A heart sound classification method and system based on hyper-dimensional computing

Abstract

The application belongs to the technical field of heart sound signals, and discloses a heart sound classification method and system based on hyper-dimensional calculation, which comprises the following steps: extracting heart sound signal features through MFCCs and DWT; encoding the extracted features into hyper-vectors; generating a plurality of sub-class hyper-vectors through hyper-vector clustering; and calculating the similarity between a test sample hyper-vector and the sub-class hyper-vectors to perform classification. The application can efficiently perform a heart sound classification task on a resource-limited device. The method generates sub-class hyper-vectors through hyper-dimensional encoding and clustering, and performs classification by using hyper-vector similarity calculation. Compared with traditional deep learning methods, the application has the advantages of high calculation efficiency, low energy consumption, low data requirement, strong hardware fault tolerance, and the like, and is suitable for intelligent medical devices and embedded diagnosis systems.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of heart sound signal technology, and particularly relates to a heart sound classification method and system based on hyperdimensional computation. Background Technology

[0002] Heart sound analysis, as a non-invasive diagnostic method, has important applications in clinical medicine. In recent years, deep learning methods have been widely used for heart sound classification and have made significant progress in accuracy. However, deep learning models are typically computationally expensive, requiring high-performance computing resources, which limits their application on resource-constrained embedded devices. Furthermore, deep learning methods are highly dependent on data quality and the number of training samples, making them difficult to handle noisy environments and limited training data. Summary of the Invention

[0003] To address the problems existing in current technologies, this invention proposes a heart sound classification method based on hyperdimensional computation. Hyperdimensional computation is a computational paradigm inspired by the brain's information processing mechanisms. It can represent and process data using high-dimensional hypervectors, offering advantages such as high computational speed, low energy consumption, and strong robustness to noise. The goal of this invention is to improve computational efficiency, reduce energy consumption, and enhance tolerance to hardware failures while maintaining classification accuracy, thereby achieving efficient heart sound classification in resource-constrained environments.

[0004] This invention is implemented as follows: a heart sound classification method based on hyperdimensional computation, the method comprising:

[0005] S1: Signal preprocessing: Extracting Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) features from the heart sound signal;

[0006] S2: Hyperdimensional encoding: Encode the extracted features into a sample hypervector Sample-HV, so that similar heart sound samples correspond to similar hypervectors;

[0007] S3: Classification: Cluster the sample hypervectors to divide each heart sound category into multiple subcategories, and generate a subclass hypervector Subclass-HV for each subcategory;

[0008] S4: Classification decision: Determine the category to which a test sample belongs by calculating the similarity between its hypervector and the hypervectors of each subclass;

[0009] S5: Subclass Hypervector Retraining: Adjust the subclass hypervectors based on misclassified samples to further improve classification accuracy.

[0010] Furthermore, S1 specifically includes:

[0011] Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) are commonly used mathematical tools in signal processing and classification, and they have shown good performance in the field of heart sound analysis. MFCCs can capture the spectral characteristics of heart sound signals, while DWT can decompose the time-frequency characteristics of the signal. In the experiment, this invention extracted N features from each heart sound sample, two-thirds of which were MFCCs and one-third were DWT features. This combination can effectively extract key frequency information of heart sounds to accurately identify heart diseases.

[0012] Furthermore, S2 specifically includes:

[0013] In this step, each heart sound sample's N features are assigned a randomly generated feature hypervector, denoted as F_nin{-1,1}^D. These hypervectors typically have a dimension D of 10,000 and are approximately orthogonal to each other, meaning the positional difference between hypervectors is close to 50%.

[0014] Since the eigenvalues ​​are continuous, this invention linearly quantizes each eigenvalue into M discrete levels and assigns a level-HV to each level. Specifically, this invention first determines the minimum and maximum values ​​of each feature and divides 96% of the data range into M discrete intervals, with the lowest 2% and the highest 2% assigned to the lowest and highest levels, respectively. Then, a hypervector is assigned to each level, such that the hypervectors of adjacent levels have high similarity, while the hypervectors of the furthest level are completely orthogonal.

[0015] The feature hypervector (Feature-HV) and its corresponding level hypervector (Level-HV) are combined through a binding operation to generate the final feature value hypervector. The binding operation is calculated using a bitwise XOR operation, as shown in the following formula:

[0016] H_n=F_nastL_{i,n}

[0017] Where H_n is the final hypervector corresponding to feature n, and L_{i,n} is the level hypervector corresponding to the feature;

[0018] The hypervectors of all features are combined through a bundling operation to form the complete sample hypervector, Sample-HV, which is calculated as follows:

[0019] S_i = [H_1 + H_2 + ... + H_N]

[0020] The brackets [.] will assign a value of 1 to dimensions greater than 0 and a value of -1 to dimensions less than 0, thus binarizing the vector.

[0021] Furthermore, S3 specifically includes:

[0022] The hyperspace clustering method is used to divide the hypervectors of samples within each category into K subcategories; the specific method is as follows:

[0023] S31: Random initialization: Randomly assign the sample hypervector of each class to K subclasses;

[0024] S32: Calculate the sub-category hypervector: Perform a bundle operation on the hypervectors of all samples within the sub-category to generate the sub-category hypervector.

[0025] C_j^k=[S_1^k+S_2^k+…+S_n^k]

[0026] S33: Reassign samples: Calculate the Hamming distance between each sample hypervector and all subclass hypervectors, and assign the sample to the most similar subclass;

[0027] S34: Repeat steps S32-S33 until convergence;

[0028] Ultimately, each category is divided into K subcategories, and K subcategory hypervectors are stored as classification references.

[0029] Furthermore, S4 specifically includes:

[0030] For the test sample, the sample hypervector (Sample-HV) is first calculated using the encoder. Then, the Hamming distance between this hypervector and the hypervectors of all subclasses is calculated. The classification result depends on the most similar subclass. [

[0032] text{PredictedClass}=argmin_{substack{j,k}}d_H(S_{text{test}},C_j^k)

[0033] ].

[0034] Furthermore, S5 specifically includes:

[0035] (1) Identify the misclassified samples and record their true class and the supervector of the misclassified subclass.

[0036] (2) Adjust the sub-category hypervector:

[0037] If a subclass misclassifies a sample, then subtract the sample's hypervector from the subclass's hypervector:

[0038] C_{text{wrong}}=C_{text{wrong}}-S_{text{test}}

[0039] If a sample is misclassified, its hypervector is added back to the correct subclass:

[0040] C_{text{correct}}=C_{text{correct}}+S_{text{test}}

[0041] (3) Repeat the above process until the classification accuracy of the training set reaches a stable state.

[0042] Another objective of this invention is to provide a hyperdimensional computation-based heart sound classification system based on the aforementioned hyperdimensional computation-based heart sound classification method. This system specifically includes:

[0043] The signal preprocessing module extracts the Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) features of the heart sound signal.

[0044] The hyperdimensional encoding module, connected to the signal preprocessing module, encodes the extracted features into sample hypervectors Sample-HV, so that similar heart sound samples correspond to similar hypervectors.

[0045] The category segmentation module, connected to the hyperdimensional encoding module, clusters the sample hypervectors, dividing each heart sound category into multiple subcategories, and generating a subclass hypervector (Subclass-HV) for each subcategory;

[0046] The classification decision module, connected to the category partitioning module, determines the category to which a test sample belongs by calculating the similarity between the hypervector of the test sample and the hypervector of each subcategory.

[0047] Subclass hypervectors are retrained and connected to the classification decision module. The subclass hypervectors are adjusted based on misclassified samples to further improve classification accuracy.

[0048] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor, the memory storing a computer program, which, when executed by the processor, causes the processor to perform the steps of the heart sound classification method based on hyperdimensional computation.

[0049] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the heart sound classification method based on hyperdimensional computation.

[0050] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0051] First, the technical solution provided by this invention exhibits superior computational efficiency. Compared to traditional 1D-CNN (single-dimensional convolutional neural network), this method improves training speed by 11.76 times, significantly shortening the time required for model training and making it more advantageous in high-throughput tasks. Furthermore, the 12% improvement in classification speed means that in real-time inference scenarios, this model can complete data classification more quickly, improving system responsiveness and making it suitable for low-latency, high-concurrency application environments.

[0052] In terms of computational resource consumption, this solution significantly reduces computational costs. Traditional CNN models rely on a large number of 64-bit floating-point multiplication operations, requiring 5.23 million calculations, while this solution uses a lighter 1-bit XOR operation, requiring only 2.8 million calculations, resulting in a significant reduction in computational complexity. Since the XOR operation only involves simple binary operations, the calculation process does not rely on high-precision floating-point arithmetic units, greatly reducing hardware resource consumption and making it suitable for embedded systems or low-power devices with limited computing power.

[0053] Optimized data requirements are a major highlight of this solution. Even with a 40% reduction in training data, the model's classification accuracy remains at 96.34%, indicating that this method has low data dependence and good generalization ability. In practical applications, this means that even with high data acquisition costs or insufficient training data, this method can still maintain high-performance classification results, making it suitable for data-constrained scenarios such as medical diagnosis, industrial inspection, and artificial intelligence applications in low-resource environments.

[0054] This solution also possesses excellent hardware fault tolerance, maintaining a high classification accuracy even when 40% of the model parameters are randomly corrupted. This demonstrates the method's strong robustness to hardware defects, computational errors, and storage corruption, enabling stable operation in environments with significant computing power fluctuations or low hardware reliability. For example, on edge computing devices, low-power chips, or space-constrained smart devices, this solution can still ensure high-precision calculation results, thereby improving the system's reliability and fault tolerance.

[0055] Second, the technical solution of this invention fills a technological gap in the industry both domestically and internationally: it is the first time that hyperdimensional computing technology has been deployed in the field of heart sound recognition.

[0056] The technical solution of this invention solves a technical problem that people have long desired to solve but have never been able to achieve: it builds an energy-saving and robust heart sound classifier that can rely on simple and unreliable storage, does not require special hardware acceleration (such as GPU), and meets the increasing needs of embedded development. Attached Figure Description

[0057] Figure 1This is a flowchart of a heart sound classification method based on hyperdimensional computation provided in an embodiment of the present invention;

[0058] Figure 2 This is a flowchart of the category classification method provided in the embodiments of the present invention;

[0059] Figure 3 This is a block diagram of a heart sound classification system based on hyperdimensional computation provided in an embodiment of the present invention;

[0060] Figure 4 The following are the effects of the parameters provided in the embodiments of the present invention on model performance: (a) the effect of the hypervector dimension (D) when K=10, T=20, M=200; (b) the effect of the number of subclasses (K) of each class when D=10000, T=20, M=200; (c) the effect of the number of subclass clustering iterations (T) when D=10000, K=8, M=200; and (d) the effect of the number of levels (M) in the Level-HV dictionary when D=10000, K=8, T=10.

[0061] Figure 5 This is an example of how different numbers of features affect the results, as provided in this embodiment of the invention.

[0062] Figure 6 It is the average accuracy of the model provided in the embodiments of the present invention after retraining at different times.

[0063] Figure 7 This is the model confusion matrix on the PhysioNet dataset before and after random retraining, as provided in this embodiment of the invention.

[0064] Figure 8 This refers to the accuracy of the model on the Yaseen dataset when flipping different percentages of Subclass-HV elements, as provided in this embodiment of the invention.

[0065] Figure 9 This refers to the accuracy of the model when trained on limited data, as provided in the embodiments of the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0067] like Figure 1 As shown, a heart sound classification method based on hyperdimensional computation is proposed, which includes:

[0068] S1: Signal preprocessing: Extracting Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) features from the heart sound signal;

[0069] S2: Hyperdimensional encoding: Encode the extracted features into a sample hypervector Sample-HV, so that similar heart sound samples correspond to similar hypervectors;

[0070] S3: Classification: Cluster the sample hypervectors to divide each heart sound category into multiple subcategories, and generate a subclass hypervector Subclass-HV for each subcategory;

[0071] S4: Classification decision: Determine the category to which a test sample belongs by calculating the similarity between its hypervector and the hypervectors of each subclass;

[0072] S5: Subclass Hypervector Retraining: Adjust the subclass hypervectors based on misclassified samples to further improve classification accuracy.

[0073] This invention proposes a hyperdimensional computation-based method for classifying heart sounds, achieving efficient classification of heart sound signals by utilizing the characteristics of high-dimensional vector spaces. First, in the signal preprocessing stage (S1), feature extraction is performed on the original heart sound signals, primarily using Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT). MFCCs effectively capture the short-time spectral features of heart sounds, while DWT provides multi-scale frequency domain information. The combination of these two features provides rich descriptive features for subsequent hyperdimensional encoding.

[0074] In the hyperdimensional encoding stage (S2), the extracted features are encoded and transformed into high-dimensional sparse vectors, namely sample hypervectors (Sample-HV). The core of hyperdimensional encoding lies in mapping the original features into a high-dimensional space through random projection or binarization operations, so that heart sound samples with similar features are encoded as similar hypervectors. Due to the robustness and high-dimensional separability of hyperdimensional space, even in the presence of noise or incomplete features, the similarity measure between samples still has good discriminative ability.

[0075] Category segmentation (S3) employs a clustering method to group sample hypervectors into multiple subcategories. Specifically, each heart sound category is further divided into several subcategories, and a representative hypervector, or subclass hypervector (Subclass-HV), is generated for each subcategory. This hierarchical structure allows for more precise capture of the internal differences and similarities between different categories of heart sound signals. Furthermore, this segmentation strategy helps improve the flexibility and accuracy of classification, especially when there is overlap or high similarity between categories.

[0076] In the classification decision stage (S4), the similarity between the hypervector of the test sample and the hypervectors of each subclass is calculated using metrics such as cosine similarity or Hamming distance, and classification is performed based on the similarity scores. During this process, subclass hypervector retraining is performed through analysis and feedback on misclassified samples (S5). Specifically, retraining corrects and optimizes misclassified samples, thereby adjusting the corresponding subclass hypervectors to make the hypervector representation more accurate. The overall method improves the accuracy and robustness of heart sound classification through continuous iterative optimization.

[0077] S1 specifically includes:

[0078] Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) are commonly used mathematical tools in signal processing and classification, and they have shown good performance in the field of heart sound analysis. MFCCs can capture the spectral characteristics of heart sound signals, while DWT can decompose the time-frequency characteristics of the signal. In the experiment, this invention extracted N features from each heart sound sample, two-thirds of which were MFCCs and one-third were DWT features. This combination can effectively extract key frequency information of heart sounds to accurately identify heart diseases.

[0079] S2 specifically includes:

[0080] In this step, each heart sound sample's N features are assigned a randomly generated feature hypervector, denoted as F_nin{-1,1}^D. These hypervectors typically have a dimension D of 10,000 and are approximately orthogonal to each other, meaning the positional difference between hypervectors is close to 50%.

[0081] Since the eigenvalues ​​are continuous, this invention linearly quantizes each eigenvalue into M discrete levels and assigns a level-HV to each level. Specifically, this invention first determines the minimum and maximum values ​​of each feature and divides 96% of the data range into M discrete intervals, with the lowest 2% and the highest 2% assigned to the lowest and highest levels, respectively. Then, a hypervector is assigned to each level, such that the hypervectors of adjacent levels have high similarity, while the hypervectors of the furthest level are completely orthogonal.

[0082] The feature hypervector (Feature-HV) and its corresponding level hypervector (Level-HV) are combined through a binding operation to generate the final feature value hypervector. The binding operation is calculated using a bitwise XOR operation, as shown in the following formula:

[0083] H_n=F_nastL_{i,n}

[0084] Where H_n is the final hypervector corresponding to feature n, and L_{i,n} is the level hypervector corresponding to the feature;

[0085] The hypervectors of all features are bundled together to form the complete sample hypervector, Sample-HV, which is calculated as follows:

[0086] S_i = [H_1 + H_2 + ... + H_N]

[0087] The brackets [.] will assign a value of 1 to dimensions greater than 0 and a value of -1 to dimensions less than 0, thus binarizing the vector.

[0088] like Figure 3 As shown, S3 specifically includes:

[0089] The hyperspace clustering method is used to divide the hypervectors of samples within each category into K subcategories; the specific method is as follows:

[0090] S31: Random initialization: Randomly assign the sample hypervector of each class to K subclasses;

[0091] S32: Calculate the sub-category hypervector: Perform a bundle operation on the hypervectors of all samples within the sub-category to generate the sub-category hypervector.

[0092] C_j^k=[S_1^k+S_2^k+…+S_n^k]

[0093] S33: Reassign samples: Calculate the Hamming distance between each sample hypervector and all subclass hypervectors, and assign the sample to the most similar subclass;

[0094] S34: Repeat steps S32-S33 until convergence;

[0095] Ultimately, each category is divided into K subcategories, and K subcategory hypervectors are stored as classification references.

[0096] S4 specifically includes:

[0097] For the test sample, the sample hypervector (Sample-HV) is first calculated using the encoder. Then, the Hamming distance between this hypervector and the hypervectors of all subclasses is calculated. The classification result depends on the most similar subclass. [

[0099] text{PredictedClass}=argmin_{substack{j,k}}d_H(S_{text{test}},C_j^k)

[0100] ].

[0101] S5 specifically includes:

[0102] (1) Identify the misclassified samples and record their true class and the supervector of the misclassified subclass.

[0103] (2) Adjust the sub-category hypervector:

[0104] If a subclass misclassifies a sample, then subtract the sample's hypervector from the subclass's hypervector:

[0105] C_{text{wrong}}=C_{text{wrong}}-S_{text{test}}

[0106] If a sample is misclassified, its hypervector is added back to the correct subclass:

[0107] C_{text{correct}}=C_{text{correct}}+S_{text{test}}

[0108] (3) Repeat the above process until the classification accuracy of the training set reaches a stable state.

[0109] Example 1: Heart Sound Signal Preprocessing and Feature Extraction

[0110] In this embodiment, MFCCs and DWT features are first extracted from the heart sound signals, with 240 features extracted from each sample for classification. This step is implemented in the MATLAB environment and is applicable to different signal acquisition devices.

[0111] Example 2: Hyperdimensional Encoding

[0112] In this embodiment, hyperdimensional computation uses binary hypervectors with dimension D = 10000. Each feature is assigned an independent hypervector (Feature-HV), and the feature value hypervectors (Level-HV) are combined through a binding operation to generate the sample hypervector (Sample-HV).

[0113] Example 3: Subcategorization and Classification

[0114] This embodiment uses a clustering method to divide each category into K=8 subcategories and generates subclass hypervectors (Subclass-HV). During classification, the Hamming distance between the test sample hypervector and the hypervectors of each subcategory is calculated, and the most similar subcategory is selected for classification.

[0115] Example 4: Subclass Hypervector Retraining

[0116] In this embodiment, misclassified samples are analyzed, and the corresponding sub-category hypervectors are adjusted to optimize the classification boundary and improve classification accuracy.

[0117] Application Examples:

[0118] To comprehensively evaluate the performance of the model in this invention, it was tested on two mature, publicly available heart sound datasets: one dataset contains high-quality heart sound signals, and the other contains noisy heart sound recordings. Details of these datasets are as follows:

[0119] Yaseen's dataset contains high-quality heart sound signals collected from professional stethoscopes. The dataset includes five categories of heart sounds: one from healthy individuals and four from patients diagnosed with specific heart diseases: mitral stenosis (MS), mitral regurgitation (MR), mitral valve prolapse (MVP), and aortic stenosis (AS). The dataset contains 1000 heart sound samples, distributed on an average of 200 samples per category.

[0120] The PhysioNet / CinC Challenge 2016 provides heart sound recordings captured in various noisy environments. Some recordings are so heavily corrupted by noise that even human doctors cannot interpret them. This invention uses this dataset to evaluate the robustness of the model to noisy, low-quality samples. The dataset contains 3153 heart sound samples for training and 301 validation samples, labeled as "normal" or "abnormal" heart sounds.

[0121] Experiment Implementation

[0122] This invention implements feature extraction in MATLAB, deriving 480 MFCC and 240 DWT features from each heart sound sample. From these 720 features, this invention selects 240 to train its model on the Yasen dataset, while all 720 features are used on the PhysioNet dataset.

[0123] The HDC classifier of this invention was implemented in Python using NumPy. The CNN baseline was implemented using PyTorch. All experiments were performed on an Apple M1 Pro processor (10 cores, 3.2 GHz, 16 GB of RAM).

[0124] For each method on the Yasen dataset, this invention randomly allocates 80% of the samples for training and 20% for testing. For the PhysioNet dataset, this invention trains using all training samples and tests using the validation dataset. Each experiment is repeated 10 times to evaluate the stability of the model.

[0125] Comparison with previous studies:

[0126] On the Yasen dataset, the model of this invention achieved a peak accuracy of 98.95% and an average accuracy of 96.60% without retraining. The table below compares the accuracy of the model of this invention with that of previous methods on the Yasen dataset. Including feature extraction, the model of this invention completes training in an average of only 7.85 seconds, while the traditional 1D-CNN method requires 92.29 seconds to achieve a comparable accuracy of 96.50%. Testing speed is also improved: the model of this invention, including feature extraction, classifies 100 samples in 0.91 seconds, while the CNN baseline method requires 1.02 seconds—an improvement of 12%, as shown in the table below.

[0127]

[0128] The results were compared with previous attempts on the Yasen dataset. Our model parameters were D=10000, K=8, T=10, M=200.

[0129] Table: Yasen_Precision

[0130] The model of this invention exhibits excellent robustness in noisy environments, achieving an outstanding accuracy of 99.67% on the noisy PhysioNet2016 dataset, while the previous winning model in the PhysioNet 2016 challenge only achieved an accuracy of 86.02%, as shown in the table below.

[0131]

[0132] The results were compared with previous attempts on the PhysioNet dataset. Our model parameters were D=10000, K=20, T=20, M=200, τ=10.

[0133] Table: PhysioNet_Precision

[0134] The model of this invention significantly outperforms traditional deep learning models in terms of energy consumption. For each classification operation, the model of this invention requires approximately 2,800,000 1-bit XOR operations, while Bagheel's CNN model requires 5,237,952 64-bit floating-point multiplications. Although specific energy consumption varies depending on the hardware platform, the fundamental difference in computational complexity—simple 1-bit binary operations versus multi-bit floating-point matrix multiplications—indicates that the model of this invention consumes far less energy than traditional methods.

[0135] Tolerance for hardware failure

[0136] Neural network architectures exhibit remarkable tolerance to component failures through redundant representations, a stark contrast to the critical importance of individual bits in traditional binary computer systems. Brain-inspired HDC methods, utilizing distributed information representations of high-dimensional vectors, demonstrate even greater resilience to hardware failures.

[0137] To evaluate this robustness, this invention performed a perturbation analysis by randomly flipping elements in all trained subclass high-dimensional vectors (Subclass-HVs) (converting 1 to -1 and vice versa) to simulate hardware storage instability. The figure (model accuracy on the Yasen dataset with different percentages of Subclass-HV elements flipped) shows that when the dimensionality is sufficiently high ($D = 10000$), even flipping 40% of the elements only reduces the average accuracy by 3.25%. All dimensions tend towards 20% accuracy with a 50% perturbation, consistent with HDC theory.

[0138] subsection {Performance with limited training data}

[0139] Data scarcity is a major challenge in medical diagnosis (especially heart sound classification), and privacy concerns often limit sample availability. As shown in Figure ~ref{fig:part_train}, the model of this invention maintains excellent performance even with reduced training data. Its robustness under data-constrained conditions, as well as its computational efficiency, make it extremely promising for clinical deployment.

[0140] I. Specific application areas or related products of this invention.

[0141] like Figure 3 As shown, this embodiment of the invention provides a hyperdimensional computation-based heart sound classification system based on the aforementioned hyperdimensional computation-based heart sound classification method. The system specifically includes:

[0142] The signal preprocessing module extracts the Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) features of the heart sound signal.

[0143] The hyperdimensional encoding module, connected to the signal preprocessing module, encodes the extracted features into sample hypervectors Sample-HV, so that similar heart sound samples correspond to similar hypervectors.

[0144] The category segmentation module, connected to the hyperdimensional encoding module, clusters the sample hypervectors, dividing each heart sound category into multiple subcategories, and generating a subclass hypervector (Subclass-HV) for each subcategory;

[0145] The classification decision module, connected to the category partitioning module, determines the category to which a test sample belongs by calculating the similarity between the hypervector of the test sample and the hypervector of each subcategory.

[0146] Subclass hypervectors are retrained and connected to the classification decision module. The subclass hypervectors are adjusted based on misclassified samples to further improve classification accuracy.

[0147] This invention provides a computer device, which includes a memory and a processor. The memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the heart sound classification method based on hyperdimensional computation.

[0148] This invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the heart sound classification method based on hyperdimensional computation.

[0149] II. Evidence related to the technical effects obtained by the embodiments of the present invention.

[0150] In the experiment, this invention tested the impact of the following parameters on model performance:

[0151] Hypervector dimension (D): The classification accuracy is optimal when D = 10,000.

[0152] Number of subcategories (K): K=8 can effectively represent the internal variability of different categories.

[0153] Clustering iteration number (T): When T = 10, the sub-class hypervectors converge stably.

[0154] Number of level hypervectors (M): When M = 200, continuous features can be accurately represented.

[0155] Figure 4 The effects of (a) the hypervector dimension (D) when K=10, T=20, M=200; (b) the effect of the number of subclasses (K) per class when D=10000, T=20, M=200; (c) the effect of the number of sub-clustering iterations (T) when D=10000, K=8, M=200; and (d) the effect of the number of levels (M) in the Level-HV dictionary when D=10000, K=8, T=10. All experiments were run on the Yasen dataset. Execution time does not include feature extraction. Test time represents classification time per 100 samples. No retraining was performed.

[0156] Figure 5 This describes the impact of using different numbers of features on the results.

[0157] Figure 6It is the average accuracy of the model after retraining at different times.

[0158] Figure 7 This is the model confusion matrix on the PhysioNet dataset before and after retraining in a random manner.

[0159] Figure 8 This represents the model's accuracy on the Yaseen dataset when flipping different percentages of the Subclass-HV elements.

[0160] Figure 9 It refers to the accuracy of the model when trained on limited data.

[0161] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0162] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A heart sound classification method based on hyperdimensional computation, characterized in that, The method includes: S1: Signal preprocessing: Extracting Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) features from the heart sound signal; S2: Hyperdimensional encoding: Encode the extracted features into a sample hypervector Sample-HV, so that similar heart sound samples correspond to similar hypervectors; S3: Classification: Cluster the sample hypervectors to divide each heart sound category into multiple subcategories, and generate a subclass hypervector Subclass-HV for each subcategory; S4: Classification decision: Determine the category to which a test sample belongs by calculating the similarity between its hypervector and the hypervectors of each subclass; S5: Subclass hypervector retraining: Adjust the subclass hypervectors based on misclassified samples to further improve classification accuracy; S2 specifically includes: In this step, each of the N features of a heart sound sample is assigned a randomly generated feature hypervector, denoted as F_nin{-1,1}^D; the dimension D of these hypervectors is 10,000, and they are approximately orthogonal to each other, that is, the positional difference between the hypervectors is close to 50%. Determine the minimum and maximum values ​​for each feature, and divide 96% of the data range into M discrete intervals, assigning the lowest 2% and the highest 2% to the lowest and highest levels, respectively; then, assign a hypervector to each level, such that the hypervectors of adjacent levels have high similarity, while the hypervectors of the furthest level are completely orthogonal. The feature hypervector (Feature-HV) and its corresponding level hypervector (Level-HV) are combined through a binding operation to generate the final feature value hypervector. The binding operation is calculated using a bitwise XOR operation, as shown in the following formula: H_n=F_nastL_{i,n} Where H_n is the final hypervector corresponding to feature n, and L_{i,n} is the level hypervector corresponding to the feature; The hypervectors of all features are combined through a bundling operation to form the complete sample hypervector, Sample-HV, which is calculated as follows: S_i = [H_1 + H_2 + … + H_N] The brackets [.] will assign a value of 1 to dimensions greater than 0 and a value of -1 to dimensions less than 0, thus binarizing the vector.

2. The heart sound classification method based on hyperdimensional computation according to claim 1, characterized in that, S1 specifically includes: N features were extracted from each heart sound sample, two-thirds of which were MFCCs and one-third were DWT features.

3. The heart sound classification method based on hyperdimensional computation as described in claim 1, characterized in that, S3 specifically includes: The hyperspace clustering method is used to divide the hypervectors of samples within each category into K subcategories; the specific method is as follows: S31: Random initialization: Randomly assign the sample hypervector of each class to K subclasses; S32: Calculate the sub-category hypervector: Perform a bundle operation on the hypervectors of all samples within the sub-category to generate the sub-category hypervector. C_j^k=[S_1^k+S_2^k+…+S_n^k] S33: Reassign samples: Calculate the Hamming distance between each sample hypervector and all subclass hypervectors, and assign the sample to the most similar subclass; S34: Repeat steps S32-S33 until convergence; Ultimately, each category is divided into K subcategories, and K subcategory hypervectors are stored as classification references.

4. The heart sound classification method based on hyperdimensional computation as described in claim 1, characterized in that, S4 specifically includes: For the test sample, the sample hypervector (Sample-HV) is first calculated using the encoder. Then, the Hamming distance between this hypervector and the hypervectors of all subclasses is calculated. The classification result depends on the most similar subclass. [ text{PredictedClass}=argmin_{substack{j,k}}d_H(S_{text{test}},C_j^k) ]。 5. The heart sound classification method based on hyperdimensional computation as described in claim 1, characterized in that, S5 specifically includes: (1) Identify the misclassified samples and record their true class and the hypervector of the misclassified subclass; (2) Adjust the sub-category hypervector: If a subclass misclassifies a sample, then subtract the sample's hypervector from the subclass's hypervector: C_{text{wrong}}=C_{text{wrong}}-S_{text{test}} If a sample is misclassified, its hypervector is added back to the correct subclass: C_{text{correct}}=C_{text{correct}}+S_{text{test}} (3) Repeat the above process until the classification accuracy of the training set reaches a stable state.

6. A heart sound classification system based on hyperdimensional computation, based on the heart sound classification method based on hyperdimensional computation as described in any one of claims 1-5, characterized in that, The system specifically includes: The signal preprocessing module extracts the Mel-frequency cepstral coefficients (MFCCs) and discrete wavelet transform (DWT) features of the heart sound signal. The hyperdimensional encoding module, connected to the signal preprocessing module, encodes the extracted features into sample hypervectors Sample-HV, so that similar heart sound samples correspond to similar hypervectors. The category segmentation module, connected to the hyperdimensional encoding module, clusters the sample hypervectors, dividing each heart sound category into multiple subcategories, and generating a subclass hypervector (Subclass-HV) for each subcategory; The classification decision module, connected to the category partitioning module, determines the category to which a test sample belongs by calculating the similarity between the hypervector of the test sample and the hypervector of each subcategory. Subclass hypervectors are retrained and connected to the classification decision module. The subclass hypervectors are adjusted based on misclassified samples to further improve classification accuracy.

7. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the heart sound classification method based on hyperdimensional computation as described in any one of claims 1-5.

8. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the heart sound classification method based on hyperdimensional computation as described in any one of claims 1-5.

Images