Pruning lightweight heart sound classification method guided by intelligent agent
By employing a dual-pooling lightweight adaptive Mamba model and an agent-guided pruning quantization method, the feature extraction problem of the heart sound classification model in complex noise environments is solved, achieving efficient lightweighting on resource-constrained devices and ensuring high classification accuracy.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGBO UNIV
- Filing Date
- 2026-02-04
- Publication Date
- 2026-05-26
AI Technical Summary
Existing heart sound classification models have limited feature extraction capabilities in complex noisy environments, making it difficult to achieve high accuracy with lightweight implementation, which limits their application on resource-constrained devices.
We employ a dual-pooling lightweight adaptive Mamba model combined with an agent-guided pruning and quantization method. Feature extraction is performed through depthwise separable convolution and adaptive wavelet activation function. The model structure is optimized using an agent, and pruning and quantization compression are then carried out.
It maintains efficient and stable heart sound classification capabilities in variable environments, reduces the number of model parameters, improves operating efficiency, and is suitable for resource-constrained devices.
Smart Images
Figure CN122091193A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of medical signal processing and artificial intelligence, and in particular to an agent-guided pruning, quantization, and lightweight heart sound classification method. Background Technology
[0002] In recent years, the number of deaths caused by cardiovascular diseases has been increasing year by year, with the majority occurring in developing countries. Analysis of heart sounds can identify problems such as heart murmurs, arrhythmias, and valvular diseases, thus providing doctors with auxiliary diagnostic information. The heart ensures blood circulation and plays a decisive role in maintaining vital functions and health. Therefore, analyzing heart sounds to assist doctors in diagnosing heart sounds and promptly detecting potential cardiovascular diseases has significant practical and economic value. However, the hospital environment is complex and variable, and heart sounds generated during a heartbeat are significantly affected by various environmental factors. First, blood flow and respiration introduce interference signals; second, factors such as walking noise generate noise. These influences often cause the amplitude and frequency of heart sounds to deviate from ideal conditions, posing a significant challenge to heart sound auscultation. Finally, in the era of big data, the amount of data is increasing exponentially, making the design of a lightweight heart sound diagnostic model particularly important.
[0003] However, current common solutions have many drawbacks, including: existing traditional models have limited feature extraction capabilities, often employing a single network structure, making it difficult to robustly capture key disease features under complex noise interference such as respiration and blood flow, resulting in unstable classification accuracy; to achieve high accuracy, models are often designed to be complex and cumbersome, with a huge number of parameters and computational load, making it difficult to deploy on resource-constrained mobile or embedded medical devices; commonly used model compression methods (such as pruning and quantization) mostly rely on manual settings or globally unified strategies, lacking fine-grained differentiation of the importance of different levels, which can easily lead to a significant decrease in the accuracy of the compressed model, making it impossible to achieve effective lightweighting while maintaining high accuracy, thus limiting its practical application in real-time diagnosis and telemedicine. Summary of the Invention
[0004] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.
[0005] In view of the problems existing in the current agent-guided pruning quantization lightweight heart sound classification method, this invention is proposed.
[0006] Therefore, the purpose of this invention is to provide an agent-guided, pruning, quantization, and lightweight heart sound classification method. This method addresses the limitations of existing traditional models in feature extraction, which often employ a single network structure and struggle to robustly capture key disease features under complex noise interference such as respiration and blood flow, leading to unstable classification accuracy. To achieve higher accuracy, models are often designed to be complex and cumbersome, with a huge number of parameters and computational loads, making them difficult to deploy on resource-constrained mobile or embedded medical devices. Commonly used model compression methods (such as pruning and quantization) often rely on manual settings or globally unified strategies, lacking fine-grained differentiation of the importance of different levels, which can easily lead to a significant decrease in model accuracy after compression. This makes it impossible to achieve effective lightweighting while maintaining high accuracy, limiting its practical application in real-time diagnosis and telemedicine.
[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide an agent-guided pruning quantization lightweight heart sound classification method, which includes acquiring heart sound data from patients with different cardiovascular diseases and labeling them with categories to form a heart sound dataset; the heart sound data is a collected time-series heart sound signal; preprocessing the acquired heart sound data, the preprocessing including first performing Butterworth bandpass filtering and then normalizing the heart sound data; constructing and training a dual-pooling lightweight adaptive Mamba model; including: performing agent-guided pruning quantization on the trained model; and outputting heart sound classification results through the dual-pooling lightweight adaptive Mamba model and agent-guided pruning quantization.
[0008] As a preferred embodiment of the agent-guided pruning and quantization lightweight heart sound classification method described in this invention, the construction and training of the dual-pooling lightweight adaptive Mamba model specifically includes: taking preprocessed heart sound data as input, performing convolution and batch normalization operations sequentially to obtain preliminary features; inputting the preliminary features into the dual-pooling lightweight adaptive Mamba module for feature extraction, wherein the dual-pooling lightweight adaptive Mamba module performs max pooling and average pooling in parallel, and inputting the data after the two pooling processes into the lightweight adaptive Mamba module respectively; concatenating the outputs of the two lightweight adaptive Mamba modules to form the final feature representation for classification.
[0009] As a preferred embodiment of the agent-guided pruning and quantization lightweight heart sound classification method of the present invention, the construction of the deep learning model includes the following steps: the processing of the lightweight adaptive Mamba module includes: performing layer normalization on the input data; performing feature transformation on the normalized data sequentially through a linear layer and a depthwise separable convolutional layer; activating the output features of the depthwise separable convolutional layer using an adaptive wavelet activation function, the output expression of which is: ; in, The output features of depth-separable convolutional layers The adaptive coefficients are used; the activated features are input into the state space model to extract and model the temporal global features; the output of the state space model is processed through a linear layer and used as the output of the lightweight adaptive Mamba module.
[0010] As a preferred embodiment of the agent-guided pruning quantization lightweight heart sound classification method described in this invention, the calculation process of the state space model is implemented through the following set of formulas: ; ; ; ; ; ; in, Represents linear layer operations. and For activation function, This is the hidden state from the previous moment. These are the weighting coefficients. This indicates element-wise multiplication. This represents the summation operation, which is... , , To extract features through a linear layer Intermediate variables obtained from the mapping This is an intermediate state. The current hidden state. This is the output of the state-space model.
[0011] As a preferred embodiment of the agent-guided pruning and quantization lightweight heart sound classification method described in this invention, the agent-guided pruning and quantization process specifically includes: the agent observing the current state of the model and planning the pruning ratio and weight quantization bits for each convolutional layer in the model; and quantizing and compressing the model weights according to the quantization bits, the quantization process being performed according to the following formula: ; ; ; ; in, For the original weights, For the number of quantization bits, This is the weight matrix obtained by quantization followed by dequantization. This represents the maximum absolute value of the original weights of this layer. This represents the maximum integer quantization value for the corresponding number of quantization bits. This is the quantized integer weight matrix. The function is a rounding function; based on the pruning ratio, the quantized model weights are pruned, and unimportant weight connections are removed.
[0012] As a preferred embodiment of the agent-guided pruning and quantization lightweight heart sound classification method described in this invention, the agent is trained and makes decisions based on a reinforcement learning framework, and its reward function is defined as: , in, The classification accuracy of the lightweight model after pruning and quantization on the validation set. The size of the lightweight model, This is a scaling factor used to balance the dimensions and impact of accuracy on model size.
[0013] As a preferred embodiment of the agent-guided pruning quantization lightweight heart sound classification method described in this invention, the normalization process in the heart sound data preprocessing adopts the following formula: ; in, Represents the raw heart sound data points. This represents the total number of data points for a single heart sound signal. This represents the preprocessed heart sound data.
[0014] Secondly, to further address the aforementioned technical problems, the present invention provides an agent-guided pruning and quantization lightweight heart sound classification system, comprising: a data acquisition module for acquiring heart sound data from patients with different cardiovascular diseases and labeling them into categories to form a heart sound dataset; a data processing module for preprocessing the acquired heart sound data, including Butterworth bandpass filtering followed by normalization; a model building module for constructing and training a dual-pooling lightweight adaptive Mamba model; a model processing module for performing agent-guided pruning and quantization on the trained model; and a result output module for outputting heart sound classification results through the dual-pooling lightweight adaptive Mamba model and agent-guided pruning and quantization.
[0015] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the agent-guided pruning quantization lightweight heart sound classification method as described in the first aspect of the present invention.
[0016] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the agent-guided pruning quantization lightweight heart sound classification method as described in the first aspect of the present invention.
[0017] The beneficial effects of this invention are as follows: This invention combines a dual-pooling lightweight adaptive Mamba model to deeply extract features of different types of heart sounds. By introducing depthwise separable convolution, designing a state-space model and an adaptive wavelet activation function, it achieves accurate feature extraction of heart sounds, ensuring that it maintains efficient and stable heart sound classification capabilities even in changing environments. The agent-guided pruning quantization process further compresses the model size, reduces the number of model parameters, improves the model's running efficiency, and the model occupies less storage space in deployment devices, making the porting process more convenient. Attached Figure Description
[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. 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. Wherein: Figure 1 This is a flowchart of the heart sound classification method in an embodiment of the present invention.
[0019] Figure 2 This is a graph showing the test results of the indicators in an embodiment of the present invention.
[0020] Figure 3 This is a graph showing the test results of the quantified pruning metrics in an embodiment of the present invention.
[0021] Figure 4 This is a comparison diagram of different models in the embodiments of the present invention.
[0022] Figure 5 This is a confusion matrix diagram in an embodiment of the present invention. Detailed Implementation
[0023] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0024] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0025] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0026] Example 1 Reference Figure 1 This is the first embodiment of the present invention, which provides an agent-guided pruning and quantization lightweight heart sound classification method, including the following steps: S1: Obtain heart sound data from patients with different cardiovascular diseases and label them into categories to form a heart sound dataset.
[0027] Preferably, the heart sound data is the collected heart sound timing signal.
[0028] S2: The acquired heart sound data is preprocessed, including Butterworth bandpass filtering and normalization.
[0029] Preferably, the normalization process in the preprocessing of heart sound data adopts the following formula: ; in, Represents the raw heart sound data points. This represents the total number of data points for a single heart sound signal. This represents the preprocessed heart sound data.
[0030] Specifically, Butterworth bandpass filtering can filter out noise outside the frequency range of heart sounds, and normalization can adjust the amplitude of heart sound data to between -1 and 1.
[0031] S3: Construct and train a dual-pooling lightweight adaptive Mamba model.
[0032] Preferably, the construction and training of the dual-pooling lightweight adaptive Mamba model specifically includes: The preprocessed heart sound data is used as input, and convolution and batch normalization operations are performed sequentially to obtain preliminary features. The initial features are input into the dual-pooling lightweight adaptive Mamba module for feature extraction. The dual-pooling lightweight adaptive Mamba module performs max pooling and average pooling in parallel, and the data after the two pooling processes are input into the lightweight adaptive Mamba module respectively. The outputs of the two lightweight adaptive Mamba modules are concatenated to form the final feature representation, which is used for classification.
[0033] Furthermore, the processing steps of the lightweight adaptive Mamba module include: Perform layer normalization on the input data; The normalized data is then passed sequentially through a linear layer and a depthwise separable convolutional layer for feature transformation. The output features of the depthwise separable convolutional layer are activated using an adaptive wavelet activation function. The output expression of the adaptive wavelet activation function is as follows: ; in, The output features of depth-separable convolutional layers These are adaptive coefficients; The activated features are input into the state space model to extract and model temporal global features; The output of the state-space model is processed through a linear layer and then used as the output of the lightweight adaptive Mamba module.
[0034] Specifically, the calculation process of the state-space model is achieved through the following set of formulas: ; ; ; ; ; ; in, Represents linear layer operations. and For activation function, This is the hidden state from the previous moment. These are the weighting coefficients. This indicates element-wise multiplication. This represents the summation operation, which is... , , To extract features through a linear layer Intermediate variables obtained from the mapping This is an intermediate state. The current hidden state. This is the output of the state-space model.
[0035] S4: Perform agent-guided pruning and quantization on the trained model.
[0036] Preferably, agent-guided pruning quantization processing specifically includes: The agent observes the current state of the model and plans the pruning ratio and weight quantization bits for each convolutional layer in the model. The model weights are quantized and compressed based on the number of quantization bits. The quantization process is performed according to the following formula: ; ; ; ; in, For the original weights, For the number of quantization bits, This is the weight matrix obtained by quantization followed by dequantization. This represents the maximum absolute value of the original weights of this layer. This represents the maximum integer quantization value for the corresponding number of quantization bits. This is the quantized integer weight matrix. It is a rounding function; Based on the pruning ratio, the quantized model weights are pruned, and unimportant weight connections are removed.
[0037] Specifically, the agent is trained and makes decisions based on a reinforcement learning framework, and its reward function is defined as: , in, The classification accuracy of the lightweight model after pruning and quantization on the validation set. To reduce the size of the lightweight model, This is a scaling factor used to balance the dimensions and impact of accuracy on model size.
[0038] S5: Outputs heart sound classification results through dual-pooling lightweight adaptive Mamba model and agent-guided pruning quantization.
[0039] Specifically, the training of the dual-pooling lightweight adaptive Mamba model gradually converges, and the error rate between the output predicted label and the true label gradually decreases. By comparing the predicted label and the true label, the accuracy of the heart sound classification method proposed in this application is verified. Four evaluation metrics are given, which comprehensively evaluate the classification accuracy through accuracy, recall, precision, and F1 score. ; ; ; ; Among them, TP represents true positive, TN represents true negative, FP represents false positive, and FN represents false negative.
[0040] In summary, this invention combines a dual-pooling lightweight adaptive Mamba model to deeply extract features of different types of heart sounds. By introducing depthwise separable convolution, designing a state-space model and an adaptive wavelet activation function, it achieves accurate feature extraction of heart sounds, ensuring efficient and stable heart sound classification capabilities even in changing environments. Furthermore, the agent-guided pruning quantization process further compresses the model size, reduces the number of parameters, and improves the model's operating efficiency. The model occupies less storage space on deployment devices and is more convenient to port.
[0041] Example 2, an embodiment of the present invention, provides an agent-guided pruning and quantization lightweight heart sound classification system, comprising: a data acquisition module for acquiring heart sound data from patients with different cardiovascular diseases and labeling them into categories to form a heart sound dataset; a data processing module for preprocessing the acquired heart sound data, including Butterworth bandpass filtering and normalization; a model building module for constructing and training a dual-pooling lightweight adaptive Mamba model; a model processing module for performing agent-guided pruning and quantization on the trained model; and a result output module for outputting heart sound classification results through the dual-pooling lightweight adaptive Mamba model and agent-guided pruning and quantization.
[0042] Example 3 is an embodiment of the present invention, and its specific content is as follows: Heart sound data were collected using an electronic stethoscope. Cardiovascular diseases were classified into five categories, including one normal category (N) and four abnormal categories: aortic stenosis (AS), mitral regurgitation (MR), mitral stenosis (MS), and mitral valve prolapse (MVP). Each heart sound signal lasted about 3 seconds and was sampled at a frequency of 8 kHz. The classification performance of the dual-pooling lightweight adaptive Mamba model is as follows: Figure 2 As shown, the accuracy was 99.22%, the recall was 99.31%, the precision was 98.82%, and the F1 score was 99.04%. The classification performance of the agent-guided pruning and quantization model is as follows: Figure 3As shown, the accuracy was 94.27%, recall was 94.46%, precision was 94.47%, and F1 score was 94.27%. Compared to the model before quantization, the model after quantization had 17M Flops and 19K Parameters, which were reduced by 211M and 35K respectively. The model size ratio before and after compression was 1.81, as detailed in Table 1.
[0043] The accuracy of the heart sound classification method proposed in this application was compared with the results of four other researchers, such as... Figure 4 As shown in the figure, the heart sound classification method proposed in this application has the highest accuracy. Figure 5 It also shows the confusion matrix for classifying heart sounds for the above one normal class and four abnormal rates.
[0044] This application addresses the challenge of extracting disease-related heart sound features using depthwise separable convolutional depthwise methods in scenarios with limited heart sound data. A state-space model is designed to extract global sequence features. A lightweight adaptive Mamba model with dual pooling is applied to medical signal processing and artificial intelligence, enabling deep feature extraction of different heart sound categories. Furthermore, a designed adaptive wavelet activation function achieves accurate feature extraction of heart sounds, demonstrating excellent classification performance and ensuring efficient and stable heart sound classification even in variable environments. Agent-guided pruning quantization further compresses the model size, reduces the number of parameters, improves operational efficiency, and reduces storage space required for deployment, making porting easier.
[0045] Example 4 is an embodiment of the present invention, which differs from the previous embodiment in that: If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0046] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0047] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.
[0048] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.
[0049] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A lightweight, agent-guided, pruning-quantized heart sound classification method, characterized in that: include: Heart sound data from patients with different cardiovascular diseases were obtained and categorized to form a heart sound dataset. The heart sound data is the collected heart sound time sequence signal; The acquired heart sound data is preprocessed, including Butterworth bandpass filtering and normalization. Construct and train a lightweight adaptive Mamba model using dual pooling; including: The trained model is subjected to agent-guided pruning and quantization. Heart sound classification results are output through a dual-pooling lightweight adaptive Mamba model and agent-guided pruning quantization.
2. The agent-guided pruning, quantization, and lightweight heart sound classification method as described in claim 1, characterized in that: The construction and training of the dual-pooling lightweight adaptive Mamba model specifically includes: The preprocessed heart sound data is used as input, and convolution and batch normalization operations are performed sequentially to obtain preliminary features. The preliminary features are input into the dual-pooling lightweight adaptive Mamba module for feature extraction. The dual-pooling lightweight adaptive Mamba module performs max pooling and average pooling in parallel, and the data after the two pooling processes are input into the lightweight adaptive Mamba module respectively. The outputs of the two lightweight adaptive Mamba modules are concatenated to form the final feature representation, which is used for classification.
3. The agent-guided pruning, quantization, and lightweight heart sound classification method as described in claim 2, characterized in that: The processing steps of the lightweight adaptive Mamba module include: Perform layer normalization on the input data; The normalized data is then passed sequentially through a linear layer and a depthwise separable convolutional layer for feature transformation. The output features of the depthwise separable convolutional layer are activated using an adaptive wavelet activation function, the output expression of which is: ； in, The output features of depth-separable convolutional layers These are adaptive coefficients; The activated features are input into the state space model to extract and model temporal global features; The output of the state-space model is processed through a linear layer and then used as the output of the lightweight adaptive Mamba module.
4. The agent-guided pruning, quantization, and lightweight heart sound classification method as described in claim 3, characterized in that: The calculation process of the state-space model is achieved through the following set of formulas: ； ； ； ； ； ； in, Represents linear layer operations. and For activation function, This is the hidden state from the previous moment. These are the weighting coefficients. This indicates element-wise multiplication. This represents the summation operation, which is... , , To extract features through a linear layer Intermediate variables obtained from the mapping This is an intermediate state. The current hidden state. This is the output of the state-space model.
5. The agent-guided pruning, quantization, and lightweight heart sound classification method as described in claim 1, characterized in that: The agent-guided pruning quantization process specifically includes: The agent observes the current state of the model and plans the pruning ratio and weight quantization bits for each convolutional layer in the model. The model weights are quantized and compressed according to the quantization bit depth. The quantization process is performed according to the following formula: ； ； ； ； in, For the original weights, For the number of quantization bits, This is the weight matrix obtained by quantization followed by dequantization. This represents the maximum absolute value of the original weights of this layer. This represents the maximum integer quantization value for the corresponding number of quantization bits. This is the quantized integer weight matrix. It is a rounding function; Based on the pruning ratio, the quantized model weights are pruned, and unimportant weight connections are removed.
6. The agent-guided pruning, quantization, and lightweight heart sound classification method as described in claim 5, characterized in that: The agent is trained and makes decisions based on a reinforcement learning framework, and its reward function is defined as follows: ， in, The classification accuracy of the lightweight model after pruning and quantization on the validation set. The size of the lightweight model, This is a scaling factor used to balance the dimensions and impact of accuracy on model size.
7. The agent-guided pruning, quantization, and lightweight heart sound classification method as described in claim 1, characterized in that: The normalization process in the preprocessing of the heart sound data adopts the following formula: ； in, Represents the raw heart sound data points. This represents the total number of data points for a single heart sound signal. This represents the preprocessed heart sound data.
8. An agent-guided, pruning-quantized, lightweight heart sound classification system, based on the agent-guided, pruning-quantized, lightweight heart sound classification method according to any one of claims 1 to 7, characterized in that: include, The data acquisition module is used to acquire heart sound data from patients with different cardiovascular diseases and label them into categories to form a heart sound dataset; The data processing module is used to preprocess the acquired heart sound data. The preprocessing includes Butterworth bandpass filtering and normalization of the heart sound data. The model building module is used to build and train a dual-pooling lightweight adaptive Mamba model. The model processing module is used to perform agent-guided pruning and quantization processing on the trained model. The results output module is used to output heart sound classification results through dual-pooling lightweight adaptive Mamba model and agent-guided pruning quantization processing.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the agent-guided pruning quantization lightweight heart sound classification method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the agent-guided pruning quantization lightweight heart sound classification method according to any one of claims 1 to 7.