Method and system for constructing heart sound detection model

Abstract

The invention relates to the technical field of medical data processing, and discloses a heart sound detection model construction method and system, and the method comprises the steps: obtaining heart sound signal data of a target old group, and carrying out the preprocessing; respectively inputting the processed one-dimensional heart sound signal and a Mel time-frequency spectrogram obtained by converting the processed one-dimensional heart sound signal into a time domain feature analysis branch and a time frequency feature analysis branch in a double-branch neural network for feature extraction; and fusing the extracted two paths of high-dimensional features by using a bidirectional long-short-term memory network based on an attention mechanism, and training to obtain a heart sound detection model. The time domain and frequency domain features of the heart sound signals are fully mined through the double-branch structure, the focusing capacity of the model for key information is enhanced in combination with an attention mechanism, the heart sound classification accuracy and the model interpretability are improved, meanwhile, low calculation complexity is kept, deployment in embedded equipment is facilitated, and the method is suitable for popularization and application. Effective technical support is provided for early screening and remote auxiliary diagnosis of cardiopulmonary diseases.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method and system for constructing a heart sound detection model. Background Technology

[0002] With the increasing trend of global population aging, cardiopulmonary diseases have become one of the main factors affecting the health of the elderly. Statistics from the World Health Organization (WHO) show that cardiovascular and respiratory diseases are among the leading causes of death worldwide.

[0003] Currently, early diagnosis of cardiopulmonary diseases still largely relies on doctors' auscultation and interpretation of physiological signals. While this method plays an irreplaceable role in primary healthcare and family health settings, it also has limitations, such as the susceptibility of diagnostic results to doctor experience and the difficulty in achieving standardized recording and retrospective analysis. In recent years, some AI-based assisted diagnostic methods have been proposed, aiming to improve the consistency and efficiency of diagnosis through automated analysis. These methods typically use physiological signals collected by sensors as input, employing neural network models to achieve preliminary identification of abnormal states. However, existing AI-assisted diagnostic methods still face several common problems in practical application. For example, in complex environments, models are easily affected by various noise interferences, impacting recognition sensitivity; simultaneously, due to significant differences in physiological characteristics among individuals, the generalization ability of the models still needs improvement. Furthermore, most model structures are relatively simple, making it difficult to perform fine-grained localization and type discrimination of key components in the signal, resulting in insufficient clinical interpretability of their output results. On the other hand, high-precision models are usually accompanied by high computational complexity, limiting their deployment feasibility on resource-constrained embedded devices or mobile terminals.

[0004] Therefore, there is an urgent need in this field for an intelligent heart sound analysis solution that can maintain high accuracy while possessing good environmental robustness, strong generalization ability, high interpretability, and applicability to low-power hardware platforms, so as to promote the development of early diagnosis technology for cardiopulmonary diseases and improve the quality and coverage of primary and telemedicine services. Summary of the Invention

[0005] This invention provides a method and system for constructing a heart sound detection model to overcome the shortcomings of existing technologies.

[0006] The present invention provides a method for constructing a heart sound detection model, comprising: Acquire heart sound signal data of the target elderly population, including various heart sound signals; Based on the heart sound signal data of the target elderly group, a dual-branch neural network and attention fusion mechanism are used to train the model to learn the characteristic changes of different heart sound signals in the time domain and frequency domain, thus obtaining a heart sound detection model.

[0007] According to the method for constructing a heart sound detection model provided by the present invention, the various heart sound signals include any one of the following or any combination thereof: normal heart sounds, systolic murmurs, diastolic murmurs, and pericardial friction rubs.

[0008] According to a method for constructing a heart sound detection model provided by the present invention, the step of acquiring heart sound signal data of a target elderly population includes: Preprocessing is performed on the heart sound signal data of the target elderly population. The preprocessing includes any one or any combination of the following: adaptive noise reduction, heart sound segmentation, and quality assessment.

[0009] According to the method for constructing a heart sound detection model provided by the present invention, the heart sound signal data of the target elderly group is acquired by a dual-microphone array. The heart sound signal data includes the original heart sound signal acquired by the main microphone in the dual-microphone array and the noise reference signal acquired by the reference microphone in the dual-microphone array. The adaptive noise reduction is specifically as follows: based on the heart sound signal data of the target elderly group, adaptive noise reduction is performed through a pre-trained one-dimensional U-Net model to obtain the noise-reduced heart sound signal data. The pre-trained one-dimensional U-Net model is trained based on noisy heart sound / clean heart sound data.

[0010] According to the method for constructing a heart sound detection model provided by the present invention, heart sound segmentation specifically involves: obtaining the timestamp of each heart sound component based on the denoised heart sound signal data using a time series model, thereby realizing heart sound segmentation of the heart sound signal data. The time series model is trained based on historical heart sound signal data to learn the temporal relationship and morphological features between heart sound S1 segment and heart sound S2 segment.

[0011] According to the method for constructing a heart sound detection model provided by the present invention, the quality assessment specifically involves: calculating the Mel spectrum of each cardiac cycle signal after heart sound segmentation, and performing quality assessment through a pre-trained convolutional neural network to obtain qualified heart sound signal data.

[0012] According to a method for constructing a heart sound detection model provided by the present invention, the method involves training a model to learn the characteristic changes of different heart sound signals in the time and frequency domains based on heart sound signal data of a target elderly population, using a dual-branch neural network and an attention fusion mechanism, to obtain a heart sound detection model, comprising: Based on the heart sound signal data, the heart sound signal data is normalized through the data input module of the heart sound detection model to obtain a one-dimensional heart sound signal; Based on the heart sound signal data, the heart sound signal data is transformed into a Mel-time spectrum through the data input module of the heart sound detection model, which serves as a two-dimensional image representation of the heart sound signal data. Based on the one-dimensional heart sound signal, the first high-dimensional feature tensor is obtained by analyzing the temporal features of the two-branch neural network and using a one-dimensional residual convolutional network. Based on the Mel-time spectrum, the second high-dimensional feature tensor is obtained by analyzing the time-frequency features of the dual-branch neural network and using a two-dimensional deep separable convolutional neural network. By using a bidirectional long short-term memory network based on an attention mechanism, the first high-dimensional feature tensor and the second high-dimensional feature tensor are fused to obtain a fused feature vector; Based on the fused feature vector, the classification module of the heart sound detection model outputs the type probability of the heart sound signal to be tested.

[0013] According to a method for constructing a heart sound detection model provided by the present invention, the method utilizes a bidirectional long short-term memory network based on an attention mechanism to fuse a first high-dimensional feature tensor and a second high-dimensional feature tensor to obtain a fused feature vector, including: By concatenating the first high-dimensional feature tensor and the second high-dimensional feature tensor along the feature dimension, a preliminary fused feature is obtained. By incorporating the attention mechanism, attention weights are applied to the initially fused features to obtain weighted features. Based on the weighted features, a bidirectional long short-term memory network is used to obtain a fused feature vector that incorporates the contextual information of the entire sequence.

[0014] According to the method for constructing a heart sound detection model provided by the present invention, the expression of the weighted features is as follows: , , In the formula, GAP(·) represents global average pooling, and Linear(·) represents a linear layer. This represents the Sigmoid activation function. This indicates element-wise multiplication. Indicates preliminary fusion characteristics. Represents the attention weight matrix. This represents the weighted features.

[0015] According to the method for constructing a heart sound detection model provided by the present invention, the expression of the classification module is as follows: , in, and This represents the weights and biases of the classification module, and the output categories include: normal, aortic stenosis, mitral regurgitation, etc. Represents the fused feature vector. This represents the probability distribution for classification.

[0016] The present invention also provides a heart sound-assisted detection system, comprising: The data receiving module is used to: receive the heart sound signal of the subject from at least one terminal; The detection module is used to: based on the heart sound signal of the subject, and using the heart sound detection model obtained by any of the above-described methods for constructing the heart sound detection model, obtain the predicted type of the subject's heart sound signal; The data output module is used to send the predicted type of the heart sound signal of the subject to at least one terminal.

[0017] It should be noted that a terminal refers to an input / output device connected to a computer system. Depending on the function, terminals can be divided into various types: smart terminals or intelligent terminals, dumb terminals, interactive terminals or online terminals. Specifically, a terminal can be various mobile communication devices, such as mobile phones and tablets. This article aims to provide users with the function of inputting data and outputting data.

[0018] The present invention also provides an electronic device, including a processor and a memory storing a computer program, wherein the processor executes the computer program to implement the method for constructing the heart sound detection model described above.

[0019] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for constructing the heart sound detection model described above.

[0020] The present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer is able to execute any of the above-described methods for constructing a heart sound detection model.

[0021] The present invention provides a method and system for constructing a heart sound detection model, which can bring at least the following beneficial effects: This invention employs a dual-branch neural network combining time-domain and frequency-domain features, and introduces an attention fusion mechanism to effectively integrate the waveform temporal features of one-dimensional heart sound signals and the frequency-domain structural features of two-dimensional time-spectrum diagrams. This design not only enhances the feature representation ability of different heart sound types (such as systolic murmurs and diastolic murmurs), but also improves the accuracy of classification through deep feature-level fusion, thereby enabling more reliable identification of various heart sound abnormalities and providing high-precision auxiliary diagnostic results.

[0022] This invention employs a dual-microphone array-based data acquisition system and adaptive noise reduction based on a pre-trained one-dimensional U-Net, effectively suppressing environmental noise and breathing sounds to obtain purer heart sound signals. Simultaneously, it models data specific to the target elderly population and incorporates efficient network structures such as depthwise separable convolutions, allowing the model to focus more on learning the features of this group. This improves adaptability to individual physiological differences and enhances the model's robustness and generalization performance in real-world, complex scenarios.

[0023] By using a temporal model to accurately segment the denoised heart sounds (e.g., locating S1 and S2) and incorporating an attention mechanism, key information can be highlighted when fusing temporal and frequency domain features. This allows the heart sound detection model to not only output type probabilities but also to correlate its internal mechanisms with specific stages of the heart sound cycle. This helps to indicate to doctors the possible time and location of abnormalities (e.g., systole or diastole), thus providing more targeted and reliable auxiliary information for clinical decision-making and compensating for the lack of interpretability in traditional black-box models.

[0024] This invention employs lightweight structures such as two-dimensional depthwise separable convolutions in its network design, significantly reducing the number of model parameters and computational complexity. Combined with effective preprocessing procedures (such as quality assessment to screen qualified data) and efficient feature fusion strategies, it significantly reduces computational resource consumption during inference while maintaining high recognition accuracy. This enables the trained heart sound detection model to be deployed more smoothly on resource-constrained embedded devices or mobile terminals, providing a practical technical foundation for primary healthcare and remote home monitoring.

[0025] This invention enables a closed loop from data acquisition to result output, facilitating integration into existing medical devices or mobile health applications. It effectively improves the automation level and accessibility of early screening for cardiopulmonary diseases, and has significant practical value and promising prospects for widespread application. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating a method for constructing a heart sound detection model provided by the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of a heart sound detection model construction system provided by the present invention.

[0029] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0031] Figure 1 This is a flowchart illustrating a method for constructing a heart sound detection model according to the present invention. The execution entity of this method can be any applicable terminal-side device or network-side device, such as a heart sound detection model construction apparatus.

[0032] See Figure 1 The present invention provides a method for constructing a heart sound detection model, which may include: S110. Obtain heart sound signal data of the target elderly population, wherein the heart sound signal data includes multiple heart sound signals, and the multiple heart sound signals include any one of the following or any combination thereof: normal heart sounds, systolic murmurs, diastolic murmurs, and pericardial friction rubs.

[0033] In one embodiment, after acquiring the heart sound signal data, it can be preprocessed, wherein the preprocessing includes any one or any combination of the following: adaptive noise reduction, heart sound segmentation, and quality assessment.

[0034] In one embodiment, the heart sound signal data of the target elderly group can be acquired by a dual-microphone array based on MEMS (microelectromechanical systems) technology. The heart sound signal data includes the raw heart sound signal acquired by the main microphone in the dual-microphone array and the noise reference signal acquired by the reference microphone in the dual-microphone array.

[0035] The dual-microphone array boasts high signal-to-noise ratio (SNR) and wide frequency response, with key advantages including: a wider frequency response range (20Hz - 1000Hz and above), enabling complete capture of all effective components of heart sounds, including low-frequency third and fourth heart sounds (S3) and high-frequency noise; a higher SNR (SNR > 65dB), obtaining cleaner, less distorted signals from the source, providing high-quality raw materials for subsequent AI analysis; and excellent phase consistency, allowing for a dual-microphone array layout with one main microphone for heart sound acquisition and one reference microphone for ambient noise. By comparing the two signals, input is provided for subsequent active noise reduction algorithms.

[0036] Preprocessing can be performed by a dedicated low-power neural network processing unit (NPU) using a lightweight AI model, improving processing accuracy.

[0037] The adaptive noise reduction is specifically as follows: based on the heart sound signal data of the target elderly group, adaptive noise reduction is performed through a pre-trained one-dimensional U-Net model to obtain the noise-reduced heart sound signal data. The pre-trained one-dimensional U-Net model is trained based on noisy heart sound / clean heart sound data.

[0038] Traditional fixed filters cannot handle the ever-changing environmental noise (such as conversations and air conditioner noise) and body noise (such as breathing sounds and bowel sounds). This embodiment deploys a carefully pruned and quantized one-dimensional U-Net generative model on the NPU. This model takes the original mixed signal as input, its "encoder" part learns noise characteristics, and its "decoder" part reconstructs a pure heart sound signal. Trained on a large amount of data, this model can intelligently identify and separate heart sounds from noise, rather than simply filtering, avoiding the loss of useful frequency components and significantly improving the signal-to-noise ratio.

[0039] The heart sound segmentation is specifically as follows: based on the denoised heart sound signal data, the timestamp of each heart sound component is obtained using a time series model, thereby realizing the heart sound segmentation of the heart sound signal data. The time series model is trained based on historical heart sound signal data to learn the temporal relationship and morphological features between heart sound S1 segment and heart sound S2 segment.

[0040] The localization of heart sounds S1 and S2 is easily affected by factors such as arrhythmia and murmurs. Traditional thresholding methods based on amplitude or energy have low accuracy. This embodiment runs a miniature time-series model (miniature recurrent neural network RNN). This model takes the denoised signal as the time sequence input and learns the complex temporal relationship and morphological features between S1 and S2, thereby accurately predicting the boundary points of each heart sound component. This achieves extremely high segmentation accuracy, provides standardized data units for subsequent analysis, and retains only 1-3 complete, segmented cardiac cycle data for uploading, avoiding the uploading of long-term invalid recordings and saving more than 95% of data transmission, thus achieving low power consumption of the device.

[0041] The quality assessment specifically involves: calculating the Mel spectrum of each cardiac cycle signal after heart sound segmentation, and then performing quality assessment through a pre-trained convolutional neural network to obtain qualified heart sound signal data.

[0042] During heart sound acquisition, signal quality may be poor due to movement or unstable pressure. Uploading such signals would waste cloud resources and yield unreliable results. For each segmented cardiac cycle signal, it is converted into a microsecond-level spectrogram and input into a tiny convolutional neural network (CNN) for quality scoring (e.g., excellent, good, poor). If the score is "poor," an LED indicator on the device prompts the user to "remeasure." If the score is "excellent" or "good," the signal undergoes adaptive gain adjustment to standardize it to the optimal amplitude range.

[0043] In this embodiment, the front-end uses a combination of "professional hardware sensors + dedicated low-power AI chip (NPU) + lightweight intelligent algorithms" to achieve the transformation from "analog signal acquisition" to "digital, standardized, lightweight high-quality data packets", laying a solid and efficient foundation for the back-end deep AI analysis.

[0044] S120. Based on the heart sound signal data of the target elderly group, a dual-branch neural network and attention fusion mechanism are used to train the model to learn the characteristic changes of different heart sound signals in the time domain and frequency domain, and obtain the heart sound detection model.

[0045] In one embodiment, S120 may include: Based on the heart sound signal data, the heart sound signal data is normalized through the data input module of the heart sound detection model to obtain a one-dimensional heart sound signal. x(t) ; Based on the heart sound signal data, a short-time Fourier transform (STFT) is performed on the heart sound signal data through the data input module of the heart sound detection model, converting the heart sound signal data into a Mel-time spectrogram. M(f,t) , as a two-dimensional image representation of heart sound signal data; Based on one-dimensional heart sound signals x(t) By using the temporal feature analysis branch of a dual-branch neural network and a one-dimensional residual convolutional network (1D ResNet), the first high-dimensional feature tensor is obtained. F temp The residual structure effectively alleviates the gradient vanishing problem in deep networks, allowing networks to be built deeper to extract more complex temporal features. The first high-dimensional feature tensor... F temp It contains precise information on time-domain waveform morphology, amplitude, and temporal relationships, which is crucial for determining the intensity, splitting, and additional heart sounds (S3, S4) of the first heart sound S1 and the second heart sound S2. According to the Mel-time spectrum diagram M(f,t) By utilizing the time-frequency feature analysis branch of a dual-branch neural network and employing a 2D deepwise separable convolutional neural network (2D CNN), the number of parameters is greatly reduced while maintaining accuracy, resulting in a second high-dimensional feature tensor. F spec Second high-dimensional feature tensor F spec It contains rich global and local contextual information in the frequency domain and excels at capturing the frequency distribution and morphological characteristics of noise. A bidirectional long short-term memory (LSTM) network based on an attention mechanism is used to fuse the first and second high-dimensional feature tensors to obtain a fused feature vector. The attention mechanism teaches the model to "pay attention" to the most likely abnormal periods within the current cardiac cycle (such as systole or diastole) and assign higher weights to features from these periods. The Bi-LSTM is responsible for learning the temporal contextual dependencies of heart sound signals, such as the dynamic changes from S1 to S2 and then to the next S1 within a cycle. Specifically, the first and second high-dimensional feature tensors are concatenated along their feature dimensions to obtain the preliminary fused features. F fus By incorporating the attention mechanism, attention weights are applied to the initially fused features to obtain weighted features. By utilizing a bidirectional long short-term memory network, a fused feature vector that incorporates contextual information from the entire sequence is obtained. Since heart sounds are strictly temporal signals, the weighted features... F att A bidirectional long short-term memory (Bi-LSTM) network is input in time slices. The Bi-LSTM learns the long-term temporal dependencies of heart sound signals in both forward and backward directions, such as the impact of the interval between S1 and S2, and the duration of diastole, on the overall cycle determination. The output of the last hidden layer of the Bi-LSTM... H lstm It gathers contextual information from the entire sequence; The expression for the weighted feature is: , , In the formula, GAP(·) represents global average pooling, and Linear(·) represents a linear layer. This represents the Sigmoid activation function. This indicates element-wise multiplication. Indicates preliminary fusion characteristics. Represents the attention weight matrix. Indicates the weighted features; Based on the fused feature vector, the classification module of the heart sound detection model outputs the type probability of the heart sound signal to be detected. Specifically, it will... H lstm Inputting a fully connected layer with a Softmax activation function yields the final classification probability distribution, where the expression for the classification module is: , in, and This represents the weights and biases of the classification module, and the output categories include: normal, aortic stenosis, mitral regurgitation, etc. Represents the fused feature vector. This represents the probability distribution for classification.

[0046] In terms of interpretability, the attention weight matrix A calculated earlier in the feature fusion stage is upsampled along the temporal dimension (if necessary) to make its length consistent with the original input signal x(t). This weight sequence A(t) directly reflects the degree of "attention" the model pays to different time points of the original heart sound signal when making classification decisions. Regions with high weights are those that the model considers to be abnormal (such as the systolic or diastolic phase of a murmur). On the user end (App or Web interface), this A(t) is overlaid on the original heart sound waveform in the form of a heatmap, intuitively telling doctors or patients where the problem is and what the problem might be, thus achieving visualization and interpretability of the diagnostic results.

[0047] The heart sound-assisted detection system provided by the present invention will be described below. The heart sound-assisted detection system described below and the construction method of the heart sound detection model described above can be referred to in correspondence.

[0048] See Figure 2 The present invention provides a heart sound-assisted detection system, which may include: The data receiving module is used to: receive the heart sound signal of the subject from at least one terminal; The detection module is used to: based on the heart sound signal of the subject, and using the heart sound detection model obtained by any of the above-described methods for constructing the heart sound detection model, obtain the predicted type of the subject's heart sound signal; The data output module is used to send the predicted type of the heart sound signal of the subject to at least one terminal.

[0049] In one embodiment, the heart sound-assisted detection system includes hardware and software components, the hardware component including a smart stethoscope head, a mobile terminal (such as a smartphone), and a cloud server.

[0050] ① Intelligent stethoscope head, including: Sensing module: Employs an InvenSense ICS-43434 MEMS digital microphone with a frequency response of 20Hz - 20kHz and a signal-to-noise ratio (SNR) greater than 65dB. It uses a dual-microphone array layout, with the main microphone located at the bottom of the stethoscope head cavity and a reference microphone located on the side for ambient noise acquisition.

[0051] Preprocessing module: Employs a low-power NPU that supports convolutional neural network acceleration, such as the Ambiq Apollo4Plus SoC. This chip integrates a Cortex-M4 core and a neural network accelerator (Tensilica-VP6), with an operating power consumption of less than 1mW.

[0052] Communication module: Uses TI's CC2642R chip, which supports Bluetooth 5.2 Low Energy (BLE) protocol.

[0053] Power management module: Powered by a 200mAh lithium polymer battery and equipped with corresponding charging management circuitry.

[0054] ② Mobile terminal Install a dedicated app to receive auscultation head data, preview it locally, encrypt and upload the data to the cloud, and receive and display the final diagnostic report.

[0055] ③ Cloud server A computing cluster equipped with high-performance GPUs (such as NVIDIA V100) is used to deploy and run the heart sound detection model of this invention.

[0056] Regarding the software algorithm section Step 1: Signal Acquisition and Front-End Preprocessing (executed on the intelligent stethoscope head NPU), detailed as follows: The main microphone and the reference microphone synchronously acquire the raw signal at a sampling rate of 4kHz. x raw (t) and noise reference signal nref (t) .

[0057] Adaptive noise reduction, x raw (t) and n ref (t) Input a lightweight one-dimensional U-Net model pre-installed in the NPU. This model has been trained on a large amount of noisy / clean heart sound data and has undergone quantization and compression. The model outputs the denoised heart sound signal. x clean (t) .

[0058] Heart sounds segmented, x clean (t) Input a miniature gated recurrent unit (GRU) network. This GRU network has 3 hidden layers, each with 32 units. The network outputs a probability sequence of events S1 and S2, from which the precise timestamp t for each heart sound component can be obtained through post-processing. s1 start ,t s1 end ,t s2 start ,t s2 end This completes the segmentation.

[0059] Quality assessment: For each segmented cardiac cycle signal, its Mel spectrum is calculated and input into a lightweight convolutional neural network (4 convolutional layers) for binary classification ("qualified" or "unqualified"). If unqualified, the user is alerted by the flashing red LED on the stethoscope head; if qualified, the signal amplitude is normalized.

[0060] Step 2: Data Upload The mobile app receives qualified cardiac cycle data packets (usually containing 3 consecutive cycles) via BLE (Bluetooth Low Energy), encrypts the data using the AES-256 encryption algorithm, and then uploads it to the designated cloud server API interface via the HTTPS protocol.

[0061] Step 3: In-depth analysis in the cloud (executed on a cloud server) Data preparation: After receiving the data in the cloud, the one-dimensional signal is standardized. Simultaneously, a corresponding Mel-time spectrum is generated for this signal, with parameters set as follows: window length 1024, hop length 256, and number of Mel filters 128.

[0062] Heart sound detection model inference: one-dimensional signal x(t)The input is a time-domain branch, which is a 1D ResNet-18 network with an input size of (4000, 1) and an output of a 512-dimensional feature vector. F temp .

[0063] Mel spectrum M(f,t) The input time-frequency branch is an EfficientNet-B0-based network with an input size of (128, 128, 1) and an output 512-dimensional feature vector. F spec .

[0064] Feature fusion will F temp and F spec The features are then concatenated to obtain a 1024-dimensional fused feature. F fus .

[0065] Attention and context modeling will F fus Input an attention module (consisting of a fully connected layer and a sigmoid activation function) and generate attention weights A. Then, weight the features... F att Input a bidirectional LSTM (128 hidden layer dimensions) sequentially, and extract its output at the last time step. H lstm As the final feature.

[0066] The output will be H lstm Input a fully connected classifier and use the Softmax function to obtain the disease classification probability. P(y∣ x) Simultaneously, the attention weight A is upsampled back to the original signal length to generate a saliency map for visualization.

[0067] Step 4: Result Generation and Sending The cloud-based system assembles the analysis results (including classification labels, confidence scores, and significance graph data) into a JSON report and sends it to the mobile app. After parsing, the app displays the data in a graphical interface, such as displaying waveforms and highlighting areas of high model interest with a red heatmap, while also providing a "Play Abnormal Segments" button.

[0068] Experimental data verification and results To verify the effectiveness of this embodiment, a comprehensive experiment was conducted using the PhysioNet / Circulation 2016 publicly available heart sound dataset (available on the PhysioNet website: https: / / www.physionet.org / content / challenge-2016 / 1.0.0) and a self-built clinical dataset for comparison and analysis.

[0069] Comparison Experiment of Different Model Architectures This experiment aims to demonstrate the superiority of the dual-branch multimodal structure (DB-SAFN, i.e., the heart sound detection model in this embodiment) over the single-modal model. All models used the same training and test sets, and the comparative verification results are shown in Table 1 below: Table 1

[0070] Note: Traditional SVM is based on handcrafted features, one-dimensional CNN only supports the temporal branch, and two-dimensional CNN only supports the time-frequency branch.

[0071] Experimental results show that the proposed dual-branch fusion model (DB-SAFN) significantly outperforms the single-modality model across all evaluation metrics, achieving a sensitivity of 99.1%, demonstrating its ability to detect almost all abnormal heart sounds with very few missed diagnoses, which is crucial for early disease screening. The AUC (area under the curve) is as high as 0.995, indicating excellent classification performance.

[0072] ablation experiment This experiment aims to demonstrate the necessity of key components (attention mechanism, front-end preprocessing) in the model. Using the complete DB-SAFN model as a baseline, key components were removed sequentially, and comparative experiments were conducted to demonstrate the contribution of each innovative component to the results. Please refer to Table 2.

[0073] Table 2

[0074] Ablation experiments show that the attention mechanism effectively improves model performance, and the interpretability it brings is an added value; the dual-branch structure is the core of performance improvement, and the absence of either branch will lead to performance loss; intelligent preprocessing at the front end is the cornerstone of high-performance analysis at the back end. Without it, model performance will be severely degraded by noise interference.

[0075] Robustness testing in noisy environments The model's performance in noisy environments was tested by simulating additive white Gaussian noise of varying intensities. This experiment aimed to demonstrate the system's reliability in real-world applications (such as homes and community clinics). The model's performance retention rate (%) under different signal-to-noise ratio (SNR) conditions is shown in Table 3 below. Table 3

[0076] Note: Performance retention rate = (Indicators under current conditions / Indicators under clean signal) * 100%. Data in the table represents accuracy / sensitivity. Thanks to the intelligent noise reduction of the front-end NPU, this system exhibits strong robustness in various noisy environments. Even under harsh conditions with a signal-to-noise ratio as low as 5dB, the system maintains over 90% sensitivity, far superior to the baseline model without preprocessing. This demonstrates that the present invention is highly suitable for deployment in real-world environments with varying conditions. In addition, this system can also be used in remote consultation scenarios. After primary care physicians use the aforementioned intelligent stethoscope head to collect patients' heart sounds, they can directly share the report with remote experts via an app. Upon receiving the structured report, which includes waveforms, heat map annotations, and preliminary AI conclusions, experts can quickly grasp the key points of the patient's condition and make a final diagnosis based on their own experience, greatly improving the efficiency and accuracy of consultations.

[0077] The present invention provides a method and system for constructing a heart sound-assisted detection system. The dual-branch multimodal deep learning model network can capture subtle abnormalities in heart sounds from multiple dimensions, significantly improving the detection rate (sensitivity) of early lesions. The front-end AI preprocessing can effectively purify the signal and ensure the input quality of the back-end analysis model. The saliency map generated by the attention mechanism makes the AI ​​judgment process transparent and easy for medical professionals to understand and accept. Through a heterogeneous computing architecture, complex calculations are offloaded to the cloud. The front-end adopts a low-power NPU and intelligent wake-up strategy, which makes the device have a long battery life and suitable for home and community use. The complete technology stack (terminal-communication-cloud) optimization is adapted to various application scenarios from tertiary hospitals to primary clinics and even homes, which helps to alleviate the problem of uneven distribution of medical resources.

[0078] Figure 3 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 3 As shown, the electronic device may include a processor 810, a communications interface 820, a memory 830, and a communication bus 840, wherein the processor 810, the communications interface 820, and the memory 830 communicate with each other via the communication bus 840. The processor 810 can call logical instructions in the memory 830 to execute the following steps: Receive the heart sound signal of the subject from at least one terminal; Based on the heart sound signal of the subject, the heart sound detection model obtained by the heart sound detection model construction method described above is used to obtain the prediction result of the type of heart sound signal of the subject; The predicted type of the subject's heart sound signal is sent to at least one terminal.

[0079] Furthermore, the logical instructions in the aforementioned memory 830 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present 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 described in the various embodiments of the present 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.

[0080] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program, the computer program being able to be stored on a non-transitory computer-readable storage medium, and the computer program being executed by a processor, enabling the computer to perform the following steps: Receive the heart sound signal of the subject from at least one terminal; Based on the heart sound signal of the subject, the heart sound detection model obtained by the heart sound detection model construction method described above is used to obtain the prediction result of the type of heart sound signal of the subject; The predicted type of the subject's heart sound signal is sent to at least one terminal.

[0081] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps: Receive the heart sound signal of the subject from at least one terminal; Based on the heart sound signal of the subject, the heart sound detection model obtained by the heart sound detection model construction method described above is used to obtain the prediction result of the type of heart sound signal of the subject; The predicted type of the subject's heart sound signal is sent to at least one terminal.

[0082] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0083] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0084] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a heart sound detection model, characterized in that, include: Acquire heart sound signal data of the target elderly population, including various heart sound signals; Based on the heart sound signal data of the target elderly group, a dual-branch neural network and attention fusion mechanism are used to train the model to learn the characteristic changes of different heart sound signals in the time domain and frequency domain, thus obtaining a heart sound detection model.

2. The method for constructing a heart sound detection model according to claim 1, characterized in that, The acquisition of heart sound signal data of the target elderly group includes: Preprocessing is performed on the heart sound signal data of the target elderly population. The preprocessing includes any one or any combination of the following: adaptive noise reduction, heart sound segmentation, and quality assessment.

3. The method for constructing a heart sound detection model according to claim 2, characterized in that, Heart sound signal data of the target elderly population were acquired through a dual-microphone array. The heart sound signal data included the raw heart sound signal acquired through the main microphone in the dual-microphone array and the noise reference signal acquired through the reference microphone in the dual-microphone array. The adaptive noise reduction is specifically as follows: based on the heart sound signal data of the target elderly group, adaptive noise reduction is performed through a pre-trained one-dimensional U-Net model to obtain the noise-reduced heart sound signal data. The pre-trained one-dimensional U-Net model is trained based on noisy heart sound / clean heart sound data.

4. The method for constructing the heart sound detection model according to claim 3, characterized in that, Heart sound segmentation specifically involves: based on the denoised heart sound signal data, obtaining the timestamp of each heart sound component using a time-series model, thereby achieving heart sound segmentation of the heart sound signal data. The time-series model is trained based on historical heart sound signal data to learn the temporal relationship and morphological features between heart sound segment S1 and heart sound segment S2; and / or, The quality assessment specifically involves: calculating the Mel spectrum of each cardiac cycle signal after heart sound segmentation, and then performing quality assessment through a pre-trained convolutional neural network to obtain qualified heart sound signal data.

5. The method for constructing a heart sound detection model according to any one of claims 1-4, characterized in that, The method involves using heart sound signal data from the target elderly population, employing a dual-branch neural network and attention fusion mechanism to train a model that learns the characteristic changes of different heart sound signals in the time and frequency domains, thereby obtaining a heart sound detection model, including: Based on the heart sound signal data, the heart sound signal data is normalized through the data input module of the heart sound detection model to obtain a one-dimensional heart sound signal; Based on the heart sound signal data, the heart sound signal data is transformed into a Mel-time spectrum through the data input module of the heart sound detection model, which serves as a two-dimensional image representation of the heart sound signal data. Based on the one-dimensional heart sound signal, the first high-dimensional feature tensor is obtained by analyzing the temporal features of the two-branch neural network and using a one-dimensional residual convolutional network. Based on the Mel-time spectrum, the second high-dimensional feature tensor is obtained by analyzing the time-frequency features of the dual-branch neural network and using a two-dimensional deep separable convolutional neural network. By using a bidirectional long short-term memory network based on an attention mechanism, the first high-dimensional feature tensor and the second high-dimensional feature tensor are fused to obtain a fused feature vector; Based on the fused feature vector, the classification module of the heart sound detection model outputs the type probability of the heart sound signal to be tested.

6. The method for constructing a heart sound detection model according to claim 5, characterized in that, The method utilizes a bidirectional long short-term memory network based on an attention mechanism to fuse the first high-dimensional feature tensor and the second high-dimensional feature tensor to obtain a fused feature vector, including: By concatenating the first high-dimensional feature tensor and the second high-dimensional feature tensor along the feature dimension, a preliminary fused feature is obtained. By incorporating the attention mechanism, attention weights are applied to the initially fused features to obtain weighted features. Based on the weighted features, a bidirectional long short-term memory network is used to obtain a fusion feature vector that incorporates the contextual information of the entire sequence.

7. The method for constructing a heart sound detection model according to claim 6, characterized in that, The expression for the weighted feature is: ， ， In the formula, GAP(·) represents global average pooling, and Linear(·) represents a linear layer. This represents the Sigmoid activation function. This indicates element-wise multiplication. Indicates preliminary fusion characteristics. Represents the attention weight matrix. Indicates the weighted features; The expression for the classification module is: ， in, and This represents the weights and biases of the classification module. Represents the fused feature vector. This represents the probability distribution for classification.

8. A heart sound-assisted detection system, characterized in that, include: The data receiving module is used to: receive the heart sound signal of the subject from at least one terminal; The detection module is used to: obtain a prediction result of the type of the heart sound signal of the subject based on the heart sound signal of the subject, using the heart sound detection model obtained by the construction method of the heart sound detection model according to any one of claims 1-7; The data output module is used to send the predicted type of the heart sound signal of the subject to at least one terminal.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method for constructing the heart sound detection model as described in any one of claims 1 to 7.

10. A non-transitory 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 method for constructing the heart sound detection model as described in any one of claims 1 to 7.

Images