity, thereby enabling real-time online inference. This design optimizes hardware resource utilization, enabling the system to run stably on wearable devices and edge computing platforms, meeting the needs of real-time application scenarios such as telemedicine and bedside monitoring.

[0073] A residual skip connection mechanism and feature fusion module were introduced. Based on the feature reuse and fusion strategy, this strategy not only effectively alleviates the problem of gradient vanishing in deep networks, but also enhances the overall feature expression ability by fusing redundant information extracted at each level, further improving the classification and discrimination performance of various arrhythmias.

[0074] Through testing, this application has obvious advantages in noise robustness. It combines adaptive filtering and deformable convolution technology, which can effectively deal with baseline drift and various noise interferences, while suppressing noise while ensuring that key ECG features are not over-filtered. Compared with pure Transformer or large deep convolutional networks, this application adopts a lightweight convolution design, which greatly reduces the parameter scale of the convolution kernel, thereby achieving real-time operation on general-purpose CPUs or mobile devices, meeting the needs of telemedicine and bedside monitoring. In addition, this application realizes end-to-end automatic feature extraction, avoiding reliance on tedious manual threshold settings and manual feature engineering, enabling the system to identify multiple types of arrhythmias, with greater versatility and scalability.

[0075] Furthermore, this application was tested on the publicly available PTB-XL arrhythmia database. Experimental results demonstrate that this method achieves an accuracy of 98.67% in multi-category detection tasks, including atrial fibrillation and premature ventricular contractions, and outperforms most existing comparison algorithms in both sensitivity and specificity. In simulations with baseline drift and power-frequency interference noise, this application maintains high classification accuracy and demonstrates more stable anti-interference capabilities under extreme interference conditions than existing methods.

[0076] Furthermore, this application can adapt to different application requirements through module replacement, diagnostic model optimization, and expansion. The deformable convolution submodule can be replaced with other deformable or dynamic convolution structures, and the non-local attention mechanism can be replaced by a Transformer encoder to achieve global dependency modeling. At the same time, the filtering method of the preprocessing module can be adjusted according to needs. To reduce the scale of the diagnostic model and inference latency, knowledge distillation or low-bit quantization can be used to improve deployment efficiency while maintaining accuracy.

[0077] A third object of an embodiment of the present application is to provide an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the aforementioned multi-module collaborative rapid electrocardiogram automatic diagnosis method. The electronic device also includes a communication interface and a bus.

[0078] The fourth object of the embodiment of the present application is to provide a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, it implements the above-mentioned multi-module collaborative rapid electrocardiogram automatic diagnosis method.

[0079] The fifth object of the embodiment of the present application is to provide a computer program product, which includes computer instructions, and the computer instructions instruct the computer to execute the above-mentioned multi-module collaborative rapid electrocardiogram automatic diagnosis method.

[0080] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0082] The present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, readable storage media, optical storage, etc.) containing computer-usable program code.

[0083] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0084] Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of this application.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit them. Although the present application has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present application can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present application should be included in the scope of protection of the claims of the present application.

Claims

1. A multi-module collaborative rapid electrocardiogram automatic diagnosis system, characterized by: include: A signal preprocessing module, used for preprocessing the original ECG signal; Multi-level feature extraction module, including lightweight convolution submodule, deformable convolution submodule and non-local attention submodule; It takes the preprocessed ECG signal as input, and after being processed sequentially by the lightweight convolution submodule, the deformable convolution submodule, and the non-local attention submodule, it outputs multi-level features containing local details and global temporal dependencies; The feature fusion and classification module is used to fuse multi-level features and feed them into the classifier to complete the prediction of heart rhythm type and obtain the prediction probability of each category; An online reasoning and real-time monitoring module is used to send fixed-length ECG signal segments to the diagnostic model for reasoning using a sliding window method, so that the diagnostic model can continuously output diagnostic results and promptly output alarms or diagnostic prompts based on the diagnostic results; The lightweight convolution submodule uses depthwise separable convolution to achieve feature extraction. The specific method is as follows: Where x(t) represents the preprocessed ECG signal input, W is the weight parameter of the lightweight convolution kernel, b is the bias term, BN represents the batch normalization operation, ReLU is the activation function, and y(t) represents the final output feature sequence of the lightweight convolution submodule, which can be subsequently input into the deformable convolution submodule. The deformation convolution submodule adopts the method of adaptively adjusting the convolution sampling points. The specific method is as follows: Where t represents the time index of the current convolution center point, w i is the weight parameter of the convolution kernel at index i, p i is a fixed offset, Δp i is the learnable deformation offset; The non-local attention submodule is used to capture the long-range dependencies across cycles or abnormal bands in the ECG signal, construct a global correlation matrix for each sampling point in the sequence, and obtain the features of the diagnostic time series segment through reweighting operations. The specific method is as follows: Where x i 、x j denote the feature vectors of the ECG input at index i and j, respectively, f(x i ,x j ) is a similarity function used to measure x i with x j The degree of correlation between g(x j ) is the feature transformation function, x j Perform linear mapping to extract fusible information, where C(x) is the normalization factor; The feature fusion and classification module is specifically used to use point convolution to merge and compress multi-level features in the channel dimension, retain low-frequency and high-frequency features, and unify the dimensionality for fusion; then introduce the residual skip connection mechanism to reuse the early shallow features of the multi-level features of the unified dimension for feature fusion; the final fused features enter the classification layer through a lightweight classification head, and the output scores are normalized using the Softmax function to obtain the prediction probability of each category.

2. A multi-module collaborative rapid electrocardiogram automatic diagnosis system according to claim 1, characterized in that: The signal preprocessing module is specifically used to: perform noise reduction, filtering and amplitude normalization on the original ECG signal, The filtering adopts an adaptive filter to filter out low-frequency baseline drift, a band-stop filter to suppress power frequency interference, and an empirical mode decomposition method to eliminate myoelectric interference; The amplitude normalization process obtains a signal including a P wave, a QRS complex, and a T wave.

3. The multi-module collaborative rapid electrocardiogram automatic diagnosis system according to claim 1, characterized in that: The diagnostic model is compressed into half-precision floating point numbers or integers for real-time online diagnosis on the CPU or edge devices.

4. A multi-module collaborative rapid electrocardiogram automatic diagnosis method, based on the multi-module collaborative rapid electrocardiogram automatic diagnosis system according to any one of claims 1 to 3, characterized in that: include: Preprocessing the raw ECG signal; The preprocessed ECG signal is used as input and processed sequentially by the lightweight convolution submodule, the deformable convolution submodule, and the non-local attention submodule to output multi-level features containing local details and global temporal dependencies. The multi-level features are fused and fed into the classifier to complete the prediction of the heart rhythm type and obtain the prediction probability of each category; Using a sliding window approach, ECG signal segments of fixed length are sent to the diagnostic model for inference, so that the diagnostic model can continuously output diagnostic results and output alarms or diagnostic prompts in a timely manner based on the diagnostic results.

5. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the multi-module collaborative rapid electrocardiogram automatic diagnosis method according to claim 4 is implemented.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the multi-module collaborative rapid electrocardiogram automatic diagnosis method according to claim 4 is implemented.

Images