Intelligent heart sound analysis system and method

Abstract

The application provides a kind of intelligent heart sound analysis system and method, the system includes heart sound acquisition module, for obtaining original heart sound signal and pre-processing, to obtain heart sound signal;Heart sound adaptive enhancement module is used to obtain quality evaluation feature vector based on heart sound signal, and quality evaluation feature vector is input into evaluation model to obtain the quality level of heart sound signal, and then according to the quality level, the corresponding enhancement operation is triggered, and the clean heart sound signal is obtained;Heart sound analysis module is used to obtain heart sound analysis result and heart sound consensus ratio based on clean heart sound signal, and whether to trigger analysis alarm is judged according to heart sound consensus ratio;Heart sound update module is used to obtain professional quality label when triggering analysis alarm, and the evaluation model is updated according to professional quality label, by professional marking to clean heart sound signal.The application can be deployed at low cost, has environmental self-adaptive ability, significantly improves the practicability and reliability of clinical scene heart sound analysis system.

Description

Technical Field

[0001] This application belongs to the field of heart sound technology and relates to an intelligent heart sound analysis system and method. Background Technology

[0002] Heart sound analysis, a key tool for cardiovascular disease screening, has seen significant advancements in intelligent analysis technology in recent years. Current research primarily focuses on processing and modeling heart sound signals acquired under ideal or low-noise environments. A common approach in existing technologies or traditional methods is to employ a standardized preprocessing procedure after signal acquisition, such as filtering and noise reduction, followed by direct feature extraction and diagnostic analysis using deep learning models. This standardized processing path achieves good recognition results under controlled conditions, laying an important foundation for automated heart sound analysis.

[0003] However, existing technologies have many obvious limitations in practical applications. For example, heart sound signals in real clinical environments are often mixed with various interferences such as breathing sounds, conversations, and instrument noise. Ideal preprocessing procedures are insufficient to effectively handle complex noise backgrounds, leading to decreased model robustness. Furthermore, uniform filtering and denoising operations may filter out certain weak features important for pathological diagnosis, resulting in information loss. Therefore, how to provide a reliable and practical intelligent heart sound analysis system is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This application provides an intelligent heart sound analysis system and method to solve the technical problem of how to provide a reliable and practical intelligent heart sound analysis system.

[0005] The first aspect of this application provides an intelligent heart sound analysis system, the system comprising a heart sound acquisition module, a heart sound adaptive enhancement module, a heart sound analysis module, and a heart sound update module:

[0006] The heart sound acquisition module is used to acquire the raw heart sound signal and perform preprocessing to obtain the heart sound signal;

[0007] The adaptive enhancement module for heart sounds is used to obtain a quality assessment feature vector based on the heart sound signal, input the quality assessment feature vector into the assessment model to obtain the quality level of the heart sound signal, and then trigger a corresponding enhancement operation according to the quality level to obtain a clean heart sound signal.

[0008] The heart sound analysis module is used to obtain the heart sound analysis result and heart sound consensus ratio based on the heart sound clean signal, and to determine whether to trigger an analysis alarm based on the heart sound consensus ratio.

[0009] The heart sound update module is used to professionally label the clean heart sound signal to obtain a professional quality label when the analysis alarm is triggered, and update the evaluation model according to the professional quality label.

[0010] In some implementations of the first aspect, the adaptive heart sound enhancement module includes a feature extraction unit, an evaluation unit, and an adaptive enhancement unit;

[0011] The feature extraction unit is used to extract quality assessment features based on the heart sound signal, so as to obtain the quality assessment feature vector based on the quality assessment features; the quality assessment features include at least one of the following: amplitude kurtosis, amplitude skewness, short-time zero-crossing rate variance, physiological energy ratio, high-frequency energy ratio, spectral flatness, heart sound periodicity, autocorrelation main peak ratio, frequency smoothing envelope variance, Shannon entropy, and modal energy entropy.

[0012] The evaluation unit is used to input the quality evaluation feature vector into the evaluation model, so as to classify and evaluate the quality evaluation feature vector through the evaluation model and obtain the corresponding quality level; the quality level includes high-quality signal, acceptable signal and poor-quality signal;

[0013] The adaptive enhancement unit is used to perform adaptive enhancement processing on the acceptable signal to enhance the acceptable signal into a high-quality processed signal, and together with the high-quality signal, serve as the clean heart sound signal.

[0014] In some implementations of the first aspect, the adaptive enhancement unit is used to perform adaptive continuous wavelet denoising on the acceptable signal and transient artifact suppression on the denoised acceptable signal to enhance the acceptable signal into the high-quality processed signal.

[0015] In some implementations of the first aspect, the adaptive enhancement unit is further configured to discard the acceptable signal along with the inferior signal when the acceptable signal cannot be enhanced into a high-quality processing signal.

[0016] In some implementations of the first aspect, the heart sound analysis module includes a division unit, an analysis unit, and a divergence unit;

[0017] The segmentation unit is used to divide the clean heart sound signal into several sub-segments, each of which includes at least one complete cardiac cycle.

[0018] The analysis unit is used to extract original time-domain features and frequency-domain Mel frequency cepstral coefficients based on the sub-segments, to obtain corresponding weighted fusion features based on the original time-domain features and the frequency-domain Mel frequency cepstral coefficients, and to predict and output sub-analysis results of each sub-segment according to the weighted fusion features, so as to obtain the heart sound analysis results and the heart sound consensus ratio according to each sub-analysis result;

[0019] The divergence unit is used to determine whether the consensus ratio meets the preset divergence trigger boundary conditions. If so, the analysis alarm is triggered and the heart sound clean signal is transmitted to the divergence buffer.

[0020] In some implementations of the first aspect, the analysis unit includes an adaptive feature decomposition mechanism and a deep abstract coding network; wherein the deep abstract coding network includes a noise-aware gating mechanism;

[0021] The adaptive feature decomposition mechanism is used to perform dilated convolution and residual connection processing on the sub-segments to obtain the first original temporal features, and to perform dense connection processing on the sub-segments to obtain the second original temporal features.

[0022] The noise-aware gating mechanism is used to obtain the intrinsic mode components of the sub-segment and obtain gating weights based on the energy entropy of the intrinsic mode components, so as to dynamically obtain gating features based on the gating weights.

[0023] The deep abstract coding network is used to encode the first original temporal features and the second original temporal features, and to perform noise-adaptive feature modulation based on the gated features to obtain deep features.

[0024] In some implementations of the first aspect, the deep abstract coding network further includes a clean adaptive feature fusion mechanism;

[0025] The clean adaptive feature fusion mechanism is used to extract frequency domain Mel frequency cepstral coefficients based on the sub-segments, and dynamically adjust the fusion ratio of the deep features and the frequency domain Mel frequency cepstral coefficients based on a switching adjustment function to obtain the weighted fused features; wherein, the switching adjustment function is obtained based on the output confidence of the evaluation model.

[0026] In some implementations of the first aspect, the heart sound update module is used to retrain the evaluation model based on the professional quality label to obtain new evaluation model parameters, and transmit the new evaluation model parameters to the heart sound adaptive enhancement module so that the evaluation model can be adaptively updated; wherein, the new evaluation model parameters include model support vector parameters, Lagrange multipliers, and bias terms.

[0027] In some implementations of the first aspect, the heart sound acquisition module includes a raw heart sound acquisition unit, a preprocessing unit, and an effective signal detection unit;

[0028] The original heart sound acquisition unit is used to acquire the original heart sound signal;

[0029] The preprocessing unit is used to perform resampling, amplitude normalization and filtering / denoising on the original heart sound signal to obtain a processed heart sound signal.

[0030] The effective signal detection unit is used to perform length detection and existence detection on the heart sound processing signal, and to take the heart sound processing signal whose signal length is greater than the length threshold and whose signal existence index value is greater than the existence threshold as the heart sound signal.

[0031] A first aspect of this application provides an intelligent heart sound analysis method, the method being applied to the intelligent heart sound analysis system as described in the first aspect, the method comprising:

[0032] The raw heart sound signal is acquired and preprocessed to obtain the heart sound signal;

[0033] A quality assessment feature vector is obtained based on the heart sound signal, and the quality assessment feature vector is input into the assessment model to obtain the quality level of the heart sound signal;

[0034] Trigger the corresponding enhancement operation based on the quality level to obtain a clean heart sound signal;

[0035] Based on the clean heart sound signal, obtain the heart sound analysis result and heart sound consensus ratio, and determine whether to trigger an analysis alarm based on the heart sound consensus ratio;

[0036] When the analysis alert is triggered, the quality assessment feature vector is professionally labeled to obtain professional quality labels, and the assessment model is updated based on the professional quality labels.

[0037] As described above, the intelligent heart sound analysis system and method of this application have the following beneficial effects:

[0038] 1. Significantly enhances the quality of heart sound signals: This application achieves adaptive enhancement of heart sound signals based on constructed multidimensional features. In high-noise clinical environments such as real hospitals, it effectively enhances the quality of signals with repair potential in a progressive manner, thereby avoiding the waste of effective medical information, expanding the coverage of heart sound analysis, and greatly improving the practicality and reliability of intelligent heart sound analysis.

[0039] 2. Effectively Overcoming User-Induced Domain Shift and Suppressing Catastrophic Forgetting in Model Evolution: This application proposes a consensus ratio divergence boundary triggering mechanism. This mechanism can accurately screen and intercept controversial samples with domain shift characteristics, and, in conjunction with closed-loop feedback from domain experts, dynamically maintain the statistical distribution balance of data features from both new and old environments. This not only enables the evaluation model to adapt to continuous changes in specific clinical acoustic environments at low cost, but also more effectively prevents the model from "catastrophic forgetting" of historical case features during dynamic update iterations.

[0040] 3. Adaptive avoidance of excessive smoothing of weak pathological features, improving diagnostic accuracy in high-noise environments: This application introduces a noise-sensing gating and clean feature dynamic switching mechanism. Based on the real-time signal-to-noise ratio of the input signal, it adaptively and dynamically adjusts the fusion ratio of the original time-domain features and the frequency-domain Mel-frequency cepstral coefficients. This underlying control logic effectively breaks through the limitations of fixed feature splicing, avoiding the excessive noise reduction and smoothing of weak but crucial pathological noise features as background noise by the model under extreme clinical noise interference. This significantly improves the system's high sensitivity and diagnostic robustness in harsh clinical auscultation environments. Attached Figure Description

[0041] Figure 1 A schematic diagram of the architecture of the intelligent heart sound analysis system described in an embodiment of this application is shown.

[0042] Figure 2 A schematic diagram of the architecture of the heart sound acquisition module described in an embodiment of this application is shown.

[0043] Figure 3 A schematic diagram of the architecture of the adaptive heart sound enhancement module described in an embodiment of this application is shown.

[0044] Figure 4 A schematic diagram of the architecture of the heart sound analysis module described in an embodiment of this application is shown.

[0045] Figure 5 A schematic diagram of the architecture of the analysis unit described in an embodiment of this application is shown.

[0046] Figure 6 A flowchart of the intelligent heart sound analysis system described in an embodiment of this application is shown.

[0047] Figure 7 A schematic diagram of the edge-cloud collaborative architecture of the intelligent heart sound analysis system described in an embodiment of this application is shown.

[0048] Figure 8 A flowchart illustrating an intelligent heart sound analysis method according to an embodiment of this application is shown. Detailed Implementation

[0049] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0050] It should be noted that in the following description, reference is made to the accompanying drawings, which illustrate several embodiments of this application. It should be understood that other embodiments may also be used, and changes in mechanical composition, structure, electrical system, and operation may be made without departing from the spirit and scope of this application. The following detailed description should not be considered limiting, and the scope of the embodiments of this application is defined only by the claims of the published patent. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application. Spatial terms such as “upper,” “lower,” “left,” “right,” “below,” “below,” “lower part,” “above,” “upper part,” etc., may be used herein to illustrate the relationship between one element or feature shown in the figures and another element or feature.

[0051] Furthermore, as used herein, the singular forms “a,” “an,” and “the” are intended to include the plural forms as well, unless the context indicates otherwise. It should be further understood that the terms “comprising,” “including,” indicate the presence of the stated feature, operation, element, component, item, kind, and / or group, but do not preclude the presence, occurrence, or addition of one or more other features, operations, elements, components, items, kinds, and / or groups. The terms “or” and “and / or” as used herein are to be interpreted inclusively, or mean any one or any combination thereof.

[0052] Intelligent heart sound analysis systems, as a key tool in heart sound auscultation, suffer from several significant limitations in practical applications. Firstly, in high-noise clinical environments such as real hospitals, heart sound data is highly susceptible to interference from complex environmental noise, leading to a severe deterioration in signal quality. Current methods often employ simple rejection strategies when processing low-quality signals, lacking mechanisms for progressively enhancing signals with repair potential. This not only wastes valuable medical information but also limits the scope of diagnosis.

[0053] Secondly, existing static models face a severe problem of user-induced domain shift. In real-world scenarios, the acoustic environments of different hospitals, the hardware differences of different acquisition devices, and the operating techniques of different doctors can lead to significant data distribution shifts. Traditional static intelligent models, once deployed, are highly susceptible to degradation in generalization ability and increased false positive rates when faced with such highly inconsistent clinical data. Due to the lack of a low-cost, edge-side continuous learning mechanism, existing systems struggle to dynamically adapt to this domain shift by incorporating closed-loop pathways based on feedback from clinical experts.

[0054] Furthermore, existing diagnostic networks generally suffer from static limitations at the feature fusion level. When processing damaged or pre-denoised heart sound signals, traditional diagnostic models typically employ fixed feature extraction structures, failing to dynamically adjust the dependency ratio between "physical acoustic features (such as Mel-frequency cepstral coefficients, MFCC)" and "deep abstract features" based on the real-time signal-to-noise ratio of the signal. This rigid processing mechanism leads to the model being prone to over-smoothing weak but crucial pathological noises as background noise under extreme noise interference, thus missing important diagnostic clues.

[0055] To at least address the aforementioned technical problems, embodiments of this application provide an intelligent heart sound analysis system and method that can be deployed at low cost, has environmental adaptability, and significantly improves the practicality and reliability of heart sound analysis systems in clinical settings.

[0056] Figure 1 A schematic diagram of the architecture of the intelligent heart sound analysis system described in an embodiment of this application is shown. Figure 1 As shown, the intelligent heart sound analysis system 1 provided in this application embodiment may include a heart sound acquisition module 10, a heart sound adaptive enhancement module 20, a heart sound analysis module 30, and a heart sound update module 40.

[0057] The heart sound acquisition module 10 is used to acquire the original heart sound signal and perform preprocessing to obtain the heart sound signal.

[0058] The adaptive enhancement module 20 is used to obtain a quality assessment feature vector based on the heart sound signal, input the quality assessment feature vector into the assessment model to obtain the quality level of the heart sound signal, and then trigger the corresponding enhancement operation according to the quality level to obtain a clean heart sound signal.

[0059] The heart sound analysis module 30 is used to obtain the heart sound analysis result and the heart sound consensus ratio based on the heart sound clean signal, and to determine whether to trigger an analysis alarm based on the heart sound consensus ratio.

[0060] The heart sound update module 40 is used to professionally label the clean heart sound signal to obtain a professional quality label when the analysis alarm is triggered, and update the evaluation model according to the professional quality label.

[0061] Figure 2 A schematic diagram of the architecture of the heart sound acquisition module described in an embodiment of this application is shown. Figure 2 As shown, the heart sound acquisition module 10 includes a raw heart sound acquisition unit 101, a preprocessing unit 102, and an effective signal detection unit 103.

[0062] The original heart sound acquisition unit 101 is used to acquire original heart sound signals. In some embodiments, the original heart sound acquisition unit 101 continuously acquires original heart sound signals.

[0063] The preprocessing unit 102 is used to perform resampling, amplitude normalization and filtering noise reduction on the original heart sound signal to obtain a processed heart sound signal.

[0064] In some embodiments, the preprocessing unit 102 resamples the acquired raw heart sound signals to 2000 Hz to maintain the core heart sound band information with a standardized sampling rate. Then, the preprocessing unit 102 applies equation (1) to perform amplitude normalization processing on the raw heart sound signals to eliminate sensitivity bias between different devices.

[0065] (1)

[0066] in, S is the amplitude-normalized heart sound signal, S(t) is the signal value at time t, and max(|S|) is the maximum absolute value of the signal.

[0067] Furthermore, the preprocessing unit 102 applies a Butterworth bandpass filter (frequency range set to 25-400 Hz) and applies a specific algorithm to identify and remove spikes generated by sensor movement for filtering and noise reduction, so as to finally obtain the heart sound processing signal.

[0068] It should be noted that this application does not impose any restrictions on specific filtering and denoising algorithms.

[0069] The effective signal detection unit 103 is used to perform length detection and existence detection on the heart sound processing signal, and to take the heart sound processing signal whose signal length is greater than the length threshold and whose signal existence index value is greater than the existence threshold as the heart sound signal.

[0070] In some embodiments, length detection and presence detection are performed before quality assessment of the heart sound processing signal. Only when the length of the heart sound processing signal is greater than a length threshold and the signal presence index value is greater than a presence threshold is the heart sound processing signal accepted as a heart sound signal and allowed to proceed to the subsequent processing steps.

[0071] Length detection refers to determining whether the duration of the heart sound processing signal is greater than a length threshold. Signals shorter than this threshold are marked as invalid and require re-acquisition. In some embodiments, the length threshold is 8 seconds.

[0072] In this method, the periodicity, energy ratio, and bandwidth ratio of the heart sound processing signal are used as presence indicators for presence detection. In some embodiments, if the periodicity of the heart sound processing signal is ≥1.6 and the energy ratio is ≥0.4 and the bandwidth ratio is ≥0.3, or if the periodicity is ≥3 and the energy ratio is ≥0.4 and the bandwidth ratio is ≥0.2, then the presence indicator value is considered to be greater than the presence threshold.

[0073] Figure 3 A schematic diagram of the architecture of the adaptive heart sound enhancement module described in an embodiment of this application is shown. Figure 3 As shown, the adaptive heart sound enhancement module 20 includes a feature extraction unit 201, an evaluation unit 202, and an adaptive enhancement unit 203.

[0074] The feature extraction unit 201 is used to extract quality assessment features based on the heart sound signal, and to obtain the quality assessment feature vector based on the quality assessment features; the quality assessment features include at least one of the following: amplitude kurtosis, amplitude skewness, short-time zero-crossing rate variance, physiological energy ratio, high-frequency energy ratio, spectral flatness, heart sound periodicity, autocorrelation main peak ratio, frequency smoothing envelope variance, Shannon entropy, and modal energy entropy.

[0075] In some embodiments, the feature extraction unit 201 extracts amplitude kurtosis, amplitude skewness, short-time zero-crossing rate variance, physiological energy ratio, high-frequency energy ratio, spectral flatness, heart sound periodicity, autocorrelation main peak ratio, frequency smoothing envelope variance, Shannon entropy, and modal energy entropy, and constructs an 11-dimensional quality assessment feature vector from these 11 features.

[0076] In some embodiments, the feature extraction unit 201 extracts the amplitude kurtosis K according to formula (2). This feature can reflect the tail thickness of the signal amplitude distribution. The kurtosis corresponds to the occurrence of outlier points of large amplitude, and can keenly capture the transient sharp mechanical noise generated by the friction and collision of the stethoscope probe.

[0077] (2)

[0078] In some embodiments, the feature extraction unit 201 extracts the amplitude skewness S according to equation (3), which measures the asymmetry of the amplitude distribution. Positive skewness indicates that most values ​​are below the mean (negative tailing), while negative skewness is the opposite. This index can help determine whether the signal has severe baseline drift or unilateral low-frequency motion artifacts.

[0079] (3)

[0080] in, The first segment of the heart sound signal The amplitude of each sampling point, where N is the total number of sampling points in the signal sequence. The arithmetic mean of the original signal sequence. denoted as the standard deviation of the original signal sequence.

[0081] In some embodiments, the feature extraction unit 201 extracts the short-time zero-crossing rate variance according to equation (4). This feature reflects the degree of fluctuation in the zero-crossing rate between frames. Wideband random background noise (such as white noise in a hospital ward) can fragment the waveform, causing the zero-crossing rate to jump randomly and the variance to increase in each frame. Therefore, this indicator can effectively identify this type of noise.

[0082] (4)

[0083] in, For the first frame after frame processing The short-time zero-crossing rate of the frame signal (the number of times the signal waveform crosses zero within each frame), where M is the total number of signal frames. It is the arithmetic mean of the short-time zero-crossing rates of all frames.

[0084] In some embodiments, the feature extraction unit 201 extracts the physiological energy ratio according to equation (5). This feature directly quantifies the proportion of low-frequency (physiological) energy in the signal, which is equivalent to the signal-to-noise ratio (SNR). When high-frequency ambient noise masks heart sounds, the numerator increases more slowly while the denominator increases, leading to a decrease in the PER, thus identifying the degree of noise masking.

[0085] (5)

[0086] in, The spectrum representing the heart sound signal. The signal frequency.

[0087] In some embodiments, the feature extraction unit 201 extracts the high-frequency energy ratio according to equation (6). This feature measures the proportion of energy in high-frequency components (e.g., >500Hz) of a signal. Sharp human voices or electromagnetic high-frequency interference from medical devices in the environment can significantly boost this indicator, making them more susceptible to precise interception.

[0088] (6)

[0089] in, The power spectral density of the heart sound signal. For the Nyquist frequency, The signal frequency.

[0090] In some embodiments, the feature extraction unit 201 extracts the spectral flatness according to equation (7). This feature can be used to distinguish between irregularly distributed white noise (high flatness) and valid heart sounds with clear formants.

[0091] (7)

[0092] in, Let L be the power of the k-th frequency point in the discrete frequency domain sequence, and L be the total number of discrete frequency points in the effective frequency band.

[0093] In some embodiments, the feature extraction unit 201 extracts the degree of heart sound periodicity according to equation (8). This feature quantifies the rhythmic purity of the cardiac cycle and can serve as a core veto indicator against continuous noise inundation.

[0094] (8)

[0095] in, Let be the autocorrelation function of the envelope sequence of heart sound signals. For time delay, This represents the possible range of cardiac cycles. The autocorrelation value at zero delay is the total energy of the envelope signal itself.

[0096] In some embodiments, the feature extraction unit 201 extracts the autocorrelation ratio according to equation (9). This feature measures the dominance of heart rate rhythm and screens for severe, persistent breathing sound interference.

[0097] (9)

[0098] in, The main peak of the autocorrelation function. It is the largest amplitude value among the side lobes (secondary peaks) on both sides of the main peak.

[0099] In some embodiments, the feature extraction unit 201 extracts the frequency smoothed envelope variance according to equation (10). High-quality heart sound envelopes exhibit distinct periodic peaks and troughs with a large variance; continuous noise flattens the envelope and reduces the variance. Therefore, VFSE can be used to distinguish clean heart sounds from noise-overwhelmed signals.

[0100] (10)

[0101] in, The total number of points in the envelope sequence. The envelope sequence after frequency smoothing. For time indexing, Let be the arithmetic mean of the smooth envelope sequence.

[0102] In some embodiments, the feature extraction unit 201 extracts Shannon entropy according to equation (11). This feature quantifies the disorder of energy distribution and serves as a comprehensive boundary feature for determining low-quality repairable signals.

[0103] (11)

[0104] in, X represents the probability of the normalized amplitude of the signal occurring within each quantization interval, where X is the quantization interval divided by the normalized amplitude of the signal.

[0105] In some embodiments, the feature extraction unit 201 extracts the modal energy entropy according to equation (12). This feature captures the energy dispersion caused by non-stationary time-varying noise. When noise causes energy to be dispersed across multiple eigenmodes, The distribution is more uniform, and the entropy value increases.

[0106] (12)

[0107] in, The first obtained from empirical mode decomposition The index of each intrinsic mode, where K is the total number of intrinsic modes. For the first The proportion of energy of each intrinsic mode to the total energy.

[0108] Continue reading Figure 3 As shown, the evaluation unit 202 is used to input the quality evaluation feature vector into the evaluation model, so as to classify and evaluate the quality evaluation feature vector through the evaluation model and obtain the corresponding quality level; the quality level includes high-quality signal, acceptable signal and poor signal.

[0109] In some embodiments, the evaluation unit 202 inputs an 11-dimensional quality evaluation feature vector into a pre-trained three-class SVM model, and performs classification evaluation on the quality evaluation feature vector through the three-class SVM model, thereby classifying the signal into a high-quality signal, an acceptable signal, and a low-quality signal.

[0110] Continue reading Figure 3 As shown, the adaptive enhancement unit 203 is used to perform adaptive enhancement processing on the acceptable signal to enhance the acceptable signal into a high-quality processed signal, and together with the high-quality signal, serve as the heart sound clean signal.

[0111] Furthermore, the adaptive enhancement unit 203 is used to perform adaptive continuous wavelet denoising processing on the acceptable signal, and to perform transient artifact suppression processing on the denoised acceptable signal, so as to enhance the acceptable signal into the high-quality processed signal.

[0112] Furthermore, the adaptive enhancement unit 203 is also used to discard the acceptable signal along with the inferior signal when the acceptable signal cannot be enhanced into a high-quality processing signal.

[0113] In some embodiments, corresponding enhancement operations are triggered based on different quality levels of heart sound signals. High-quality signals do not require enhancement and can be directly input as clean heart sound signals to the subsequent heart sound analysis module 30. Acceptable signals are routed to the adaptive enhancement unit 203 for adaptive enhancement processing, thereby enhancing them into high-quality signals before being input as clean heart sound signals to the subsequent heart sound analysis module 30. Low-quality signals are discarded to save computational resources.

[0114] Furthermore, for the acceptable signal, the adaptive enhancement unit 203 performs adaptive continuous wavelet denoising processing. In some embodiments, for persistent background interference (such as ambient noise or fan hum) in the acceptable signal, a continuous noise removal algorithm based on Discrete Wavelet Transform (DWT) is deployed. This algorithm utilizes adaptive decomposition scale levels of 2 to 10. Level 2 provides slight baseline denoising, which is gradually amplified to level 10 only when the background noise is needed, thereby dynamically preventing excessive smoothing of pathological noises. Subsequently, the denoised acceptable signal undergoes transient artifact suppression processing. This stage applies an algorithmic suppression sequence to short-term, high-energy transient artifacts (such as sudden friction or impact of a stethoscope), specifically isolating and removing these high-entropy distortions without altering the underlying heart rhythm.

[0115] In some embodiments, adaptive median filtering based on energy thresholds can be used for transient artifact suppression. This involves dynamically adjusting the filter window size according to the energy distribution within the signal's neighborhood and replacing the original value with the median, thereby effectively suppressing sudden noise (such as impulse interference) while preserving the signal's detailed features and avoiding excessive smoothing caused by traditional filtering. It should be noted that other algorithms can also be used for transient artifact suppression, and this application does not limit the application to these algorithms.

[0116] Furthermore, after the adaptive enhancement unit 203 performs adaptive enhancement processing on the acceptable signal, it extracts the corresponding quality assessment feature vector again and inputs it into the assessment model. The assessment model then classifies and evaluates the signal again to obtain a quality level. If the acceptable signal is enhanced to a high-quality level, it is input as a high-quality processed signal into the subsequent heart sound analysis module 30. If the acceptable signal still cannot become a high-quality signal after adaptive enhancement processing, it should be discarded, and heart sounds can be collected again.

[0117] Figure 4 A schematic diagram of the architecture of the heart sound analysis module described in an embodiment of this application is shown. Figure 3 As shown, the heart sound analysis module 30 includes a division unit 301, an analysis unit 302, and a divergence unit 303.

[0118] The segmentation unit 301 is used to divide the clean heart sound signal into several sub-segments, each of which includes at least one complete cardiac cycle.

[0119] In some embodiments, when the duration of the heart sound clean signal is not less than 6 seconds, the segmentation unit 301 divides the continuous input signal into several overlapping sub-segments, and the length of each sub-segment is strictly limited to not less than 3 seconds to ensure that it contains at least one complete cardiac cycle.

[0120] Continue reading Figure 4 As shown, the analysis unit 302 is used to extract the original time-domain features and frequency-domain Mel frequency cepstral coefficients based on the sub-segments, to obtain the corresponding weighted fusion features based on the original time-domain features and the frequency-domain Mel frequency cepstral coefficients, and to predict and output the sub-analysis results of each sub-segment according to the weighted fusion features, so as to obtain the heart sound analysis results and the heart sound consensus ratio according to each sub-analysis result.

[0121] In some embodiments, the analysis unit 302 is a D-IFGNet classification model deployed in the cloud. This model not only outputs heart sound analysis results, but also achieves sensitive capture of "user-induced domain shift" through an innovative model architecture. Figure 5 A schematic diagram of the architecture of the analysis unit described in an embodiment of this application is shown. Figure 5 As shown, the analysis unit 302 includes an adaptive feature decomposition mechanism and a deep abstract coding network; the deep abstract coding network also includes a noise-aware gating mechanism and a clean adaptive feature fusion mechanism.

[0122] In some embodiments, the adaptive feature decomposition mechanism is used to perform dilated convolution and residual connection processing on the sub-segments to obtain first original temporal features, and to perform dense connection processing on the sub-segments to obtain second original temporal features. The adaptive feature decomposition mechanism (DRD Module) is the backbone feature extraction network of the analysis unit 302, employing a dilated-residual-dense (DRD) architecture. Through specific residual connection and feature filling strategies, it enhances the model's tolerance to periodic missing heart sound signals or sudden random interference. Dilated convolution is responsible for expanding the receptive field to capture long-term rhythms, while dense connections ensure the reuse of low-level acoustic features. (Continue reading...) Figure 5 As shown, in this mechanism, the dilated-residual block (D-RBlock) expands the receptive field through dilated convolution to capture the long-term rhythmic features of the heart sound signal, while using residual connections to enhance tolerance to periodic omissions or sudden random interference, thereby obtaining the first original temporal features. Meanwhile, the densely connected block (MCDBlock) reuses low-level acoustic features through a densely connected structure, ensuring efficient gradient propagation, thereby obtaining the second original temporal features.

[0123] Hole feature extraction is achieved through a one-dimensional convolution with an expansion rate of d, as shown in equation (13). Residual connections are then performed using equation (14).

[0124] (13)

[0125] Where n is the number of sampling points, k is the index of the weight element within the convolution kernel, K is the kernel length, and d is the dilation rate. Let be the learnable weight parameters at the k-th position of the convolution kernel. This represents the actual sampling offset. The value of the input signal at the nth sampling point, This is the output value.

[0126] (14)

[0127] in, For the output features of the l-th layer, The output features of the (l-1)th layer, For the nonlinear transformation to be learned in the l-th layer, This is the activation function.

[0128] Densely connected blocks enable cross-layer reuse of low-level acoustic features, as shown in Equation (15).

[0129] (15)

[0130] in, This indicates that a nonlinear transformation is performed on the tensor after concatenating features from multiple layers along the channel dimension.

[0131] In some embodiments, the noise-aware gating mechanism is used to acquire the intrinsic mode components (IMCs) of the sub-segment and obtain gating weights based on the energy entropy of the IMCs, so as to dynamically acquire gating features based on the gating weights. The noise-aware gating mechanism uses the IMCs of each sub-segment extracted by empirical mode decomposition as auxiliary inputs, thereby enabling the noise-aware gating mechanism to analyze the energy distribution of each IMC, dynamically sense the residual noise level in the signal, and obtain the local signal-to-noise ratio (SNR) of the current sub-segment. Further, the noise-aware gating mechanism dynamically calculates the gating weights based on the SNR, so as to dynamically acquire gating features based on the gating weights.

[0132] In some embodiments, empirical mode decomposition is performed on the sub-segment signal using equation (16) to obtain the intrinsic mode components.

[0133] (16)

[0134] in, For sub-segment signals, For the i-th intrinsic mode component, This represents the residual trend term.

[0135] In some embodiments, the gating weights are obtained by equation (17) based on the energy entropy of the intrinsic mode components. .

[0136] (17)

[0137] in, For the Sigmoid function, It is a learnable linear mapping function, and its specific implementation is not limited. Let be the energy entropy of the i-th intrinsic mode component, reflecting its energy distribution and information complexity. A higher energy entropy indicates a more dispersed energy distribution (noise characteristics), and the gating weight should approach zero to suppress this intrinsic mode component.

[0138] In some embodiments, the gating features are dynamically obtained based on the gating weights using equation (18).

[0139] (18)

[0140] in, These are gated features used to modulate the feature stream of deep abstract coding networks; The specific method for feature extraction is not limited to each intrinsic mode component.

[0141] In some embodiments, the deep abstract coding network is used to encode the first original temporal features and the second original temporal features, and to perform noise-adaptive feature modulation based on the gated features to obtain deep features. (Continue reading...) Figure 5 As shown, the Deep Abstraction Coding Network (IF Net) receives the first and second original temporal features output by the adaptive feature decomposition mechanism, and sequentially passes them through a channel attention module (SE-Block), a multihead attention layer, and a noise-aware gating mechanism to complete deep abstraction coding and obtain deep features. The noise-aware gating mechanism utilizes the gating features for adaptive weighted modulation to ultimately obtain the deep features.

[0142] Continue reading Figure 5 As shown, the deep abstract coding network also includes a clean-aware switching block. This block extracts frequency-domain Mel-frequency cepstral coefficients from the sub-segments and dynamically adjusts the fusion ratio of the deep features and the frequency-domain Mel-frequency cepstral coefficients based on a switching adjustment function to obtain the weighted fused features. The switching adjustment function is obtained based on the output confidence of the evaluation model.

[0143] In some embodiments, the switch adjustment function is defined as shown in equation (19).

[0144] (19)

[0145] in, This is a probability vector used to evaluate the class confidence level of the model output. The weight matrix is ​​a learnable matrix. This is the bias vector. The Softmax function normalizes the linear mapping result into a probability distribution and outputs... ,satisfy .in Each determines the balance between "relying on abstract semantics" and "relying on original physical laws".

[0146] Continue reading Figure 5 As shown, The switching signal is input to the local temporal clean branch to control its cleaning intensity. The local temporal clean branch acquires the frequency domain Mel-frequency cepstral coefficients of the sub-segment signal and then... As a regulating input, based on The fusion ratio of the depth feature and the frequency domain Mel frequency cepstral coefficient is dynamically adjusted to obtain the weighted fusion feature, as shown in Equation (20).

[0147] (20)

[0148] in, For weighted fusion features, For depth features, This refers to the physical acoustic properties (frequency domain Mel-frequency cepstral coefficients). When the quality level Q is high... Approaching 1, the model trusts deep abstract features to leverage their powerful semantic expressiveness; when the quality level Q is low, As the value approaches 1, the model then relies on the robustness of physical features to avoid amplifying erroneous semantics, thereby dynamically adjusting the fusion ratio of deep semantics and physical laws.

[0149] Continue reading Figure 5 As shown, the analysis unit 302 further includes a pooling layer and a classification layer. Weighted fusion features. After global feature aggregation through pooling layers to reduce spatial dimensionality, the data is fed into a classification layer to predict and output the sub-analysis results of the current sub-segment, including the heart sound category (normal or abnormal). In some embodiments, the classification layer outputs the prediction results of the current sub-segment through a fully connected network and a Softmax activation function.

[0150] In some embodiments, the analysis unit 302 outputs the sub-analysis results of each sub-segment as follows: (1 indicates pathological abnormality, 0 indicates normal). Then, the heart sound analysis results and the heart sound consensus ratio are obtained based on the results of each sub-analysis. The heart sound consensus ratio can be calculated using equation (21).

[0151] (twenty one)

[0152] in, For heart sound consensus ratio, This is the result of the i-th sub-analysis. Finally, when... If the heart sound analysis result is abnormal, that is... .

[0153] For further information, please refer to [link / reference]. Figure 4 As shown, the divergence unit 303 is used to determine whether the consensus ratio meets the preset divergence trigger boundary conditions. If so, the analysis alarm is triggered and the heart sound clean signal is transmitted to the divergence buffer.

[0154] In some embodiments, a consensus tolerance threshold is set. When the heartbeat consensus ratio V meets the preset divergence trigger boundary condition, that is... When this occurs, an analysis alert is triggered. This indicates a serious discrepancy between the evaluation confidence of the adaptive heart sound enhancement module and the analytical capabilities of the heart sound analysis module. At this point, these disputed boundary samples are intercepted and autonomously routed to the divergence buffer.

[0155] Continue reading Figure 1 As shown, at this time, the heart sound update module 40 retrains the evaluation model according to the professional quality label to obtain new evaluation model parameters, and transmits the new evaluation model parameters to the heart sound adaptive enhancement module so that the evaluation model can be adaptively updated; wherein, the new evaluation model parameters include model support vector parameters, Lagrange multipliers, and bias terms.

[0156] In some embodiments, clinical experts periodically re-examine the controversial samples in the disagreement buffer using auditory feedback and perform professional manual annotation to obtain professional quality labels, including unambiguous "good," "acceptable" requiring enhancement, and "poor" not recommended for diagnosis. The quality assessment feature vectors corresponding to the annotated controversial samples are then stored in a global "Quality Feature Buffer." This buffer operates as a first-in-first-out (FIFO) queue, dynamically maintaining statistical distribution balance by combining existing open-source support vectors and newly acquired specific user-specific controversial samples to prevent catastrophic amnesia. When the number of newly annotated samples reaches a predetermined threshold, the heart sound update module 40 retrains the assessment model based on the new annotated samples to obtain new assessment model parameters.

[0157] It should be noted that, in order to minimize bandwidth consumption and computational overhead, the heart sound update module 40 does not send back massive audio files or complex neural network weights to the heart sound adaptive enhancement module 20, but only sends out the updated lightweight evaluation model parameters, including the new support vectors. Adjusted Lagrange multipliers And updated bias terms .

[0158] Figure 6 A flowchart illustrating the intelligent heart sound analysis system described in an embodiment of this application is shown. Figure 6 As shown, the original heart sound signal is acquired, preprocessed, and then detected to obtain the heart sound signal. Next, a quality assessment feature vector is obtained based on the heart sound signal, and this feature vector is input into the assessment model to classify the heart sound signal into high-quality, acceptable, and low-quality signals. Specifically, the acceptable signal undergoes adaptive continuous wavelet denoising and transient artifact suppression processing, and is then input into the assessment model again for evaluation. If the signal is enhanced to a high-quality processed signal, it is input into the analysis unit along with the high-quality signal as a clean heart sound signal for analysis, and an analysis alarm is triggered. If an analysis alarm is triggered, the disputed clean heart sound signal is routed to the divergence buffer, undergoes expert review, and is rigorously manually annotated. The corresponding quality assessment feature vector is then stored in the quality feature buffer. When the number of newly annotated samples reaches a predetermined threshold, the assessment model is retrained to adaptively update the assessment model.

[0159] In some embodiments, the intelligent heart sound analysis system provided in this application can be deployed collaboratively at the edge and in the cloud to build an asymmetric edge-cloud collaborative architecture with deep coupling of data flow and control flow. Figure 7 A schematic diagram of the edge-cloud collaborative architecture of the intelligent heart sound analysis system described in an embodiment of this application is shown. Figure 7 As shown, the edge-end adaptive evaluation subsystem can be deployed on edge computing nodes such as dedicated electronic stethoscopes and tablets. Its purpose is to perform low-latency "evaluation-then-enhancement" cascaded traffic splitting and interception control before high-performance inference in the cloud. Specifically, after the raw heart sound acquisition unit completes the raw signal acquisition and preprocesses it through the preprocessing unit, the edge computing stage does not directly upload the raw high-frequency waveform. Instead, it extracts a quality evaluation feature vector through a feature extraction unit also deployed on the edge node. This feature vector explicitly includes key acoustic features such as amplitude kurtosis representing mechanical friction distortion, physiological energy ratio assessing background noise levels, and periodicity. The quality evaluation feature vector is then input into a lightweight evaluation model deployed at the edge for evaluation. If the signal is deemed "poor quality," edge interception is triggered, and the signal is discarded directly. If it is deemed "acceptable," adaptive enhancement is performed at the edge through the adaptive enhancement unit, and the signal is evaluated again through the evaluation model. Only when the signal is deemed "high quality" is the edge allowed to pass through, and the encrypted signal is reported to the heart sound analysis module deployed in the cloud for high-precision analysis.

[0160] Continue reading Figure 7As shown, after receiving the signal for permission to pass from the edge, the cloud calls the D-IFGNet model for feature extraction and depth analysis. This model uses a DRD (Diagram-Residual-Dense) architecture as its backbone to extract features and utilizes intrinsic mode component features extracted by empirical mode decomposition to dynamically calculate gating weights to perceive the residual noise level of the input signal, thereby obtaining depth features. Furthermore, the model uses the confidence probability vector from the forward evaluation feedback at the edge to construct a switching adjustment function. This allows for dynamic weighted switching between deep features and physical acoustic features (such as frequency domain Mel frequency cepstral coefficients), effectively avoiding excessive smoothing of key pathological noises under extreme noise conditions. This enables the acquisition of weighted fusion features, and based on these features, the heart sound analysis results and heart sound consensus ratio are obtained.

[0161] Furthermore, to achieve low-cost and continuous system evolution, a strict parameter feedback control flow was established between the cloud-based analytics model and the edge-based evaluation model. (See also...) Figure 7 As shown, the divergence unit determines whether the heart sound consensus ratio meets the preset divergence trigger boundary conditions. Only when the heart sound consensus ratio falls into the ambiguous boundary region (i.e., When a discrepancy is detected, the system determines that there is a significant discrepancy between the forward confidence score at the edge and the diagnostic result in the cloud, and automatically triggers a discrepancy alarm. After the discrepancy is triggered, the disputed sample is routed to the discrepancy buffer in the cloud. Clinical experts then perform auditory re-examination and three-category manual annotation on the boundary samples in the buffer, and store the corresponding quality assessment feature vectors in the quality feature buffer. Subsequently, when the number of newly annotated samples reaches a predetermined threshold, the heart sound update module 40 retrains the assessment model based on the newly annotated samples to obtain new assessment model parameters, and sends the updated lightweight assessment model parameters (i.e., new support vectors) to the edge segment. Lagrange multipliers and bias terms Edge-end receive parameters are plug-and-play, enabling adaptive iteration of the local evaluation model without consuming extremely high downlink bandwidth.

[0162] In this way, this application significantly reduces edge-to-cloud transmission bandwidth consumption through physical isolation between lightweight quality interception at the edge and dynamic analysis in the cloud. Simultaneously, this application, relying on a divergence return mechanism triggered by strict mathematical boundaries and a lightweight parameter distribution mechanism, endows the entire system with the ability to continuously adapt and evolve in diverse clinical acoustic environments.

[0163] Figure 8 A flowchart illustrating an intelligent heart sound analysis method according to an embodiment of this application is shown. This method is applied to the intelligent heart sound analysis system described in any of the above embodiments. Figure 8 As shown, the intelligent heart sound analysis method includes steps S1 to S5.

[0164] Step S1: Obtain the raw heart sound signal and perform preprocessing to obtain the heart sound signal.

[0165] Step S2: Obtain a quality assessment feature vector based on the heart sound signal, and input the quality assessment feature vector into the assessment model to obtain the quality level of the heart sound signal.

[0166] Step S3: Trigger the corresponding enhancement operation according to the quality level to obtain a clean heart sound signal.

[0167] Step S4: Obtain the heart sound analysis result and heart sound consensus ratio based on the heart sound clean signal, and determine whether to trigger an analysis alarm based on the heart sound consensus ratio.

[0168] Step S5: When the analysis alarm is triggered, the quality assessment feature vector is professionally labeled to obtain professional quality labels, and the assessment model is updated according to the professional quality labels.

[0169] It should be noted that the steps and principles of this method can be referred to the content described in the above embodiments, and will not be repeated here.

[0170] In summary, this application enables low-cost deployment, possesses environmental adaptability, and significantly improves the practicality and reliability of heart sound analysis systems in clinical settings.

[0171] First, this application significantly enhances the quality of heart sound signals. Based on constructed multidimensional features, this application achieves adaptive enhancement of heart sound signals, effectively and progressively enhancing signals with repair potential in high-noise clinical environments such as real hospitals. This avoids wasting valuable medical information, expands the coverage of heart sound analysis, and greatly improves the practicality and reliability of intelligent heart sound analysis.

[0172] Secondly, this application effectively overcomes user-induced domain shift and suppresses catastrophic forgetting during model evolution. This application proposes a consensus ratio divergence boundary triggering mechanism. This mechanism can accurately screen and intercept controversial samples with domain shift characteristics, and, in conjunction with closed-loop feedback from domain experts, dynamically maintain the statistical distribution balance of data features from both new and old environments. This not only enables the evaluation model to adapt to continuous changes in specific clinical acoustic environments at low cost, but also more effectively prevents the model from "catastrophic forgetting" of historical case features during dynamic update iterations.

[0173] Finally, this application adaptively avoids excessive smoothing of weak pathological features, improving diagnostic accuracy in noisy environments: This application introduces a noise-sensing gating and clean feature dynamic switching mechanism, adaptively and dynamically adjusting the fusion ratio of the original time-domain features and the frequency-domain Mel-frequency cepstral coefficients based on the real-time signal-to-noise ratio of the input signal. This underlying control logic effectively breaks through the limitations of fixed feature splicing, preventing weak but crucial pathological noise features from being excessively denoised and smoothed by the model as background noise under extreme clinical noise interference, thereby significantly improving the system's high sensitivity and diagnostic robustness in harsh clinical auscultation environments.

[0174] It should also be understood that the division of modules or units in the embodiments of this application is illustrative and only represents a logical functional division. In actual implementation, there may be other division methods. Furthermore, the functional modules in the various embodiments of this application can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0175] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0176] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. An intelligent heart sound analysis system, characterized in that, The system includes a heart sound acquisition module, a heart sound adaptive enhancement module, a heart sound analysis module, and a heart sound update module. The heart sound acquisition module is used to acquire the raw heart sound signal and perform preprocessing to obtain the heart sound signal; The adaptive enhancement module for heart sounds is used to obtain a quality assessment feature vector based on the heart sound signal, input the quality assessment feature vector into the assessment model to obtain the quality level of the heart sound signal, and then trigger a corresponding enhancement operation according to the quality level to obtain a clean heart sound signal. The heart sound analysis module is used to obtain the heart sound analysis result and heart sound consensus ratio based on the heart sound clean signal, and to determine whether to trigger an analysis alarm based on the heart sound consensus ratio. The heart sound update module is used to professionally label the clean heart sound signal to obtain a professional quality label when the analysis alarm is triggered, and update the evaluation model according to the professional quality label.

2. The intelligent heart sound analysis system according to claim 1, characterized in that, The adaptive heart sound enhancement module includes a feature extraction unit, an evaluation unit, and an adaptive enhancement unit. The feature extraction unit is used to extract quality assessment features based on the heart sound signal, so as to obtain the quality assessment feature vector based on the quality assessment features; the quality assessment features include at least one of the following: amplitude kurtosis, amplitude skewness, short-time zero-crossing rate variance, physiological energy ratio, high-frequency energy ratio, spectral flatness, heart sound periodicity, autocorrelation main peak ratio, frequency smoothing envelope variance, Shannon entropy, and modal energy entropy. The evaluation unit is used to input the quality evaluation feature vector into the evaluation model, so as to classify and evaluate the quality evaluation feature vector through the evaluation model and obtain the corresponding quality level; the quality level includes high-quality signal, acceptable signal and poor-quality signal; The adaptive enhancement unit is used to perform adaptive enhancement processing on the acceptable signal to enhance the acceptable signal into a high-quality processed signal, and together with the high-quality signal, serve as the clean heart sound signal.

3. The intelligent heart sound analysis system according to claim 2, characterized in that, The adaptive enhancement unit is used to perform adaptive continuous wavelet denoising on the acceptable signal and transient artifact suppression on the denoised acceptable signal to enhance the acceptable signal into the high-quality processed signal.

4. The intelligent heart sound analysis system according to claim 2, characterized in that, The adaptive enhancement unit is also configured to discard the acceptable signal along with the inferior signal when the acceptable signal cannot be enhanced into a high-quality processing signal.

5. The intelligent heart sound analysis system according to claim 1, characterized in that, The heart sound analysis module includes a segmentation unit, an analysis unit, and a divergence unit; The segmentation unit is used to divide the clean heart sound signal into several sub-segments, each of which includes at least one complete cardiac cycle. The analysis unit is used to extract original time-domain features and frequency-domain Mel frequency cepstral coefficients based on the sub-segments, to obtain corresponding weighted fusion features based on the original time-domain features and the frequency-domain Mel frequency cepstral coefficients, and to predict and output sub-analysis results of each sub-segment according to the weighted fusion features, so as to obtain the heart sound analysis results and the heart sound consensus ratio according to each sub-analysis result; The divergence unit is used to determine whether the consensus ratio meets the preset divergence trigger boundary conditions. If so, the analysis alarm is triggered and the heart sound clean signal is transmitted to the divergence buffer.

6. The intelligent heart sound analysis system according to claim 5, characterized in that, The analysis unit includes an adaptive feature decomposition mechanism and a deep abstract coding network; wherein, the deep abstract coding network includes a noise-aware gating mechanism; The adaptive feature decomposition mechanism is used to perform dilated convolution and residual connection processing on the sub-segments to obtain the first original temporal features, and to perform dense connection processing on the sub-segments to obtain the second original temporal features. The noise-aware gating mechanism is used to obtain the intrinsic mode components of the sub-segment and obtain gating weights based on the energy entropy of the intrinsic mode components, so as to dynamically obtain gating features based on the gating weights. The deep abstract coding network is used to encode the first original temporal features and the second original temporal features, and to perform noise-adaptive feature modulation based on the gated features to obtain deep features.

7. The intelligent heart sound analysis system according to claim 6, characterized in that, The deep abstract coding network also includes a clean adaptive feature fusion mechanism; The clean adaptive feature fusion mechanism is used to extract frequency domain Mel frequency cepstral coefficients based on the sub-segments, and dynamically adjust the fusion ratio of the deep features and the frequency domain Mel frequency cepstral coefficients based on a switching adjustment function to obtain the weighted fused features; wherein, the switching adjustment function is obtained based on the output confidence of the evaluation model.

8. The intelligent heart sound analysis system according to claim 1, characterized in that, The heart sound update module is used to retrain the evaluation model based on the professional quality label to obtain new evaluation model parameters, and transmit the new evaluation model parameters to the heart sound adaptive enhancement module so that the evaluation model can be adaptively updated; wherein, the new evaluation model parameters include model support vector parameters, Lagrange multipliers, and bias terms.

9. The intelligent heart sound analysis system according to claim 1, characterized in that, The heart sound acquisition module includes a raw heart sound acquisition unit, a preprocessing unit, and an effective signal detection unit; The original heart sound acquisition unit is used to acquire the original heart sound signal; The preprocessing unit is used to perform resampling, amplitude normalization and filtering / denoising on the original heart sound signal to obtain a processed heart sound signal. The effective signal detection unit is used to perform length detection and existence detection on the heart sound processing signal, and to take the heart sound processing signal whose signal length is greater than the length threshold and whose signal existence index value is greater than the existence threshold as the heart sound signal.

10. An intelligent heart sound analysis method, characterized in that, The method is applied to the intelligent heart sound analysis system as described in any one of claims 1 to 9, and the method includes: The raw heart sound signal is acquired and preprocessed to obtain the heart sound signal; A quality assessment feature vector is obtained based on the heart sound signal, and the quality assessment feature vector is input into the assessment model to obtain the quality level of the heart sound signal; Trigger the corresponding enhancement operation based on the quality level to obtain a clean heart sound signal; Based on the clean heart sound signal, obtain the heart sound analysis result and heart sound consensus ratio, and determine whether to trigger an analysis alarm based on the heart sound consensus ratio; When the analysis alert is triggered, the quality assessment feature vector is professionally labeled to obtain professional quality labels, and the assessment model is updated based on the professional quality labels.

Images