Heart sound signal acquisition and automatic classification method based on deep learning

Abstract

The invention discloses a heart sound signal acquisition and automatic classification method based on deep learning, and aims to solve the problems that wireless transmission time delay of heart sound signals is high, weak pathological features are difficult to extract and classification accuracy is low. The method comprises the following steps: acquiring signals of four auscultation areas by using a high-fidelity wireless digital stethoscope; high-speed and low-delay transmission of medical-level signals is realized by utilizing a star flash transmission module and combining predictive coding and Huffman coding compression mechanisms; in the preprocessing stage, self-adaptive wavelet denoising based on signal energy entropy is adopted, a multi-resolution fusion feature extraction strategy is provided, the problem of time-frequency resolution tradeoff is solved through narrow window and wide window dual-channel parallel processing, and transient positioning features and pathological frequency domain textures are captured at the same time; the classification model adopts a multi-scale parallel convolution structure and is matched with a recurrent neural network and an attention mechanism, so that the problem that transient and continuous heart sound features are difficult to consider at the same time in a single scale is solved. According to the method, the accuracy and robustness of heart sound classification are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of heart sound signal collection and processing, and particularly relates to a heart sound signal automatic classification method based on deep learning. BACKGROUND

[0002] The present application relates to the technical field of medical electronics and artificial intelligence assisted diagnosis, and particularly relates to a heart sound signal collection and classification method and system. Cardiovascular disease is one of the main causes of death worldwide. Cardiac auscultation, that is, analyzing heart sound signals through a stethoscope, is the most commonly used, most convenient and lowest cost early screening method for cardiovascular disease in clinical practice. Normal heart sounds mainly include first heart sound and second heart sound, while the presence of abnormal heart sounds usually indicates pathological conditions of heart valves or structures. However, the traditional acoustic stethoscope relies heavily on the subjective experience and hearing acuity of doctors. Heart sound signals are very weak in themselves and are easily disturbed by patient breathing sounds, body friction sounds and environmental noise. This results in a highly unstable accuracy of artificial auscultation, especially in primary medical institutions, the lack of experienced cardiologists leads to a high rate of missed diagnosis and misdiagnosis. In order to overcome the above-mentioned defects, computer-aided diagnosis systems have been proposed in the art. Early computer-aided diagnosis systems usually adopt a scheme of "manual feature engineering plus traditional machine learning". The scheme first denoises and segments the collected heart sound signals, then manually designs and extracts a series of time domain features, frequency domain features or time-frequency features, and finally uses classifiers such as support vector machines and K-nearest neighbors for classification. The performance of such methods is severely limited by the quality and representation ability of the extracted features, and it is difficult to capture all the pathological information in complex heart sound signals, so the classification accuracy is difficult to meet the clinical requirements. With the development of electronic technology, digital electronic stethoscopes and heart sound collection devices have appeared. These devices have realized the digitization of signals, but there are still deficiencies in signal collection and transmission. For example, many devices only collect in a single auscultation area, ignoring the fact that different pathologies have different optimal auscultation positions in different auscultation areas, resulting in the loss of information. In addition, in existing wireless heart sound collection, Bluetooth or wireless local area network technology is mostly used for data transmission. The heart sound signals collected by the present application need to be input into a deep learning model for fine analysis, which requires the transmission link to provide an original data stream that has not been lossy compressed. The traditional Bluetooth technology has weak anti-interference ability in complex electromagnetic environments such as hospitals, and is prone to data packet loss or transmission jitter; the wireless local area network technology uses a competitive channel access mechanism, which is prone to non-deterministic delay and burst packet loss in complex electromagnetic environments such as hospitals. For real-time heart sound stream data with strict requirements, such jitter can cause discontinuity or artifacts in the waveform, and thus cause misdiagnosis. Therefore, there is an urgent need in the art for a new heart sound signal collection and classification method and system that can not only collect heart sound signals in multiple areas with high fidelity, but also transmit them in a more stable and efficient manner, and can also objectively and accurately automatically classify the signals through more advanced intelligent algorithms. SUMMARY

[0003] The present application solves the technical problems of the prior art and provides a heart sound signal collection and automatic classification method based on deep learning.

[0004] To address the aforementioned technical problems, this invention discloses a method for acquiring and automatically classifying heart sound signals based on deep learning, comprising the following steps:

[0005] Step 1: Use a wireless digital stethoscope device to collect data from the target area, and collect heart sound signals through a vibration sensor probe.

[0006] Step 2: Repeat the single acquisition process described in Step 1 continuously to obtain the original analog signal, and process and transmit it using high-speed, low-latency transmission hardware to obtain a high-amplitude digital signal;

[0007] Step 3: Preprocess the received heart sound signals;

[0008] Step 4: Input the preprocessed signal into a pre-trained deep learning classification model to obtain the classification result of the heart sound signal as "normal" or "abnormal".

[0009] In step 1, the heart sound signals are collected in the four main auscultation areas of the human body, including the aortic valve area, pulmonary valve area, tricuspid valve area, and mitral valve area.

[0010] The method for collecting heart sound signals described in step 2 is to press the probe of a handheld digital stethoscope onto the auscultation area to collect the signal.

[0011] Step 2, the high-speed low-latency transmission hardware processing method, includes:

[0012] Step 2-1: The low-noise signal amplification circuit initially amplifies the weak original heart sound signal, significantly improving its signal-to-noise ratio and amplitude, laying the foundation for subsequent processing.

[0013] Step 2-2: The high-precision analog-to-digital converter samples and quantizes the amplified analog signal, converting it into a high-fidelity digital signal to ensure the integrity of the heart sound waveform characteristics and facilitate subsequent digital signal processing.

[0014] Steps 2-3: The processed digital heart sound signal is transmitted through the StarFlash communication module.

[0015] The StarFlash communication module described in step 2 of this invention provides a high-speed, extremely low-latency, and stable wireless data transmission link for the core, thanks to its high bandwidth, low power consumption, and strong anti-interference capability. In order to adapt to the characteristics of large instantaneous data volume and high latency requirements of heart sound signals, a compression mechanism combining predictive coding and Huffman coding, as well as a transmission mechanism of gateway polling and packet transmission, are added, which effectively meets the real-time and reliability requirements of heart sound signal transmission.

[0016] The heart sound signal preprocessing method described in step 3 includes:

[0017] Step 3-1: The original signal is denoised by a combination of bandpass filtering and wavelet transform. Bandpass filtering effectively filters out environmental and physiological interference outside the main frequency band of heart sounds, while wavelet transform can further eliminate residual transient noise in the time domain, thereby significantly improving the signal-to-noise ratio.

[0018] Step 3-2: Perform peak normalization on the denoised signal to uniformly scale its amplitude to a predetermined range, so as to eliminate amplitude deviation caused by differences in acquisition conditions and provide a stable data foundation for model training.

[0019] Step 3-3: Use block and sliding window techniques to perform data augmentation on the normalized continuous heart sound signal to generate a large number of one-dimensional time domain signal segments with temporal context information.

[0020] Steps 3-4 involve feature extraction using a time-frequency complementary feature extraction method based on a narrow-wide-window dual-channel parallel processing approach. The preprocessed heart sound signal blocks are input into the dual-channel parallel processing stream. In the first channel, a short-time Fourier transform is performed on the signal using a narrow time window to capture the transient temporal localization features of the first and second heart sounds, generating a high-temporal-resolution spectrogram. In the second channel, a short-time Fourier transform is performed on the same signal using a wide time window to capture the fine-grained frequency-domain texture features of cardiac pathological murmurs, generating a high-frequency-resolution spectrogram. The resulting high-temporal-resolution spectrogram and the high-frequency-resolution spectrogram are then mapped using a Mel filter bank and spliced ​​and fused along the channel dimension to form a composite feature map containing multi-scale time-frequency information. The energy value of the composite feature map is converted to a logarithmic scale to obtain the final features input to the model. The feature extraction technology proposed in this invention differs from the feature extraction methods using a single window length in existing technologies. It is a feature extraction strategy based on multi-resolution fusion, which effectively solves the trade-off problem between time and frequency resolution. This allows the back-end neural network to simultaneously acquire accurate cardiac cycle localization information and delicate pathological noise spectrum features, thereby significantly improving the classification accuracy of faint noises.

[0021] In the wavelet transform denoising step described in step 3 of this invention, an adaptive threshold selection strategy based on signal energy entropy is employed. The energy entropy of the wavelet coefficients at each decomposition layer of the heart sound signal is calculated, and the threshold function is dynamically adjusted according to the magnitude of the energy entropy. When the energy entropy is high, it is identified as a complex pathological signal or strong noise, and a softer threshold is used to preserve details. When the energy entropy is low, it is identified as a background silence period, and a harder threshold is used to thoroughly filter out noise. Thus, while removing environmental noise, the characteristics of weak heart murmurs are preserved to the greatest extent possible.

[0022] The heart sound signal classification method described in step 4 includes:

[0023] Step 4-1 involves using a convolutional neural network module, which employs a multi-scale parallel convolutional structure. This structure includes at least two parallel convolutional branches: the first branch uses a small-sized convolutional kernel to capture the transient impact features of the first and second heart sounds; the second branch uses a large-sized convolutional kernel or dilated convolution to capture the long-term persistent texture features of heart murmurs. The output features of the two branches are concatenated and fused along the channel dimension before being input into subsequent pooling layers. This structure solves the problem that a single-sized convolutional kernel cannot simultaneously capture both transient and persistent features of heart sounds.

[0024] Step 4-2: The feature sequence is processed in depth through the recurrent neural network module. The long-range temporal dependencies and dynamic evolution patterns between events in the heart sound cycle are captured through its internal gating mechanism, so as to model the overall rhythm and pathological pattern of the heart sound signal.

[0025] Step 4-3: Through the temporal attention mechanism module, the features of all time steps output by the recurrent neural network are adaptively weighted, amplifying the contribution of important moments while suppressing irrelevant information, thereby generating a context vector focused on discriminative information as the final classification basis;

[0026] Step 4-4: Through a fully connected classifier, nonlinear transformation and feature integration are used to output the precise probability distribution of whether the heart sound signal belongs to "normal" or "abnormal", thus completing the automated classification.

[0027] Beneficial effects: By employing an adaptive threshold selection strategy based on signal energy entropy, environmental noise is removed while the characteristics of weak heart murmurs are preserved to the greatest extent, solving the problem that previous methods could not effectively distinguish between useful signals and murmurs. The feature extraction method based on a narrow-wide-window dual-channel parallel time-frequency complementary feature extraction method effectively solves the trade-off between time-frequency resolution and significantly improves the classification accuracy of weak murmurs. The multi-scale parallel convolutional structure solves the problem that a single-size convolutional kernel cannot simultaneously capture both transient and persistent features of heart sounds. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the present invention.

[0029] Figure 2 This is a schematic diagram of the hardware structure for heart sound acquisition and transmission.

[0030] Figure 3 This is a flowchart of the heart sound signal preprocessing process.

[0031] Figure 4 This is a comparison of heart sound signals before and after preprocessing.

[0032] Figure 5This is a schematic diagram of the network structure of a deep learning classification model. Detailed Implementation

[0033] This embodiment is applied to the detection of abnormal heart sounds. In the field of abnormal heart sound detection, research still faces many challenges due to the weak heart sound signals and susceptibility to noise interference. In the past, most methods were based on manual feature engineering plus traditional machine learning. However, the performance limit of such methods is severely limited by the quality and representation ability of the extracted features, making it difficult to capture all the pathological information in complex heart sound signals. Therefore, the classification accuracy is difficult to meet clinical needs.

[0034] The method for acquiring and automatically classifying heart sound signals based on deep learning described in this embodiment has the following system design flow: Figure 1 This includes the following steps:

[0035] Step 1: Use a wireless digital stethoscope device to collect data from the target area, and collect heart sound signals through a piezoelectric film vibration sensor probe;

[0036] Step 2: Repeat the single acquisition process described in Step 1 continuously to obtain the analog heart sound signal, and process and transmit the original analog signal to obtain a high-amplitude digital heart sound signal at the volt level;

[0037] Step 3: Preprocess the received digital heart sound signal, and extract features from the preprocessed signal using a time-frequency complementary feature extraction method based on a narrow-window dual-channel parallel approach.

[0038] Step 4: Input the extracted features into a pre-trained deep learning classification model to obtain a classification result of whether the heart sound signal is "normal" or "abnormal".

[0039] In step 1, the heart sound signals are collected in the four main auscultation areas of the human body, including the aortic valve area, pulmonary valve area, tricuspid valve area, and mitral valve area.

[0040] The method for acquiring heart sound signals described in step 2 is to press the handheld piezoelectric film vibration sensor probe against the auscultation area to acquire the signal.

[0041] Step 2, the processing method, includes:

[0042] Step 2-1: A low-noise signal amplification circuit constructed using the AD620 instrumentation amplifier chip is used to initially amplify the weak original heart sound signal, with an amplification factor of up to 200 times that of the original signal, thereby improving its signal-to-noise ratio and amplitude, laying the foundation for subsequent processing.

[0043] Step 2-2: A high-precision analog-to-digital converter circuit composed of an independent analog-to-digital converter chip ADC0832 is used to sample and quantize the amplified analog signal, converting it into a high-fidelity digital signal. The sampling frequency is 2500 Hz and the quantization level is 256, thereby ensuring the integrity of the heart sound waveform characteristics and facilitating subsequent digital signal processing.

[0044] Steps 2-3: The processed digital heart sound signal is transmitted through the StarFlash communication module based on the WS63 chip.

[0045] The StarScan communication module includes a compression mechanism combining predictive coding and Huffman coding, as well as a transmission mechanism of gateway polling and packet-based transmission. With its high bandwidth, low power consumption, and strong anti-interference capabilities, the StarScan communication module provides a high-speed, extremely low-latency, and stable wireless data transmission link for the core. To adapt to the characteristics of large instantaneous data volume and high latency requirements of heart sound signals, a compression mechanism combining predictive coding and Huffman coding, along with a transmission mechanism of gateway polling and packet-based transmission, has been added. The compressed byte count is 20% of the uncompressed count, effectively meeting the real-time and reliability requirements of heart sound signal transmission.

[0046] The heart sound signal preprocessing method in step 3 includes:

[0047] Step 3-1-1 involves denoising the original signal using a combination of bandpass filtering and an adaptive wavelet threshold modulation method based on multi-scale energy entropy measurement. Bandpass filtering effectively removes environmental and physiological interference outside the main heart sound frequency band, while wavelet transform further eliminates residual transient noise in the time domain, thus significantly improving the signal-to-noise ratio. The bandpass filter uses a finite impulse response digital filter with strictly linear phase characteristics, and the passband frequency range is 20 Hz to 600 Hz. The adaptive wavelet threshold modulation method based on multi-scale energy entropy measurement uses the Dobermann wavelet basis function, and the decomposition level is set to 5 levels.

[0048] Step 3-1-2: Perform peak normalization on the denoised signal to uniformly scale its amplitude to the closed interval [-1,1] to eliminate amplitude deviation caused by differences in acquisition conditions and provide a stable data foundation for model training.

[0049] Step 3-1-3: Use block segmentation and sliding window to perform data augmentation on the normalized continuous heart sound signal to generate a large number of one-dimensional time domain signal segments with temporal context information; specifically, the time length of each one-dimensional time domain signal segment is set to 5 seconds, and the step size of the sliding window is 2.5 seconds.

[0050] The feature extraction in step 3 specifically involves:

[0051] Step 3-2-1: Input the preprocessed one-dimensional time-domain signal segments into the dual-channel parallel processing stream respectively;

[0052] In the first channel, a narrow time window is used to perform a short-time Fourier transform on the signal. The window length of the narrow time window is set to a Hanning window with 64 sampling points, the frame shift is set to 32 sampling points, and the number of fast Fourier transform points is set to 1024, in order to capture the transient temporal localization features of the first and second heart sounds and generate a high temporal resolution spectrogram with an effective temporal resolution of 25.6 milliseconds.

[0053] In the second channel, a wide time window is used to perform a short-time Fourier transform on the same signal. The window length of the wide time window is set to a Hanning window with 256 sampling points, the frame shift is set to 32 sampling points, and the number of fast Fourier transform points is set to 1024 to capture the fine frequency domain texture features of cardiac pathological noise and generate a high-frequency resolution spectrogram with a physical frequency resolution of 9.7 Hz.

[0054] Step 3-2-2: The obtained high temporal resolution spectrogram and high frequency resolution spectrogram are mapped using a Mel filter bank, and then spliced ​​and fused along the channel dimension to form a composite feature map containing multi-scale time-frequency information. The first scale is a micro-temporal scale based on a narrow window with 64 sampling points; the second scale is a macro-frequency scale based on a wide window with 256 sampling points. The number of filters in the Mel filter bank is set to 128, thereby generating a feature map with a height of 128 pixels.

[0055] Step 3-2-3: Convert the energy value of the composite feature map to a logarithmic scale using a logarithmic decibel conversion to obtain the final input features of the model.

[0056] The feature extraction technique proposed in this invention differs from existing methods that use a single window length. It is a time-frequency complementary feature extraction method based on a narrow-wide-window dual-channel parallel approach. The narrow window is used to analyze transient signals, while the wide window analyzes steady-state continuous signals. This combination addresses the problem of simultaneously analyzing complex pathological signals where transient and steady-state signals coexist, effectively resolving the trade-off between time-frequency resolution. This allows the backend neural network to simultaneously acquire accurate cardiac cycle localization information and subtle pathological murmur spectral features, significantly improving the classification accuracy of faint murmurs. Specifically, this invention is specifically optimized for the challenging pathological scenario of mitral valve prolapse accompanied by regurgitation. In this scenario, the heart sound signal characteristics are: a high-frequency, non-ejector click lasting less than 30 milliseconds during mid-systole, followed by a noise-like late-systolic regurgitation murmur that lasts until the second heart sound. In existing technologies, while a single wide window can clearly present the frequency domain texture of regurgitation noise, the extremely short clicks are often masked or confused with other noises due to temporal smoothing effects, leading to missed diagnoses of prolapse. Conversely, a single narrow window, while capturing the click's timing, suffers from extremely low frequency resolution, presenting the regurgitation noise as blurry broadband noise, making it difficult to determine the degree of regurgitation. This invention, through channel splicing and fusion, generates a composite feature map on the same time axis, preserving both the sharp temporal edges of the clicks extracted by the narrow window channel and the fine frequency harmonic structure of the regurgitation noise extracted by the wide window channel. This time-frequency complementary feature representation allows the backend neural network to make accurate classifications with clinical and pathological evidence based on the precise temporal relationship between clicks and other noises—something that single-scale feature extraction methods cannot achieve.

[0057] In step 3-1-1 of the wavelet transform denoising process, an adaptive threshold selection strategy based on signal energy entropy is adopted. First, the threshold of the wavelet decomposition is calculated. Normalized energy entropy of layer detail coefficients The calculation formula is:

[0058]

[0059] in For the first The energy percentage of each coefficient This represents the total length of the coefficients in this layer. The normalized energy entropy. The value range is [0,1]; the general denoising threshold is... , in This represents the total number of sampling points for the current one-dimensional time-domain heart sound signal segment. This represents the standard deviation estimate of the background noise, calculated by selecting the high-frequency detail coefficient vector of the first layer of wavelet decomposition. Calculate the median of its absolute value and divide it by the Gaussian noise correction factor of 0.6745. Set the energy entropy judgment threshold to 0.6; when the calculated energy entropy... When the threshold is greater than 0.6, the current one-dimensional time-domain signal segment is judged to contain rich information, belonging to a complex pathological signal or a region with strong noise. In this case, a softer threshold is used to preserve details. Specifically, the general denoising threshold is multiplied by an adjustment factor of 0.6 to obtain the final threshold. When the calculated energy entropy... When the threshold is ≤0.6, the current one-dimensional time-domain signal segment is determined to be sparse and belongs to a background silence period. At this time, a harder threshold is used to thoroughly filter out noise. Specifically, the general denoising threshold is multiplied by an adjustment factor of 1.0 to obtain the final threshold; thus, while removing environmental noise, the faint heart murmur characteristics are preserved to the greatest extent. For example... Figure 4 As shown, the high-frequency random noise superimposed in the original signal is effectively filtered out, and the baseline drift is corrected. At the same time, the morphological features of the first and second heart sounds and the details of faint pathological murmurs in the heart sound signal are completely preserved, verifying the effectiveness of the energy entropy-based adaptive threshold selection strategy in improving the signal-to-noise ratio and preserving key features.

[0060] The heart sound signal classification method in step 4 includes:

[0061] Step 4-1: A convolutional neural network module is used, which employs a multi-scale parallel convolutional structure. This structure includes at least two parallel convolutional branches: the first branch uses a small-sized convolutional kernel (3×3) with a stride of 1 to capture the transient impact features of the first and second heart sounds; the second branch uses a large-sized convolutional kernel or dilated convolution (11×11) with a stride of 1 to capture the long-term persistent texture features of heart murmurs. The output features of the two branches are concatenated and fused along the channel dimension, and then input into the subsequent pooling layer. The number of output channels of the convolutional layer is set to 32, and the pooling layer uses max pooling with a kernel size of 2×2. This structure solves the problem that a single-sized convolutional kernel cannot simultaneously capture both transient and persistent features of heart sounds.

[0062] Step 4-2: The feature sequence is processed in depth through a bidirectional LSTM recurrent neural network module. The module captures the long-range temporal dependencies and dynamic evolution patterns between events in the heart sound cycle through its internal gating mechanism, thereby modeling the overall rhythm and pathological pattern of the heart sound signal. The module contains two stacked bidirectional long short-term memory networks, each containing 128 hidden units.

[0063] Step 4-3: Using the temporal attention mechanism module, the similarity between the hidden state at each time step output by the recurrent neural network module and the trainable context vector is calculated, and then transformed into a normalized probability weight distribution using a normalized exponential function. This weight value dynamically reflects the contribution of the current time step feature to the final classification result. During training, the model automatically learns to assign high weights to effective signal segments containing the first / second heart sounds and pathological murmurs, while assigning low weights to background noise or silent periods. The high weight values ​​typically range from 0.1 to 0.5, making the feature vector at that time step play a dominant role in feature fusion; the low weight values ​​typically range from 0.0001 to 0.02. Finally, by weighted summing of the features from all time steps, a global context vector focusing on key pathological features is generated, thereby effectively suppressing irrelevant interference information.

[0064] Step 4-4: Through a fully connected classifier, which contains 64 neurons, the heart sound signal is output as a precise probability distribution of "normal" or "abnormal" through nonlinear transformation and feature integration, thus completing the automatic classification with a classification accuracy of 90%.

[0065] Furthermore, in one implementation, step 2 includes:

[0066] A handheld piezoelectric film vibration sensor probe is stably pressed against a designated auscultation area on the chest to acquire raw analog heart sound signals. This signal is then processed by high-speed, low-latency transmission hardware: First, a low-noise signal amplification circuit built with the AD620 instrumentation amplifier chip initially amplifies the weak raw heart sound signal, achieving a magnification factor of 200 times, significantly improving its signal-to-noise ratio and amplitude, laying the foundation for subsequent processing. Next, a high-precision analog-to-digital converter circuit using an independent ADC0832 chip samples and quantizes the amplified analog signal at a sampling frequency of 2500 Hz and a quantization level of 256, converting it into a high-fidelity digital signal. This ensures the integrity of the heart sound waveform characteristics and facilitates subsequent digital signal processing. Finally, the processed digital heart sound signal is transmitted via a StarFlash communication module based on the WS63 chip. With its high bandwidth, low power consumption, and strong anti-interference capabilities, this star-flash module provides a high-speed, extremely low-latency, and stable wireless data transmission link for the core. To adapt to the characteristics of large instantaneous data volume and high latency requirements of heart sound signals, a compression mechanism combining predictive coding and Huffman coding, as well as a transmission mechanism of gateway polling and packet transmission, have been added. The number of bytes after compression is 20% of that of the uncompressed module, effectively meeting the real-time and reliability requirements of heart sound signal transmission.

[0067] Furthermore, in one implementation, step 3 includes:

[0068] At the receiving end, a series of key preprocessing steps are performed on the transmitted digital heart sound signal to improve signal quality and construct feature representations suitable for deep learning model analysis. First, a combination of bandpass filtering and an adaptive wavelet threshold modulation method based on multi-scale energy entropy is used to denoise the original signal. The bandpass filter employs a finite impulse response digital filter with strictly linear phase characteristics, and the passband frequency range is 20 Hz to 600 Hz. The adaptive wavelet threshold modulation method based on multi-scale energy entropy uses the Dobermann wavelet basis function, with a decomposition level of 5. The bandpass filtering effectively filters out environmental and physiological interference outside the main heart sound frequency band. The wavelet transform adopts an adaptive threshold selection strategy based on signal energy entropy. First, the threshold value of the wavelet decomposition is calculated... Normalized energy entropy of layer detail coefficients The calculation formula is:

[0069]

[0070] in For the first The energy percentage of each coefficient This represents the total length of the coefficients in this layer. The normalized energy entropy. The value range is [0,1]; the general denoising threshold is... , in This represents the total number of sampling points for the current one-dimensional time-domain heart sound signal segment. This represents the standard deviation estimate of the background noise, calculated by selecting the high-frequency detail coefficient vector of the first layer of wavelet decomposition. Calculate the median of its absolute value and divide it by the Gaussian noise correction factor of 0.6745. Set the energy entropy judgment threshold to 0.6; when the calculated energy entropy... When the threshold is greater than 0.6, the current one-dimensional time-domain signal segment is judged to contain rich information, belonging to a complex pathological signal or a region with strong noise. In this case, a softer threshold is used to preserve details. Specifically, the general denoising threshold is multiplied by an adjustment factor of 0.6 to obtain the final threshold. When the calculated energy entropy... When the threshold is ≤0.6, the current one-dimensional time-domain signal segment is determined to be sparse and belongs to a background silence period. At this time, a harder threshold is used to completely filter out noise. Specifically, the general denoising threshold is multiplied by an adjustment factor of 1.0 to obtain the final threshold. This removes environmental noise while preserving the faint heart murmur characteristics to the greatest extent. Subsequently, the peak value of the denoised signal is normalized to uniformly scale its amplitude to the closed interval [-1,1] to eliminate amplitude deviations caused by differences in acquisition conditions and provide a stable data foundation for model training. Furthermore, to expand the dataset and enhance the robustness of the model, block segmentation and sliding window techniques are used to augment the normalized continuous heart sound signal, generating a large number of one-dimensional time-domain signal segments with temporal context information. Specifically, the time length of each one-dimensional time-domain signal segment is set to 5 seconds, and the step size of the sliding window is 2.5 seconds. Finally, feature extraction was performed using a time-frequency complementary feature extraction method based on a narrow-wide-window dual-channel parallel processing method. The preprocessed one-dimensional time-domain signal segments were input into the dual-channel parallel processing stream. In the first channel, a narrow time window was used to perform a short-time Fourier transform on the signal. The window length of the narrow time window was set to a Hanning window with 64 sampling points, the frame shift was set to 32 sampling points, and the number of fast Fourier transform points was set to 1024 to capture the transient temporal localization features of the first and second heart sounds, generating a high-temporal-resolution spectrogram with an effective temporal resolution of 25.6 milliseconds. In the second channel, a wide time window was used to perform a short-time Fourier transform on the same signal. The window length of the wide time window was set to 2... A Hanning window with 56 sampling points, a frame shift of 32 sampling points, and a Fast Fourier Transform (FFT) of 1024 points are used to capture the fine frequency-domain texture features of cardiac pathological murmurs, generating a high-frequency resolution spectrogram with a physical frequency resolution of 9.7 Hz. The resulting high-temporal resolution spectrogram and high-frequency resolution spectrogram are then mapped using a Mel filter bank and fused along the channel dimension to form a composite feature map containing multi-scale time-frequency information. The Mel filter bank has 128 filters, generating a feature map with a height of 128 pixels. The energy value of the composite feature map is converted to a logarithmic scale to obtain the features for the final input model. This invention's feature extraction technique differs from existing methods that use a single window length. It is a time-frequency complementary feature extraction method based on a narrow-wide-window dual-channel parallel architecture, effectively solving the trade-off between time-frequency resolution. This allows the backend neural network to simultaneously acquire accurate cardiac cycle localization information and fine pathological murmur spectrum features, significantly improving the classification accuracy of faint murmurs.

[0071] Furthermore, in one implementation, step 4 includes:

[0072] The preprocessed two-dimensional Mel-frequency feature map is fed into a carefully designed deep learning classification model for accurate classification. This process begins with a convolutional neural network module employing a multi-scale parallel convolutional structure. This structure contains at least two parallel convolutional branches: the first branch uses a small-sized convolutional kernel (3×3) with a stride of 1 to capture the transient impact features of the first and second heart sounds; the second branch uses a large-sized convolutional kernel or dilated convolution (11×11) with a stride of 1 to capture the long-term persistent texture features of heart murmurs. The output features of the two branches are concatenated and fused along the channel dimension before being input into subsequent pooling layers. The number of output channels in the convolutional layers is set to 32, and max pooling with a kernel size of 2×2 is used. This structure solves the problem that a single-sized convolutional kernel cannot simultaneously capture both transient and persistent features of heart sounds. Subsequently, a bidirectional LSTM recurrent neural network module performs deep processing on the feature sequence. Through its internal gating mechanism, it captures the long-range temporal dependencies and dynamic evolution patterns between events in the heart sound cycle, achieving modeling of the overall rhythm and pathological patterns of the heart sound signal. This module contains two stacked bidirectional long short-term memory networks, each with 128 hidden units. To further focus on key temporal segments most relevant to pathological diagnosis, a temporal attention mechanism module is introduced. This module calculates the similarity between the hidden state at each time step output by the recurrent neural network module and the trainable context vector, and transforms it into a normalized probability weight distribution using a normalized exponential function. This weight value dynamically reflects the contribution of the current time step feature to the final classification result. During training, the model automatically learns to assign high weights to effective signal segments containing the first / second heart sound and pathological murmurs, while assigning low weights to background noise or silent periods. The high weight values ​​typically range from 0.1 to 0.5, making the feature vector at that time step play a dominant role in feature fusion; the low weight values ​​typically range from 0.0001 to 0.02. Finally, by weighted summing of features from all time steps, a global context vector focusing on key pathological features is generated, thereby effectively suppressing irrelevant interference information. This context vector is then fed into a fully connected classifier, which contains 64 neurons. Through nonlinear transformation and feature integration, it outputs a precise probability distribution of whether the heart sound signal is "normal" or "abnormal," achieving automated classification with an accuracy of 90%.

[0073] This invention provides a method and approach for the acquisition and automatic classification of heart sound signals based on deep learning. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment of the invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for acquiring and automatically classifying heart sound signals based on deep learning, characterized in that, Including the following step: Step 1: Use a wireless digital stethoscope device to collect data from the target area, and collect heart sound signals through a vibration sensor probe. Step 2: Repeat the acquisition process described in Step 1 continuously to obtain the original analog signal, and process and transmit the original analog signal to obtain a high-amplitude digital signal; Step 3: Preprocess the received digital heart sound signal, and extract features from the preprocessed signal based on a multi-resolution fusion feature extraction strategy. Step 4: Input the extracted features into a pre-trained deep learning classification model to obtain the classification result of the heart sound signal.

2. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 1, characterized in that, In step 1, the heart sound signals are collected in the four main auscultation areas of the human body, including the aortic valve area, pulmonary valve area, tricuspid valve area, and mitral valve area.

3. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 1, characterized in that, The method for collecting heart sound signals described in step 2 is to press the probe of a handheld wireless stethoscope onto the auscultation area to collect the signal.

4. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 1, characterized in that, Step 2, the processing method, includes: Step 2-1: The low-noise signal amplification circuit initially amplifies the weak original heart sound signal, improving its signal-to-noise ratio and amplitude. Step 2-2: The high-precision analog-to-digital converter samples and quantizes the amplified analog signal, converting it into a high-fidelity digital signal to ensure the integrity of the heart sound waveform characteristics and facilitate subsequent digital signal processing. Steps 2-3: The processed digital heart sound signal is transmitted through the StarFlash communication module.

5. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 4, characterized in that, The StarScan communication module includes a compression mechanism combining predictive coding and Huffman coding, as well as a transmission mechanism of gateway polling and packet-based transmission.

6. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 1, characterized in that, The heart sound signal preprocessing method in step 3 includes: Step 3-1-1: The original signal is denoised by combining bandpass filtering with an adaptive wavelet threshold modulation method based on multi-scale energy entropy metric. Step 3-1-2: Perform peak normalization on the denoised signal to uniformly scale its amplitude to a predetermined range in order to eliminate amplitude deviation caused by differences in acquisition conditions. Step 3-1-3: Use block segmentation and sliding window to perform data augmentation on the normalized continuous heart sound signal to generate a large number of one-dimensional time-domain signal segments with temporal context information.

7. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 6, characterized in that, The feature extraction in step 3 specifically involves: Step 3-2-1: Input the preprocessed one-dimensional time-domain signal segments into the dual-channel parallel processing stream respectively; In the first channel, a narrow time window is used to perform a short-time Fourier transform on the signal to capture the transient temporal localization features of the first and second heart sounds and generate a high temporal resolution spectrogram. In the second channel, a short-time Fourier transform is performed on the same signal using a wide time window to capture the fine frequency domain texture features of cardiac pathological murmurs and generate a high-frequency resolution spectrogram. Step 3-2-2: Map the obtained high temporal resolution spectrogram and the high frequency resolution spectrogram respectively through the Mel filter bank, and then splice and fuse them in the channel dimension to form a composite feature map containing multi-scale time and frequency information. Step 3-2-3: Convert the energy value of the composite feature map to a logarithmic scale to obtain the features of the final input model.

8. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 6, characterized in that, The original signal described in step 3-1-1 is denoised by using an adaptive threshold selection strategy based on signal energy entropy. The energy entropy of the wavelet coefficients of each decomposition layer of the heart sound signal is calculated, and the threshold function is dynamically adjusted according to the magnitude of the energy entropy. When the energy entropy is high, it is determined to be a complex pathological signal or strong noise, and a softer threshold is used to preserve details. When the energy entropy is low, it is determined to be a background silent period, and a harder threshold is used to completely filter out noise.

9. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 1, characterized in that, The heart sound signal classification method in step 4 includes: Step 4-1: The convolutional neural network module employs a multi-scale parallel convolutional structure. This structure includes at least two parallel convolutional branches: the first branch uses a small-sized convolutional kernel to capture the transient impact features of the first and second heart sounds; the second branch uses a large-sized convolutional kernel or dilated convolution to capture the long-term persistent texture features of the heart murmur. The output features of the two branches are concatenated and fused in the channel dimension to form a feature sequence, which is then input into the subsequent pooling layer. Step 4-2: The feature sequence is processed in depth through the recurrent neural network module. The long-range temporal dependencies and dynamic evolution patterns between events in the heart sound cycle are captured through its internal gating mechanism, so as to model the overall rhythm and pathological pattern of the heart sound signal. Step 4-3: Through the temporal attention mechanism module, the features of all time steps output by the recurrent neural network are adaptively weighted, amplifying the contribution of important moments while suppressing irrelevant information, thereby generating a context vector focused on discriminative information as the final classification basis; Step 4-4: Through a fully connected classifier, the probability distribution of heart sound signals is output via nonlinear transformation and feature integration to complete automated classification.

10. The method for acquiring and automatically classifying heart sound signals based on deep learning according to claim 9, characterized in that, The categories for classification include: normal and abnormal.

Images