A child congenital heart disease intelligent identification method based on heart sound double-flow feature cooperation

Abstract

The application discloses a kind of based on heart sound double-flow feature coordination's intelligent identification method for children's congenital heart disease, belong to biomedical signal processing technical field.The method obtains child heart sound audio signal and carries out pretreatment, respectively extracts heart sound time-frequency energy feature and heart sound gamma pass feature, is coupled and is constructed double-flow multidimensional heart sound feature map after time domain resolution synchronous processing;Characteristic map is input hierarchical time-frequency analysis network, with noise suppression unit filters out environmental disturbance, with key feature focusing unit highlights abnormal heart sound feature;Combined with difficult and easy sample balance strategy training model, output classification prediction probability.The application is improved through double-flow feature complementation and multi-stage feature enhancement, heart sound feature expression and anti-interference ability, help to improve the stability and reliability of the intelligent identification of children's heart sound anomaly.

Description

Technical Field

[0001] This invention belongs to the field of biomedical signal processing technology, specifically relating to an intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features. Background Technology

[0002] Congenital heart disease is the most common congenital malformation in childhood. If it is detected early and intervened in a timely manner, most children can receive effective treatment. Currently, cardiac auscultation based on heart sound signals is the main method for early screening of congenital heart disease in children. However, heart sound collection is easily affected by environmental noise, clothing friction, and the child's crying. Manual auscultation relies on the doctor's experience, and faint pathological murmurs are prone to being missed or misdiagnosed.

[0003] Most existing automatic heart sound recognition methods only employ a single type of acoustic feature, failing to consider both the time-frequency energy distribution and auditory detail of the heart sound signal. Furthermore, most methods use traditional network structures, lacking feature enhancement mechanisms to address noise interference and weak abnormal signals, resulting in insufficient anti-interference capabilities in complex acquisition environments and low accuracy in recognizing weak pathological heart sounds. In addition, congenital heart disease heart sound data generally suffers from an imbalanced positive and negative sample problem, which current training strategies have not specifically optimized for, further affecting recognition stability and sensitivity.

[0004] Therefore, how to fully extract complementary features of heart sounds in noisy environments, enhance the expression of weak abnormal signals, and adapt to unbalanced samples to achieve stable recognition is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This invention provides the following technical solution:

[0006] This invention provides an intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features. The method includes the following steps:

[0007] S1: Acquire the raw heart sound audio signal of the child to be tested;

[0008] S2: Perform signal denoising and standardization processing on the original heart sound audio signal to obtain the preprocessed heart sound signal;

[0009] S3: Construct a heterogeneous heart sound feature space that complements time-frequency and auditory features: Extract the heart sound time-frequency energy features and heart sound gamma-pass features of the preprocessed heart sound signals respectively, and eliminate the time dimension difference between the two through synchronous processing with time-domain resolution, perform deep coupling of heterogeneous features, and construct a dual-stream multidimensional heart sound feature map;

[0010] S4: The dual-stream multidimensional heart sound feature map is input into a hierarchical time-frequency analysis network integrating a multi-stage feature enhancement unit. The hierarchical time-frequency analysis network includes an input noise suppression unit, a block feature mapping unit, a hierarchical feature encoder, and a classification decision unit. The input noise suppression unit filters out environmental noise, the hierarchical feature encoder captures temporal dependencies, and the key feature focusing unit adaptively captures weak abnormal heart sound components.

[0011] S5: Output the abnormal classification prediction probability of the heart sound signal to be tested.

[0012] Furthermore, S2 includes the following sub-steps:

[0013] S21: Resample the original heart sound audio signal to a preset frequency to unify the time resolution;

[0014] S22: Construct a bandpass filter to filter the resampled signal, retain the preset effective frequency band containing the main energy of the heart sound, and filter out out-of-band noise and baseline drift;

[0015] S23: Employ a spike removal algorithm to identify and eliminate high-amplitude pulse artifacts in the filtered signal;

[0016] S24: Obtain the extreme value of the current heart sound segment amplitude and use a normalization algorithm to map the signal amplitude to the standard range.

[0017] Furthermore, S3 includes the following sub-steps:

[0018] S31: Construct a time-frequency energy feature map of the heart sound reflecting the energy distribution for the preprocessed heart sound signal, and process the signal using short-time Fourier transform and log-Mel filter bank to generate log-Mel spectral features;

[0019] S32: Construct a heart sound gamma-ton feature map representing auditory characteristics for the preprocessed heart sound signal, and use a gamma-ton filter bank to filter and perform discrete cosine transform processing on the signal to generate gamma-ton frequency cepstral coefficient features.

[0020] S33: To address the difference in sampling rates between the two types of features on the time axis, perform temporal resolution synchronization processing on the heart sound gamma-pass feature map, mapping its time dimension to the time scale of the heart sound time-frequency energy feature map to achieve frame-level alignment.

[0021] S34: Perform deep heterogeneous feature coupling between the aligned heart sound gamma-pass feature map and the heart sound time-frequency energy feature map to construct a dual-stream multidimensional heart sound feature map containing time-frequency details and auditory characteristics.

[0022] Furthermore, the hierarchical time-frequency analysis network integrating multi-stage feature enhancement units includes the following modules cascaded in sequence:

[0023] Input noise suppression unit: placed at the front end of the network, used to calculate the importance weights of features, suppress the response of background noise channels, and preserve the feature representation of effective heart sound signals;

[0024] Block feature mapping unit: used to divide a two-dimensional feature map into a sequence of non-overlapping feature blocks and linearly map it to a high-dimensional feature space;

[0025] Hierarchical feature encoder: It contains multiple cascaded encoding stages for extracting temporal features of heart sound signals layer by layer; key feature focusing units are embedded between the encoding stages.

[0026] Classification Decision Unit: Used to aggregate deep features and map them to the classification prediction probability that the sample to be tested belongs to the category of congenital heart disease abnormality.

[0027] Furthermore, during the training phase of the hierarchical time-frequency analysis network, to address the imbalance in the distribution of positive and negative samples in the heart sound anomaly recognition scenario, an adaptive loss calculation mechanism based on the difficulty of sample classification is used to optimize the model:

[0028] The weight of easily classified samples in gradient updates is dynamically reduced, while the weight of difficult-to-classify samples containing weak abnormal heart sound features is increased, so that the optimization focus of the model is on low-confidence samples.

[0029] Furthermore, in step S5, the specific decision-making and execution process for predicting the abnormal classification probability of the output heart sound signal to be tested includes:

[0030] A threshold determination mechanism is adopted to compare the classification prediction probability finally output by the hierarchical time-frequency analysis network with a preset threshold. If the classification prediction probability exceeds the preset threshold, the current heart sound to be tested is determined to be an abnormal heart sound category, and the corresponding prompt information is output.

[0031] The beneficial effects of this invention are:

[0032] (1) It enriches the feature expression of heart sound signals. By constructing a dual-stream feature space that complements the time-frequency distribution and auditory characteristics, weak pathological heart sound features can be effectively extracted, increasing the diversity and completeness of heart sound features;

[0033] (2) Enhanced resistance to noise interference. Integrating multi-stage feature enhancement units into the feature extraction network can improve the model recognition performance under low signal-to-noise ratio conditions by suppressing various types of interference noise responses and preserving abnormal heart sound feature responses;

[0034] (3) The model performance under imbalanced sample distribution was optimized. An adaptive loss calculation mechanism based on classification difficulty was adopted to dynamically adjust the weights of different samples during training, which can improve the model's detection effect on difficult-to-classify abnormal samples;

[0035] (4) Improved objectivity in identifying abnormal heart sounds. By outputting quantified classification prediction probabilities, the reliance on subjective human auscultation experience in heart sound identification can be reduced, providing objective data reference for heart sound detection. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the overall process of the intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features provided in an embodiment of the present invention;

[0037] Figure 2 A schematic diagram illustrating the principle of synchronous processing of the heart sound dual-stream feature space construction and temporal resolution, which complements the time-frequency distribution and auditory characteristics, provided in an embodiment of the present invention.

[0038] Figure 3 A schematic diagram of the overall architecture of a hierarchical time-frequency analysis network integrating multi-stage feature enhancement units provided in an embodiment of the present invention;

[0039] Figure 4 This is a schematic diagram of the input noise suppression unit and the key feature focusing unit in an embodiment of the present invention. Detailed Implementation

[0040] The technical solution of the present invention will be further described below with reference to the accompanying drawings, but the scope of protection of the present invention is not limited to the following description.

[0041] like Figure 1 As shown, a method for intelligent identification of congenital heart disease in children based on the synergy of dual-flow heart sound features includes the following steps.

[0042] Step S1: Acquire the raw heart sound audio signal of the child to be tested. Specifically, collect the child's heart sound data in a clinical or home environment using a digital stethoscope or medical acquisition terminal. Considering the common environmental noise (such as talking) and interference noise from the examinee (such as crying, clothing rubbing, and physical agitation) in pediatric examination scenarios, it is necessary to acquire raw acoustic data containing the complete cardiac cycle to provide a data foundation for subsequent signal purification and feature extraction.

[0043] Step S2: Perform signal denoising and standardization to generate preprocessed heart sound signals. To extract effective heart sound components from background noise, this invention employs a multi-step signal preprocessing workflow, specifically including the following functional steps.

[0044] Time-domain resolution standardization: The original signal is resampled to 2000Hz. This frequency setting aims to create a unified time reference that covers the high-frequency details of heart sounds while avoiding computational redundancy caused by excessively high sampling rates.

[0045] Bandpass filtering: Construct a fourth-order Butterworth bandpass filter to retain the frequency band from 25Hz to 400Hz.

[0046] Peak removal processing: To address interference caused by body movement or friction, a Schmidt peak removal algorithm is used to identify and remove abnormal high-amplitude data segments by calculating the statistical threshold of the signal.

[0047] Global amplitude dynamic normalization: To eliminate gain differences caused by different acquisition devices, a linear mapping algorithm is used to uniformly constrain the signal amplitude to the standard range of [-1, 1]. This step is achieved through the following amplitude normalization formula:

[0048] ;

[0049] In the formula, Characterizing the heart sound signal after filtering and artifact removal. This represents the peak absolute amplitude of the sequence within the current time window. This is the final output standardized preprocessed signal.

[0050] Step S3: As Figure 2 As shown, a dual-stream feature space for heart sounds, which complements the time-frequency distribution and auditory characteristics, is constructed to generate a dual-stream multidimensional heart sound feature map. To balance the high-frequency energy distribution and low-frequency detail information of the heart sound signal, this invention adopts a dual-stream feature complementarity strategy, extracting in parallel the log-Mel features reflecting the heart sound energy distribution and the gamma-ton features capturing low-frequency auditory characteristics. The specific implementation process is as follows.

[0051] Heart sound time-frequency energy feature extraction: The log-Mel spectrum is extracted as the first feature branch to characterize the energy distribution of the heart sound signal in the time-frequency domain. To conform to the nonlinear nature of human ear frequency perception (i.e., greater sensitivity to low frequencies), linear physical frequencies are used... Mapping to Mel scale The conversion formula is:

[0052] ;

[0053] In the formula, Represents the linear physical frequency of the input signal (unit: Hz). This represents the perceived Mel frequency after mapping. The constants 2595 and 700 are frequency conversion coefficients measured in auditory psychology. By projecting the spectral energy onto the Mel scale through this mapping and taking the logarithm, a time-frequency energy characteristic map of the heart sound is obtained.

[0054] Heart sound gamma-tone feature extraction: Gamma-tone frequency cepstral coefficients (GFCCs) are extracted as the second feature branch. GFCCs are calculated based on a gamma-tone filter bank, which has excellent time-frequency joint resolution and is particularly suitable for processing non-stationary heart sound signals. Its time-domain impulse response function... Defined as:

[0055] ;

[0056] In the formula, Indicates time, The center frequency of the filter, This represents the amplitude coefficient, used to normalize the filter gain to ensure signal energy stability; is the base of the natural logarithm; The bandwidth attenuation factor determines the bandwidth of the filter; The filter order is taken in this embodiment, based on the auditory physiology model. The shape of the auditory filter conforms to the basilar membrane of the human ear; This represents phase information. The heart sound gamma-ton features generated based on this formula have higher resolution in the low-frequency region than the log-Melogram.

[0057] Temporal resolution synchronization and deep coupling: During the construction of the dual-stream feature space, the extraction step size settings for log-Mel and gamma-ton frequency cepstral coefficients may differ, leading to a difference in the number of sampling points on the time axis (i.e., frame rate mismatch). To address this, this invention employs temporal resolution synchronization processing technology (specifically implemented as a linear interpolation algorithm) to map the temporal dimension of the heart sound gamma-ton features to the temporal scale of the heart sound time-frequency energy features, achieving frame-level alignment between the two. For the target time... The characteristic response value after synchronous mapping The calculation is based on the following linear interpolation model:

[0058] ;

[0059] In the formula, This indicates the target time on the time axis of the heart sound time-frequency energy characteristic map for synchronization; and Representing the time immediately adjacent to the target in the original heart sound gamma-ton feature space respectively The time coordinates of the previous and next frames; and The corresponding known feature vectors are represented; the difference term in the formula represents the gradient of local feature changes, and the fractional term represents the weighting coefficients based on time distance. This model enables the mapping of heterogeneous heart sound gamma-pass features to a unified temporal resolution benchmark.

[0060] Finally, deep coupling of heterogeneous features is performed on the two types of features after synchronization. Specifically, feature splicing operations can be performed in the channel dimension to generate a dual-stream multidimensional heart sound feature map that includes the time-frequency energy distribution of the heart sound signal and the low-frequency auditory characteristics.

[0061] Step S4: Conventional deep learning models are easily affected by interference when processing weak signals against a noisy background. To address this issue, this invention constructs a hierarchical time-frequency analysis network integrating multi-stage feature enhancement units. For example... Figure 3 The diagram shown is a schematic of the overall architecture of the intelligent recognition network constructed in this embodiment. The working methods of each module are as follows.

[0062] Input noise suppression unit: such as Figure 3 As shown, this unit is placed at the very front of the network, before the first encoding stage. This unit uses the same underlying attention architecture as the subsequent key feature focusing unit. It utilizes a channel attention mechanism to perform global average pooling and global max pooling on the input two-stream feature maps, and combines this with a shared multilayer perceptron to perform weight calculations, adaptively adjusting the channel weights of the input feature maps.

[0063] Block Feature Mapping Unit: In this embodiment, the unit performs block processing on the input feature map using non-overlapping convolution. The convolution kernel size is matched with the stride. After dividing the two-dimensional feature map into a sequence of non-overlapping feature blocks, it is mapped to a preset high-dimensional vector space through linear projection to generate an initial one-dimensional feature sequence. This completes the dimensional transformation from two-dimensional time-frequency features to one-dimensional sequence features, adapting to the input requirements of subsequent hierarchical coding structures.

[0064] The hierarchical feature encoder comprises multiple cascaded feature encoding stages, namely the first, second, third, and fourth encoding stages. Each encoding stage consists of several stacked, repetitive feature transform blocks. In this embodiment, the number of transform blocks for the four stages is configured as 2, 2, 6, and 2, respectively. Within a single transform block, layer normalization, a moving window multi-head self-attention module, and a two-layer perceptron are integrated sequentially. A feature downsampling fusion layer is placed between adjacent encoding stages. Through spatial size reduction and channel dimension expansion, the temporal features of heart sounds are extracted layer by layer, and a global long-range dependency relationship of the signal is constructed.

[0065] Key feature focusing unit: such as Figure 3 As shown, this unit is positioned between the first and second encoding stages. It employs a serial channel-spatial collaborative attention architecture, sequentially performing channel weight activation and spatial region weight localization. First, global pooling is performed on the feature sequence, and channel weight coefficients are generated using a multilayer perceptron. Then, a spatial local weight map is generated based on convolution operations, and feature processing is completed through element-wise weighting.

[0066] Classification Decision Unit: Finally, the classification decision unit performs global aggregation and dimensionality compression on the deep feature sequence output by the network to generate a unified global feature vector. Then, through the fully connected layer and the normalized exponential activation function, the classification prediction probability of the heart sound abnormality category corresponding to the test sample is obtained.

[0067] Furthermore, regarding the specific implementation architecture of the aforementioned input noise suppression unit and key feature focusing unit, both are composed of channel attention submodules and spatial attention submodules connected in series, forming a collaborative feature weighting structure as a whole.

[0068] Channel attention submodule: such as Figure 4 As shown, this submodule is used to evaluate the importance of feature channels. Specifically, global average pooling and global max pooling operations can be performed in parallel on the input feature map, and then the pooling results are mapped using a shared multilayer perceptron (MLP) to learn the channel weights. Assume the input feature map is... Channel attention weight map The calculation formula is as follows:

[0069] ;

[0070] In the formula, express Activation function and These represent global average pooling and global max pooling, respectively. This process achieves compression and activation of the feature map along the channel dimension.

[0071] Spatial attention submodule: such as Figure 4 As shown, this submodule is used to locate key time-frequency regions. Specifically, it can be used to locate the features after channel weighting. Based on this, average pooling and max pooling are performed along the channel axis to generate spatial statistical descriptors. The resulting statistical feature maps are then concatenated and passed through convolutional layers to generate a two-dimensional spatial weight map. Its mathematical expression is as follows:

[0072] ;

[0073] In the formula, This indicates a convolution operation with a kernel size of 7×7, used to extract local spatial context.

[0074] In the specific forward propagation process, the two sub-modules mentioned above can process the features sequentially through element-wise multiplication. The final output feature... Represented as:

[0075] ;

[0076] In the formula This indicates element-wise multiplication.

[0077] Step S5: Perform classification decision and model training optimization. The classification decision logic adopts a threshold determination mechanism, comparing the classification prediction probability output by the model with a preset classification threshold (e.g., 0.5). If the probability exceeds the threshold, it is determined to be a heart sound abnormality category and a prompt message is output.

[0078] To address the imbalance in the distribution of positive and negative samples in pediatric examination scenarios, an adaptive loss calculation mechanism based on the difficulty of sample classification is employed, aiming to dynamically adjust the weight contribution of different samples in model updates.

[0079] As a preferred implementation, this embodiment can use a focal loss function to implement the above mechanism, the mathematical expression of which is:

[0080] ;

[0081] In the formula: This represents the model's predicted probability of the true class. This represents the category balance factor, used to adjust the basic weight ratio of positive and negative samples; This represents the focus modulation index (preferably set to 2 in this embodiment); through this function... As a weight adjuster, the term can automatically reduce the weight of easily classified samples, forcing the model to focus its optimization efforts on low-confidence, difficult-to-classify samples (such as faint abnormal heart sounds), thus completing the gradient update of the model parameters.

Claims

1. A method for intelligent identification of congenital heart disease in children based on the synergy of dual-flow heart sound features, characterized in that, The method includes: S1: Acquire the raw heart sound audio signal of the child to be tested; S2: Perform signal denoising and standardization processing on the original heart sound audio signal to obtain the preprocessed heart sound signal; S3: Construct a heterogeneous heart sound feature space that is complementary to time-frequency and auditory features, extract the heart sound time-frequency energy features and heart sound gamma-pass features of the preprocessed heart sound signal respectively, and eliminate the time dimension difference between the two through synchronous processing of time-domain resolution, perform deep coupling of heterogeneous features, and construct a dual-stream multidimensional heart sound feature map. S4: The dual-stream multidimensional heart sound feature map is input into a hierarchical time-frequency analysis network integrating multi-stage feature enhancement units. This network includes a cascaded input noise suppression unit, a block feature mapping unit, a hierarchical feature encoder, and a classification decision unit. The input noise suppression unit, located at the front of the network, calculates the importance weights of the features, suppresses the response of background noise channels, and highlights the feature representation of effective heart sound signals. The block feature mapping unit divides the two-dimensional feature map into a non-overlapping sequence of feature blocks and linearly maps it to a high-dimensional feature space. The hierarchical feature encoder contains multiple cascaded encoding stages for extracting temporal features of the heart sound signal layer by layer, with key feature focusing units embedded between the encoding stages. The classification decision unit aggregates deep features and maps them to the classification prediction probability of the heart sound abnormality category of the test sample. S5: Output the abnormal classification prediction probability of the heart sound signal to be tested.

2. The intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features according to claim 1, characterized in that, The signal denoising and standardization process specifically includes the following steps: S21: Resample the original heart sound audio signal to a preset frequency to unify the time resolution; S22: Construct a bandpass filter to filter the resampled signal, retain the preset effective frequency band containing the main energy of the heart sound, and filter out out-of-band noise and baseline drift; S23: Employ a spike removal algorithm to identify and eliminate high-amplitude pulse artifacts in the filtered signal; S24: Obtain the extreme value of the current heart sound segment amplitude and use a normalization algorithm to map the signal amplitude to the standard range.

3. The intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features according to claim 1, characterized in that, The construction of the heterogeneous heart sound feature space that complements time-frequency and auditory perception specifically includes the following sub-steps: S31: Construct a time-frequency energy feature map of the heart sound reflecting the energy distribution for the preprocessed heart sound signal, and process the signal using short-time Fourier transform and log-Mel filter bank to generate log-Mel spectral features; S32: Construct a heart sound gamma-ton feature map representing auditory characteristics for the preprocessed heart sound signal, and use a gamma-ton filter bank to filter and perform discrete cosine transform processing on the signal to generate gamma-ton frequency cepstral coefficient features. S33: To address the difference in sampling rates between the two types of features on the time axis, perform temporal resolution synchronization processing on the heart sound gamma-pass feature map, mapping its time dimension to the time scale of the heart sound time-frequency energy feature map to achieve frame-level alignment. S34: Perform deep heterogeneous feature coupling between the aligned heart sound gamma-pass feature map and the heart sound time-frequency energy feature map to construct a dual-stream multidimensional heart sound feature map containing time-frequency details and auditory characteristics.

4. The intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features according to claim 1, characterized in that, The key feature focusing unit is composed of a channel attention submodule and a spatial attention submodule connected in series: The channel attention submodule evaluates the expression weights of different channels of the feature map for heart sound rhythm components through global pooling and multilayer perceptron network, thereby increasing the feature weights containing effective heart sound energy and suppressing irrelevant background noise channels in the channel dimension. Based on the channel-weighted feature map, the spatial attention submodule generates a two-dimensional spatial weight map through pooling and convolution operations along the channel axis, which adaptively locates and focuses on key local regions containing weak abnormal heart sound features in the time-frequency domain.

5. The intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features according to claim 1, characterized in that, During the training phase of the hierarchical time-frequency analysis network, to address the imbalance in the distribution of positive and negative samples in the heart sound abnormality recognition scenario, an adaptive loss calculation mechanism based on the difficulty of sample classification is used to optimize the model. The weight of easily classified samples in gradient updates is dynamically reduced, while the weight of difficult-to-classify samples containing weak abnormal heart sound features is increased, so that the optimization focus of the model is on low-confidence samples.

6. The intelligent identification method for congenital heart disease in children based on the synergy of dual-flow heart sound features according to claim 1, characterized in that, In step S5, the specific decision-making and execution process for predicting the abnormal classification probability of the output heart sound signal to be tested includes: A threshold determination mechanism is adopted to compare the classification prediction probability finally output by the hierarchical time-frequency analysis network with a preset threshold. If the classification prediction probability exceeds the preset threshold, the current heart sound to be tested is determined to be an abnormal heart sound category, and the corresponding prompt information is output.

Images