Shock classification method, device, equipment and medium based on heart sound analysis

Abstract

The present application is about a shock classification method, device, equipment and medium based on heart sound analysis, and specifically relates to the field of shock warning. The method includes: obtaining pre-processed heart sound data, the heart sound data includes multiple heart sound data segments; dividing the multiple heart sound data segments into interference-free heart sound data segments, recoverable heart sound data segments and irrecoverable heart sound data segments; performing signal recovery on the recoverable heart sound data segments through a heart sound signal recovery module that shares a complex-valued coding submodule with a heart sound data quality assessment module, and outputting the recovered heart sound data segments, wherein the interference-free heart sound data segments and the recovered heart sound data segments constitute the available heart sound data segments; inputting each of the available heart sound data segments into a shock classification unit to obtain the corresponding shock classification result; performing a majority vote on the multiple shock classification results to obtain the final shock automatic classification result.

Description

Technical Field

[0001] The present invention relates to the field of shock warning, and in particular to a shock classification method, device, equipment and medium based on heart sound analysis. Background Art

[0002] Shock is an acute state of insufficient tissue perfusion and impaired oxygen metabolism in the body. It has various causes and is an emergency that can easily lead to circulatory failure and seriously endanger life. Early detection and treatment of shock are of great significance. For example, among different types of shock, the incidence of hypovolemic shock caused by traumatic blood loss is relatively high, about 10% to 30%. The first hour after injury of these patients is called the "golden hour", and the mortality rate is highest within 6 hours, so timely treatment is crucial. For the four different types of shock, the treatment plans are different, so early and rapid classification and identification of shock and taking reasonable treatment measures are of great significance to reducing mortality.

[0003] In hospitals or places with medical staff, patients with shock can be detected early, but in many places outside hospitals, it is difficult to detect shock early. At this time, portable monitoring equipment is particularly important. Heart sound is an important biological signal of the human body. Portable heart sound monitoring and intelligent analysis systems have begun to be used in clinical practice.

[0004] Therefore, how to automatically and efficiently extract heart sound features for analysis, provide a scientific basis for early identification of shock, accurately determine the type of shock, so as to implement reasonable anti-shock treatment measures and improve patient prognosis is a problem that needs to be solved urgently. Summary of the invention

[0005] The present application provides a shock classification method, device, equipment and medium based on heart sound analysis, and the technical solution is as follows.

[0006] In one aspect, a shock classification method based on heart sound analysis is provided, the method comprising:

[0007] Acquiring preprocessed heart sound data, wherein the heart sound data includes a plurality of heart sound data segments;

[0008] Performing quality assessment on the heart sound data by a heart sound data quality assessment module, dividing the plurality of heart sound data segments into interference-free heart sound data segments, recoverable heart sound data segments and irrecoverable heart sound data segments, wherein the heart sound data quality assessment module includes a complex-valued coding submodule;

[0009] The heart sound signal recovery module of the complex-valued coding submodule is shared with the heart sound data quality assessment module to perform signal recovery on the recoverable heart sound data segment and output a recovered heart sound data segment, wherein the interference-free heart sound data segment and the recovered heart sound data segment constitute an available heart sound data segment;

[0010] Inputting each of the available heart sound data segments into a shock classification unit to obtain a corresponding shock classification result, wherein the shock classification unit includes a parallel processing time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork;

[0011] A majority vote is performed on the plurality of shock classification results to obtain a final automatic shock classification result.

[0012] In a possible implementation, the performing of quality assessment on the heart sound data by a heart sound data quality assessment module, and dividing the plurality of heart sound data segments into non-interference heart sound data segments, recoverable heart sound data segments, and irrecoverable heart sound data segments, includes:

[0013] For any target heart sound data segment in the heart sound data, transform the target heart sound data segment into the time-frequency domain to obtain a corresponding target complex time-frequency spectrum matrix;

[0014] The target complex time-frequency spectrum matrix is ​​used as the input of the complex-valued coding submodule, and the corresponding target complex-valued coding features are output. The complex-valued coding submodule includes: a complex two-dimensional convolutional layer, S levels of complex Transformer and downsampling, and a complex Transformer;

[0015] The absolute value of the target complex-valued coding feature is taken as input, and is sequentially passed through a fully connected layer and ReLU, a fully connected layer and Softmax to divide the target heart sound data segment into one of an undisturbed heart sound data segment, a recoverable heart sound data segment and an irrecoverable heart sound data segment.

[0016] In a possible implementation, the performing signal recovery on the recoverable heart sound data segment through the heart sound signal recovery module that shares the complex-valued coding submodule with the heart sound data quality assessment module and outputting the recovered heart sound data segment includes:

[0017] In the case where the target heart sound data segment is divided into the recoverable heart sound data segment, the target complex-valued coding feature output by the complex-valued coding submodule of the target heart sound data segment is reused, the target complex-valued coding feature is used as the input of the complex-valued decoding submodule, and the corresponding target complex-valued decoding feature is output;

[0018] transforming the target complex-valued decoding feature into the time domain to obtain the restored heart sound data segment;

[0019] The complex-valued decoding submodule includes, in sequence: S levels of upsampling and complex Transformer, and a complex two-dimensional convolutional layer.

[0020] In a possible implementation, the input of the complex Transformer of the kth level of the complex-valued decoding submodule is: the concatenation of the upsampling result of the kth level and the output result of the complex Transformer of the S-k+1th level in the complex-valued encoding submodule.

[0021] In one possible implementation, the complex Transformer is composed of a cascade of a multi-head self-attention submodule for complex numbers and a gated feedforward submodule for complex numbers.

[0022] In a possible implementation, inputting each of the available heart sound data segments into the shock classification unit to obtain a corresponding shock classification result includes:

[0023] For any target available heart sound data segment in the available heart sound data, the target available heart sound data segment is respectively input into a time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork to obtain a time-frequency domain extraction feature and a time-domain extraction feature;

[0024] After the time-frequency domain extracted features and the time domain extracted features are input into a feature fusion submodule, they are sequentially passed through a fully connected layer and ReLU, a fully connected layer and Softmax to obtain the shock classification result corresponding to the target available heart sound data segment.

[0025] In a possible implementation, the time-frequency domain Transformer feature extraction subnetwork includes: a cascaded short-time Fourier transform module, a multi-head self-attention module, and a gated feedforward module;

[0026] The time-domain Transformer feature extraction subnetwork includes: a cascaded one-dimensional convolution layer, Q stacked hole convolution residual modules, a multi-head self-attention module and a gated feedforward module.

[0027] In a possible implementation, the processing of the feature fusion submodule includes:

[0028] Performing layer normalization and 1×1 convolution processing on the time-frequency domain extraction features and the time-domain extraction features, respectively, to obtain time-frequency domain intermediate features and time-domain intermediate features;

[0029] Obtaining the time-frequency domain intermediate features and the attention weight matrix of each of the time-domain intermediate features;

[0030] Using an attention interaction mechanism, the time-frequency domain intermediate features, the time domain intermediate features, and two attention weight matrices are operated to obtain fused features;

[0031] The fused features are sequentially passed through a depth-wise separable convolutional layer and ReLU, layer normalization, and the feature fusion result is output.

[0032] In another aspect, a shock classification device based on heart sound analysis is provided, the device comprising:

[0033] A data preprocessing unit, used to obtain preprocessed heart sound data, wherein the heart sound data includes a plurality of heart sound data segments;

[0034] a data quality control unit, configured to perform quality assessment on the heart sound data through a heart sound data quality assessment module, and divide the plurality of heart sound data segments into non-interference heart sound data segments, recoverable heart sound data segments, and irrecoverable heart sound data segments, wherein the heart sound data quality assessment module includes a complex-valued coding submodule; perform signal recovery on the recoverable heart sound data segments through a heart sound signal recovery module that shares the complex-valued coding submodule with the heart sound data quality assessment module, and output a recovered heart sound data segment, wherein the non-interference heart sound data segment and the recovered heart sound data segment constitute an available heart sound data segment;

[0035] A shock classification unit, used to take each of the available heart sound data segments as input to obtain a corresponding shock classification result, wherein the shock classification unit includes a parallel processing time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork;

[0036] The fusion output unit is used to perform majority voting on the plurality of shock classification results to obtain a final shock automatic classification result.

[0037] On the other hand, a computer device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by the processor to implement the above-mentioned shock classification method based on heart sound analysis.

[0038] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set or instruction set is stored in the computer-readable storage medium, and the at least one instruction, at least one program, code set or instruction set is loaded and executed by a processor to implement the above-mentioned shock classification method based on heart sound analysis.

[0039] In another aspect, a computer program product or a computer program is provided, the computer program product or the computer program comprising computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the above-mentioned shock classification method based on heart sound analysis.

[0040] The technical solution provided by this application may have the following beneficial effects:

[0041] The preprocessed heart sound data are automatically acquired, the quality of each heart sound data segment in the heart sound data is judged, the signal is restored for the heart sound data segments judged as recoverable heart sound data segments, and the available heart sound data segments consisting of the recovered heart sound data segments and the non-interference heart sound data segments are input into the shock classification unit. The shock classification result corresponding to each available heart sound data segment is obtained through the time-frequency domain Transformer feature extraction subnetwork and the time domain Transformer feature extraction subnetwork processed in parallel in the shock classification unit. Finally, multiple shock classification results are combined to obtain the final shock classification result. While realizing the early automatic classification and recognition of shock in various application scenarios inside and outside the hospital, the reliability of the early automatic classification of shock and the efficiency of data use are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the specific implementation methods of the present application or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0043] Figure 1 The present invention is a flowchart of a shock classification method based on heart sound analysis according to an exemplary embodiment.

[0044] Figure 2 The figure is a schematic diagram of waveforms corresponding to different heart sounds according to an exemplary embodiment.

[0045] Figure 3 It is a structural block diagram of a data quality control unit according to an exemplary embodiment.

[0046] Figure 4 The figure is a structural block diagram of a shock classification unit according to an exemplary embodiment.

[0047] Figure 5 The invention is a structural block diagram of a shock classification device based on heart sound analysis according to an exemplary embodiment.

[0048] Figure 6 is a schematic diagram of a computer device provided according to an exemplary embodiment. DETAILED DESCRIPTION

[0049] The technical solution of the present application will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present application.

[0050] It should be understood that the "indication" mentioned in the embodiments of the present application can be a direct indication, an indirect indication, or an indication of an association relationship. For example, A indicates B, which can mean that A directly indicates B, for example, B can be obtained through A; it can also mean that A indirectly indicates B, for example, A indicates C, B can be obtained through C; it can also mean that there is an association relationship between A and B.

[0051] In the description of the embodiments of the present application, the term "corresponding" may indicate a direct or indirect correspondence between two items, or an association relationship between the two items, or a relationship between indication and being indicated, configuration and being configured, and the like.

[0052] In an embodiment of the present application, "predefinition" can be achieved by pre-saving corresponding codes, tables or other methods that can be used to indicate relevant information in a device (for example, including a terminal device and a network device). The present application does not limit its specific implementation method.

[0053] According to the changes in hemodynamics and microcirculation, the development of shock can be divided into three stages: early shock, shock period, and late shock. The early shock is the body's compensatory period, during which the body's blood is redistributed under the stimulation of shock factors to ensure blood supply to important organs. The shock period is the body's decompensatory period, at which time organ damage continues to worsen due to obvious tissue hypoxia and a significant decrease in venous return. The late shock is the irreversible stage of shock, at which time cells necrotize, autolyze, and disseminate bleeding after persistent hypoxia, and important organs throughout the body fail. Therefore, early identification and warning of shock are of great significance.

[0054] Currently, there are a small number of patent applications for early prediction of shock, risk warning, staging, decision support, and prognosis based on clinical in-hospital data or data collected in combination with commonly used portable devices, including:

[0055] The invention patent application "Real-time risk warning monitoring method and system for shock based on medical Internet of Things time series data and deep learning algorithm" (application number 202311690984.6) takes a time series data set including the patient's clinical time series data and labels as the processing object, performs sparse representation and data reconstruction to obtain enhanced data, and inputs it into several types of conventional machine learning models to obtain real-time risk warning monitoring results of the disease.

[0056] The invention patent application "An early prediction system for septic shock based on machine learning" (application number 202310332897.7) extracts several predictive feature variables from previous electronic medical records or medical data sets, and uses an improved fusion algorithm to build a prediction model, thereby achieving the effect of using previous vital signs to determine whether there will be septic shock in the future.

[0057] The invention patent application "A modeling method and system for predicting the prognosis of septic shock" (application number 202310862389.X) selected relatively easy-to-obtain in-hospital indicators, and used plasma heparin-binding protein, procalcitonin, C-reactive protein and arterial blood lactate as predictive indicators of the convolutional neural network model, which can achieve rapid prediction of the prognosis of septic shock in patients within 24 hours of admission.

[0058] The invention patent applications "A real-time risk warning monitoring system, equipment and storable medium for cardiogenic shock" (application number 202310783406.0), "A rapid staging and triage system, equipment and storable medium for cardiogenic shock" (application number 202310862389.X), "A clinical decision support system, equipment and storable medium for cardiogenic shock" (application number 202310697356.4), and "A cardiogenic shock prognosis prediction and warning system, equipment and storable medium" (application number 202310782125.3) use in-hospital clinical data and indicators to perform risk warning, staging and triage, clinical decision support and prognosis prediction for a special type of shock such as cardiogenic shock.

[0059] The invention patent application "Portable Shock Detection Device" (application number 202111046782.9) uses body surface vibration signals and compares the heart rate and respiratory rate calculated therefrom with the threshold to give the risk of shock.

[0060] The main disadvantages of the above existing patents are:

[0061] (1) The data used are mainly clinical indicators and monitoring data in hospitals, which makes it difficult to achieve automatic classification and identification of shock in the early stage in out-of-hospital application scenarios;

[0062] (2) Using a small number of other portable devices to collect signals for shock detection, the low-quality data collected by the portable devices will cause errors in classification and recognition results;

[0063] (3) When extracting features from data, the extracted features are subjective or fail to fully consider the characteristics of the time series signal itself to extract features for shock classification and identification;

[0064] (4) The decision method for shock classification and recognition is simple, and the final shock classification and recognition performance is poor.

[0065] In view of the above defects, in an embodiment of the present application, a scheme for automatic classification and identification of early shock based on heart sound signal analysis is provided. Please refer to the following embodiment for details.

[0066] Figure 1 is a method flow chart of a shock classification method based on heart sound analysis according to an exemplary embodiment. The method is applied in a computer device, which may be a portable heart sound device, such as Figure 1 As shown, the shock classification method based on heart sound analysis may include the following steps:

[0067] Step 110: Acquire pre-processed heart sound data, where the heart sound data includes a plurality of heart sound data segments.

[0068] The heart sound data is a digitized heart sound signal picked up by sampling at the apex of the heart; and the heart sound data segments are data segments of equal length formed after segmenting the heart sound data.

[0069] In an embodiment of the present application, heart sound data is collected by a portable heart sound device, and the heart sound data is pre-processed such as filtering to obtain a plurality of heart sound data segments of equal length.

[0070] For example, at the apex of the heart with a sampling rate f s (f s ≥2000Hz) to obtain a digitized heart sound signal x of T (T≥4) seconds in length; the heart sound signal x is passed through a K-order forward and backward Butterworth bandpass filter with a passband of 25-500Hz, and then divided into M heart sound data segments x of equal length with 50% overlap m ,m=1,2,…,M, where the length of each heart sound data segment is N sampling points.

[0071] Step 120: Perform quality assessment on the heart sound data through a heart sound data quality assessment module, and divide the multiple heart sound data segments into interference-free heart sound data segments, recoverable heart sound data segments, and irrecoverable heart sound data segments. The heart sound data quality assessment module includes a complex-valued coding submodule.

[0072] Among them, the heart sound data quality assessment module is a neural network module used to classify the quality type of heart sound data segments, which includes a complex-valued encoding submodule, which is a submodule that can encode complex-valued data such as heart sound data.

[0073] In an embodiment of the present application, three quality types of heart sound data segments are pre-set: interference-free heart sound data segments, recoverable heart sound data segments, and irrecoverable heart sound data segments. A heart sound data quality assessment module is used to perform quality assessment on each heart sound data segment in the heart sound data to assess the quality type of each heart sound data segment.

[0074] In a possible implementation, the heart sound data quality assessment module is composed of a three-classification deep neural network with trained weights. The training process of the three-classification deep neural network corresponding to the heart sound data quality assessment module is as follows: construct a heart sound data set with labels, each segment of data is of equal length to N sampling points, where the label [1,0,0] T Indicates a heart sound data segment without interference, label [0,1,0] T Indicates that the heart sound data segment can be restored, and the label [0,0,1] T It indicates an irrecoverable heart sound data segment, and the sample size of the data corresponding to the three types of labels is not less than 1000; the weights of the complex-valued coding submodule in the network use the trained weights of the complex-valued coding submodule in the heart sound signal recovery module in step 130. During the training process, this part of the weights is frozen, and only the weights of the remaining part of the network are trained; the loss function used in training is the cross entropy function, and training is performed by back propagation.

[0075] Step 130: Perform signal recovery on the recoverable heart sound data segments through the heart sound signal recovery module that shares the complex-valued coding submodule with the heart sound data quality assessment module, and output the recovered heart sound data segments. The interference-free heart sound data segments and the recovered heart sound data segments constitute the available heart sound data segments.

[0076] Among them, the heart sound signal recovery module is a neural network module used to recover the signal of the recoverable heart sound data segment, and it shares a complex-valued coding submodule with the heart sound data quality assessment module.

[0077] In an embodiment of the present application, a heart sound data quality assessment module is used to perform signal recovery on heart sound data segments whose quality type belongs to recoverable heart sound data segments, thereby obtaining recovered heart sound data segments, thereby effectively utilizing the data of the recoverable heart sound data segments, and combining the interference-free heart sound data segments and the recovered heart sound data segments into usable heart sound data segments.

[0078] In a possible implementation, the heart sound signal recovery module is composed of a complex restorer with trained weights, and the training process of the complex restorer corresponding to the heart sound signal recovery module is: constructing a noisy heart sound signal recovery data set, wherein the noisy heart sound data is a mixture of labeled heart sound data and noise in different proportions, the labeled heart sound data is noise-free heart sound data, and each segment of data has N sampling points of equal length; the data sample size is not less than 5000 cases; the loss function used for training is the superposition of mean square error and mutual correlation coefficient, and training is performed by back propagation.

[0079] Step 140: Input each of the available heart sound data segments into the shock classification unit to obtain the corresponding shock classification result. The shock classification unit includes a parallel processing time-frequency domain Transformer feature extraction subnetwork and a time domain Transformer feature extraction subnetwork.

[0080] Among them, the time-frequency domain Transformer feature extraction subnetwork is a neural network module with the ability to extract features in the time-frequency domain, and the time-domain Transformer feature extraction subnetwork is a neural network module with the ability to extract features in the time domain.

[0081] In the embodiment of the present application, the shock classification result is identified for each heart sound data segment in the available heart sound data segments by using a shock classification unit having a time-frequency domain Transformer feature extraction subnetwork and a time domain Transformer feature extraction subnetwork.

[0082] In one possible implementation, the shock classification unit is composed of a two-way Transformer feature fusion deep neural network with trained weights. The training process of the two-way Transformer feature fusion deep neural network is as follows: construct a labeled heart sound dataset, each segment of data is of equal length to N sampling points, where the label [1, 0, 0, 0, 0] T Indicates hypovolemic shock, [0,1,0,0,0] T Indicates obstructive shock, [0,0,1,0,0] T Indicates distributed shock, [0,0,0,1,0] T Indicates cardiogenic shock, [0,0,0,0,1] T Indicates a non-shock state. The sample size of the data corresponding to each of the five labels is not less than 2000. The loss function used in training is the cross entropy function, and training is performed by back propagation.

[0083] For example, from Figure 2Select N = 16512 sampling points from any of the 5 heart sounds belonging to different categories and input them into the shock classification unit. The network output result of (a) the hypovolemic shock heart sound segment is [1, 0, 0, 0, 0] T , (b) The network output of the obstructive shock heart sound segment is [0,1,0,0,0] T , (c) The network output of the distributed shock heart sound segment is [0,0,1,0,0] T , (d) The network output of the cardiogenic shock heart sound segment is [0,0,0,1,0] T , (d) The network output of the heart sound segment in the non-shock state is [0,0,0,0,1] T .

[0084] Step 150: Perform majority voting on multiple shock classification results to obtain a final shock automatic classification result.

[0085] In an embodiment of the present application, a plurality of shock classification results output by a shock classification unit from a plurality of available heart sound data segments are subjected to majority voting to give a final 5-classification result: hypovolemic shock, obstructive shock, distributive shock, cardiogenic shock or non-shock state.

[0086] Among them, hypovolemic shock refers to shock caused by a decrease in total blood volume, which is often caused by traumatic blood loss, fluid loss, burns, etc.; obstructive shock refers to shock caused by blood flow obstruction due to extracardiac factors such as pulmonary vascular embolism and pulmonary hypertension; distributive shock refers to shock caused by abnormal blood distribution causing a large amount of blood to stagnate in the dilated small blood vessels and a decrease in effective circulating blood volume. It is common in infectious and non-infectious shock caused by systemic inflammatory response syndrome, anaphylactic shock, etc.; cardiogenic shock refers to heart pump failure caused by the heart itself or extracardiac reasons, which causes a decrease in effective circulating blood volume and microcirculation perfusion and leads to shock. It is common in patients with acute myocardial infarction, heart failure, fulminant myocarditis, etc.

[0087] In summary, the shock classification method based on heart sound analysis provided in this embodiment automatically obtains preprocessed heart sound data, performs quality judgment on each heart sound data segment in the heart sound data, performs signal recovery on the heart sound data segment judged as a recoverable heart sound data segment, and inputs the available heart sound data segments consisting of the recovered heart sound data segments and the non-interference heart sound data segments into the shock classification unit, and obtains the shock classification result corresponding to each available heart sound data segment through the time-frequency domain Transformer feature extraction subnetwork and the time domain Transformer feature extraction subnetwork processed in parallel in the shock classification unit, and finally combines multiple shock classification results to obtain the final shock classification result, which improves the reliability of the early automatic classification of shock and the efficiency of data use while realizing the early automatic classification and recognition of shock in various application scenarios inside and outside the hospital.

[0088] In an exemplary embodiment, Figure 3 As shown, the quality classification and quality recovery of the heart sound data segments are performed through the heart sound data quality assessment module and the heart sound signal recovery module of the shared complex-valued coding submodule.

[0089] Specifically, step 120 includes:

[0090] Step 121: for any target heart sound data segment in the heart sound data, transform the target heart sound data segment into the time-frequency domain to obtain a corresponding target complex time-frequency spectrum matrix.

[0091] In the embodiment of the present application, firstly, the target heart sound data segment is transformed into the time-frequency domain using the short-time Fourier transform (STFT), and the target complex time-frequency spectrum matrix outputted by the STFT is used.

[0092] Step 122: Use the target complex time-frequency spectrum matrix as the input of the complex-valued coding submodule, and output the corresponding target complex-valued coding features. The complex-valued coding submodule includes: a complex two-dimensional convolutional layer, S levels of complex Transformer and downsampling, and a complex Transformer.

[0093] In an embodiment of the present application, the target complex time-frequency spectrum matrix is ​​passed through a complex-valued coding submodule, which includes a complex two-dimensional convolution layer, S levels of complex Transformer and maximum pooling downsampling, and a complex Transformer in sequence to output the target complex-valued coding features.

[0094] Among them, the complex two-dimensional convolutional layer is used to map the single-channel target complex time-frequency spectrum matrix into the initial heart sound features of multiple channels; the S-level complex Transformer and maximum pooling downsampling are used to extract and compress the S-level heart sound time series features respectively; the complex Transformer is used to extract the final heart sound features as the target complex-valued coding features.

[0095] Step 123: The absolute value of the target complex-valued encoding feature is used as input, and is sequentially passed through a fully connected layer and ReLU, a fully connected layer and Softmax to divide the target heart sound data segment into one of an undisturbed heart sound data segment, a recoverable heart sound data segment and an irrecoverable heart sound data segment.

[0096] In the embodiment of the present application, one fully connected layer (Linear) and ReLU, one fully connected layer (Linear) and Softmax are passed in sequence, and finally the probability of each of the three categories (undisturbed heart sound data segment, recoverable heart sound data segment, and irrecoverable heart sound data segment) is output in the form of a three-dimensional vector. The largest probability is set to 1 and the others are set to zero, thereby obtaining the quality judgment result of the target heart sound data segment.

[0097] Among them, the fully connected layer and ReLU are used to further extract and integrate the amplitude of the target complex-valued coding feature extracted by the complex Transformer to obtain the first feature for distinguishing the quality of heart sounds; the fully connected layer is used to further calculate the first feature for distinguishing the quality of heart sounds to obtain the second feature for distinguishing the quality of heart sounds; Softmax is used to calculate the probability of the three categories of quality distinction of the target heart sound data segment (undisturbed heart sound data segment, recoverable heart sound data segment, and irrecoverable heart sound data segment) based on the second feature for distinguishing the quality of heart sounds and output them.

[0098] Specifically, step 130 includes:

[0099] Step 131: When the target heart sound data segment is divided into recoverable heart sound data segments, the target complex-valued coding features output by the complex-valued coding submodule of the target heart sound data segment are reused, and the target complex-valued coding features are used as inputs of the complex-valued decoding submodule to output corresponding target complex-valued decoding features.

[0100] The complex-valued decoding submodule includes: S levels of upsampling and complex Transformer, and a complex two-dimensional convolutional layer.

[0101] Among them, S levels of upsampling and complex Transformer are used to expand and extract S levels of pure heart sound time series features respectively; the complex two-dimensional convolutional layer is used to map the pure heart sound time series features of the Sth level into a single-channel restored heart sound complex time-frequency spectrum matrix as the target complex-valued decoding feature.

[0102] In the embodiment of the present application, the complex-valued coding submodule in the heart sound data quality assessment module and the complex-valued coding submodule in the heart sound signal recovery module have exactly the same structure and share weights. Therefore, when a target heart sound data segment is identified as a recoverable heart sound data segment by the heart sound data quality assessment module, the target complex-valued coding feature of the target heart sound data segment output by the complex-valued coding submodule is directly input into the complex-valued decoding submodule in the heart sound signal recovery module to obtain the target complex-valued decoding feature.

[0103] Furthermore, the input of the complex Transformer of the kth level in the complex-valued decoding submodule is the concatenation of the upsampling result of the kth level and the output result of the complex Transformer of the S-k+1th level in the complex-valued encoding submodule. Using the above concatenation result as input can make full use of the features of each level, avoid information loss, and reduce the risk of overfitting in network training.

[0104] Step 132: transform the target complex-valued decoding feature into the time domain to obtain a restored heart sound data segment.

[0105] In the embodiment of the present application, the output of the complex-valued decoding submodule is transformed into the time domain using an inverse short-time Fourier transform (ISTFT), and the restored heart sound data segment is output.

[0106] In one possible implementation, the complex Transformer in the above steps is composed of a cascade of a multi-head self-attention submodule for complex numbers and a gated feedforward submodule for complex numbers.

[0107] Among them, the complex multi-head self-attention sub-module is used for complex time series feature extraction; the complex gated feedforward sub-module is used for further refinement and integration of complex time series features.

[0108] Among them, the multi-head self-attention submodule includes H-head self-attention output submodules, and the output result of the h-th self-attention output submodule is Q h , K h 、V h are the query matrix, key matrix, and value matrix calculated by the h-th self-attention output submodule using three complex depthwise separable convolutional layers, d h is the dimension of the bond matrix, (·) H is the conjugate transpose of the matrix, ⊙ is the element-wise multiplication, and the output of the multi-head self-attention submodule is the layer-normalized result of the concatenation of the output results of all H-head self-attention output submodules.

[0109] Among them, the gated feedforward submodule includes three processing paths: the first path is a direct connection, the second path is a complex depthwise separable convolutional layer, and the third path is a complex depthwise separable convolutional layer and a Gaussian error linear unit. The output result of the gated feedforward submodule is the layer normalization result after the path fusion feature and the first path feature are concatenated. The path fusion feature is the feature obtained by element-wise multiplication of the calculation results of the second path and the third path, and then through 1×1 convolution.

[0110] In summary, in the shock classification method based on heart sound analysis provided in this embodiment, during the process of quality assessment of heart sound data segments and signal recovery, the heart sound data quality assessment module and the heart sound signal recovery module share a complex-valued coding submodule based on a complex Transformer and its weights, thereby achieving the effect of streamlining the network and improving network utilization efficiency.

[0111] In an exemplary embodiment, Figure 4 As shown in the figure, the function of the shock classification unit is realized by fusing the dual-path Transformer features into a deep neural network.

[0112] Specifically, step 140 includes:

[0113] Step 141: For any target available heart sound data segment in the available heart sound data, the target available heart sound data segment is input into the time-frequency domain Transformer feature extraction subnetwork and the time domain Transformer feature extraction subnetwork respectively to obtain the time-frequency domain extraction features and the time domain extraction features.

[0114] In an embodiment of the present application, the time-frequency domain Transformer feature extraction subnetwork and the time-domain Transformer feature extraction subnetwork are two parallel feature extraction subnetworks. By inputting any target available heart sound data segment into the time-frequency domain Transformer feature extraction subnetwork and the time-domain Transformer feature extraction subnetwork respectively, two types of features can be extracted: time-frequency domain extraction features and time domain extraction features.

[0115] Furthermore, the time-frequency domain Transformer feature extraction sub-network includes: a cascaded short-time Fourier transform module (STFT), a multi-head self-attention module and a gated feedforward module.

[0116] Among them, the short-time Fourier transform module is used to extract the time-frequency spectrum feature matrix of the available heart sound data segment and obtain its real time-frequency amplitude spectrum matrix; the multi-head self-attention module is used to extract the time-frequency features of time series data; and the gated feedforward module is used to further refine and integrate the time-frequency features of time series data.

[0117] Furthermore, the time-domain Transformer feature extraction sub-network includes: a cascaded one-dimensional convolutional layer, Q stacked dilated convolutional residual modules, a multi-head self-attention module and a gated feedforward module.

[0118] Among them, the one-dimensional convolution layer is used to map the single-channel input time series data into the initial heart sound time series features of multiple channels; the qth atrous convolution residual module is used to extract the qth large receptive field heart sound time series features; the multi-head self-attention module is used for heart sound time domain feature extraction; and the gated feedforward module is used for further refinement and integration of heart sound time domain features.

[0119] Among them, the multi-head self-attention submodule includes H-head self-attention output submodules, and the output result of the h-th self-attention output submodule is Q h , K h 、V h are the query matrix, key matrix, and value matrix calculated by the h-th self-attention output submodule using three depth-wise separable convolutional layers, respectively, and d h is the dimension of the bond matrix, (·) T is the transpose of the matrix, ⊙ is the element-wise multiplication, and the output of the multi-head self-attention submodule is the layer-normalized result of the concatenation of the output results of all H-head self-attention output submodules;

[0120] Among them, the gated feedforward submodule includes three processing paths: the first path is a direct connection, the second path is a depth-wise separable convolutional layer, and the third path is a depth-wise separable convolutional layer and a Gaussian error linear unit. The output of the gated feedforward submodule is the layer normalization result after the path fusion feature and the first path feature are concatenated. The path fusion feature is the feature obtained by element-wise multiplication of the calculation results of the second path and the third path, and then through 1×1 convolution.

[0121] Step 142: After the time-frequency domain extracted features and the time domain extracted features are input into a feature fusion submodule, they are sequentially passed through a fully connected layer and ReLU, a fully connected layer and Softmax to obtain the shock classification result corresponding to the target available heart sound data segment.

[0122] In an embodiment of the present application, after the time-frequency domain extraction features and the time domain extraction features are simultaneously input into a feature fusion submodule, they are sequentially passed through 1 fully connected layer (Linear) and ReLU, 1 fully connected layer (Linear) and Softmax, and finally the probabilities of the five categories (hypovolemic shock, obstructive shock, distributive shock, cardiogenic shock and non-shock state) are output in the form of a 5-dimensional vector. The maximum probability is set to 1 and the others are set to zero, thereby obtaining the shock classification result.

[0123] Among them, the fully connected layer and ReLU are used to further extract and integrate the fused features output by the feature fusion submodule to obtain the first feature of shock classification heart sounds; the fully connected layer is used to further calculate the first feature of shock classification heart sounds to obtain the second feature of shock classification heart sounds; Softmax is used to calculate the probability of each of the five shock categories (hypovolemic shock, obstructive shock, distributive shock, cardiogenic shock and non-shock state) of the target heart sound data segment based on the second feature of shock classification heart sounds and output them.

[0124] Furthermore, the processing of the feature fusion submodule includes:

[0125] (1) Perform layer normalization and 1×1 convolution processing on the time-frequency domain extracted features and the time domain extracted features respectively to obtain the time-frequency domain intermediate features and the time domain intermediate features.

[0126] (2) Obtain the attention weight matrices of the intermediate features in the time-frequency domain and the intermediate features in the time domain.

[0127] (3) The attention interaction mechanism is used to operate the intermediate features in the time-frequency domain, the intermediate features in the time domain, and the two attention weight matrices to obtain the fused features.

[0128] (4) The fused features are passed through a depth-wise separable convolutional layer and ReLU,layer normalization, and the feature fusion result is output.

[0129] Among them, the depth-wise separable convolutional layer and ReLU are used to simplify the weight size and output feature scale, and accelerate forward operations and backward propagation training; layer normalization is used to accelerate backward propagation training and improve the generalization ability of the network.

[0130] In summary, the shock classification method based on heart sound analysis provided in this embodiment fully considers the characteristics of heart sound signals as time series signals when designing the shock classification unit, utilizes the expansion of the receptive field of the time series signal by the dilated convolution residual module and the high efficiency of the transformer module in extracting features of the time series signal, fully extracts features from both the time-frequency domain and the time domain, and the features extracted by the two feature extraction subnetworks are refined through a feature fusion submodule by using the two-way attention interactive synthesis method on the two paths to improve the final shock classification accuracy.

[0131] Figure 5 The present invention is a block diagram of a shock classification device based on heart sound analysis according to an exemplary embodiment. The device comprises:

[0132] A data preprocessing unit 501 is used to obtain preprocessed heart sound data, wherein the heart sound data includes a plurality of heart sound data segments;

[0133] The data quality control unit 502 is configured to perform quality assessment on the heart sound data through a heart sound data quality assessment module, and divide the plurality of heart sound data segments into non-interference heart sound data segments, recoverable heart sound data segments, and non-recoverable heart sound data segments, wherein the heart sound data quality assessment module includes a complex-valued coding submodule; perform signal recovery on the recoverable heart sound data segments through a heart sound signal recovery module that shares the complex-valued coding submodule with the heart sound data quality assessment module, and output a recovered heart sound data segment, wherein the non-interference heart sound data segment and the recovered heart sound data segment constitute an available heart sound data segment;

[0134] A shock classification unit 503, used to take each of the available heart sound data segments as input to obtain a corresponding shock classification result, wherein the shock classification unit 503 includes a parallel processing time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork;

[0135] The fusion output unit 504 is used to perform majority voting on the plurality of shock classification results to obtain a final shock automatic classification result.

[0136] In a possible implementation, the heart sound data quality assessment module is used to:

[0137] For any target heart sound data segment in the heart sound data, transform the target heart sound data segment into the time-frequency domain to obtain a corresponding target complex time-frequency spectrum matrix;

[0138] The target complex time-frequency spectrum matrix is ​​used as the input of the complex-valued coding submodule, and the corresponding target complex-valued coding features are output. The complex-valued coding submodule includes: a complex two-dimensional convolutional layer, S levels of complex Transformer and downsampling, and a complex Transformer;

[0139] The absolute value of the target complex-valued coding feature is taken as input, and is sequentially passed through a fully connected layer and ReLU, a fully connected layer and Softmax to divide the target heart sound data segment into one of an undisturbed heart sound data segment, a recoverable heart sound data segment and an irrecoverable heart sound data segment.

[0140] In a possible implementation, the heart sound signal recovery module is used to:

[0141] In the case where the target heart sound data segment is divided into the recoverable heart sound data segment, the target complex-valued coding feature output by the complex-valued coding submodule of the target heart sound data segment is reused, the target complex-valued coding feature is used as the input of the complex-valued decoding submodule, and the corresponding target complex-valued decoding feature is output;

[0142] transforming the target complex-valued decoding feature into the time domain to obtain the restored heart sound data segment;

[0143] The complex-valued decoding submodule includes, in sequence: S levels of upsampling and complex Transformer, and a complex two-dimensional convolutional layer.

[0144] In a possible implementation, the input of the complex Transformer of the kth level of the complex-valued decoding submodule is: the concatenation of the upsampling result of the kth level and the output result of the complex Transformer of the S-k+1th level in the complex-valued encoding submodule.

[0145] In one possible implementation, the complex Transformer is composed of a cascade of a multi-head self-attention submodule for complex numbers and a gated feedforward submodule for complex numbers.

[0146] In a possible implementation, the shock classification unit 503 is used to:

[0147] For any target available heart sound data segment in the available heart sound data, the target available heart sound data segment is respectively input into a time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork to obtain a time-frequency domain extraction feature and a time-domain extraction feature;

[0148] After the time-frequency domain extracted features and the time domain extracted features are input into a feature fusion submodule, they are sequentially passed through a fully connected layer and ReLU, a fully connected layer and Softmax to obtain the shock classification result corresponding to the target available heart sound data segment.

[0149] In a possible implementation, the time-frequency domain Transformer feature extraction subnetwork includes: a cascaded short-time Fourier transform module, a multi-head self-attention module, and a gated feedforward module;

[0150] The time-domain Transformer feature extraction subnetwork includes: a cascaded one-dimensional convolution layer, Q stacked hole convolution residual modules, a multi-head self-attention module and a gated feedforward module.

[0151] In a possible implementation, the feature fusion submodule is used to:

[0152] Performing layer normalization and 1×1 convolution processing on the time-frequency domain extraction features and the time-domain extraction features, respectively, to obtain time-frequency domain intermediate features and time-domain intermediate features;

[0153] Obtaining the time-frequency domain intermediate features and the attention weight matrix of each of the time-domain intermediate features;

[0154] Using an attention interaction mechanism, the time-frequency domain intermediate features, the time domain intermediate features, and two attention weight matrices are operated to obtain fused features;

[0155] The fused features are sequentially passed through a depth-wise separable convolutional layer and ReLU, layer normalization, and the feature fusion result is output.

[0156] It should be noted that the shock classification device based on heart sound analysis provided in the above embodiment is only illustrated by the division of the above functional modules. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the device and method embodiments provided in the above embodiment belong to the same concept, and the specific implementation process is detailed in the method embodiment, which will not be repeated here.

[0157] See also Figure 6 , which is a schematic diagram of a computer device provided according to an exemplary embodiment of the present application, the computer device may be a portable heart sound device, the computer device includes a memory and a processor, the memory is used to store a computer program, and when the computer program is executed by the processor, the above-mentioned shock classification method based on heart sound analysis is implemented.

[0158] The processor may be a central processing unit (CPU). The processor may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or a combination of the above chips.

[0159] As a non-transitory computer-readable storage medium, the memory can be used to store non-transitory software programs, non-transitory computer executable programs and modules, such as program instructions / modules corresponding to the method in the embodiment of the present invention. The processor executes various functional applications and data processing of the processor by running the non-transitory software programs, instructions and modules stored in the memory, that is, implementing the method in the above method embodiment.

[0160] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application required for at least one function; the data storage area may store data created by the processor, etc. In addition, the memory may include a high-speed random access memory, and may also include a non-transitory memory, such as at least one disk storage device, a flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory may optionally include a memory remotely disposed relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0161] In an exemplary embodiment, a computer-readable storage medium is also provided, which is used to store at least one computer program, and the at least one computer program is loaded and executed by a processor to implement all or part of the steps in the above method. For example, the computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, etc.

[0162] Those skilled in the art will readily appreciate other embodiments of the present application after considering the specification and practicing the invention disclosed herein. The present application is intended to cover any variations, uses or adaptations of the present application, which follow the general principles of the present application and include common knowledge or customary techniques in the art that are not disclosed in the present application. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present application are indicated by the following claims.

[0163] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A shock classification method based on heart sound analysis, characterized in that: The method comprises: Acquiring preprocessed heart sound data, wherein the heart sound data includes a plurality of heart sound data segments; Performing quality assessment on the heart sound data by a heart sound data quality assessment module, dividing the plurality of heart sound data segments into interference-free heart sound data segments, recoverable heart sound data segments and irrecoverable heart sound data segments, the heart sound data quality assessment module comprising a complex-valued coding submodule; The step of performing quality assessment on the heart sound data by using a heart sound data quality assessment module and dividing the plurality of heart sound data segments into non-interference heart sound data segments, recoverable heart sound data segments and irrecoverable heart sound data segments comprises: For any target heart sound data segment in the heart sound data, transform the target heart sound data segment into the time-frequency domain to obtain a corresponding target complex time-frequency spectrum matrix; The target complex time-frequency spectrum matrix is ​​used as the input of the complex-valued coding submodule, and the corresponding target complex-valued coding features are output. The complex-valued coding submodule includes: a complex two-dimensional convolution layer, a complex Transformer of layers and downsampling, and a complex Transformer; Taking the absolute value of the target complex-valued encoding feature as input, sequentially passing through a fully connected layer and ReLU, a fully connected layer and Softmax, the target heart sound data segment is divided into one of an undisturbed heart sound data segment, a recoverable heart sound data segment and an irrecoverable heart sound data segment; The heart sound signal recovery module of the complex-valued coding submodule is shared with the heart sound data quality assessment module to perform signal recovery on the recoverable heart sound data segment and output a recovered heart sound data segment, wherein the interference-free heart sound data segment and the recovered heart sound data segment constitute an available heart sound data segment; The heart sound signal recovery module that shares the complex-valued encoding submodule with the heart sound data quality assessment module performs signal recovery on the recoverable heart sound data segment and outputs the recovered heart sound data segment, including: In the case where the target heart sound data segment is divided into the recoverable heart sound data segment, the target complex-valued coding feature output by the complex-valued coding submodule of the target heart sound data segment is reused, the target complex-valued coding feature is used as the input of the complex-valued decoding submodule, and the corresponding target complex-valued decoding feature is output; transforming the target complex-valued decoding feature into the time domain to obtain the restored heart sound data segment; Inputting each of the available heart sound data segments into a shock classification unit to obtain a corresponding shock classification result, wherein the shock classification unit includes a parallel processing time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork; A majority vote is performed on the plurality of shock classification results to obtain a final automatic shock classification result.

2. The method according to claim 1, characterized in that: The complex-valued decoding submodule includes: upsampling and complex Transformer at each level, and a complex two-dimensional convolutional layer.

3. The method according to claim 2, characterized in that The input of the complex Transformer of the th level in the complex-valued decoding submodule is: the concatenation of the upsampling result of the th level and the output result of the complex Transformer of the th level in the complex-valued encoding submodule.

4. The method according to claim 1 or 2, characterized in that: The complex Transformer is composed of a cascade of a multi-head self-attention submodule for complex numbers and a gated feedforward submodule for complex numbers.

5. The method according to claim 1, characterized in that The step of inputting each of the available heart sound data segments into the shock classification unit to obtain a corresponding shock classification result includes: For any target available heart sound data segment in the available heart sound data, the target available heart sound data segment is respectively input into a time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork to obtain a time-frequency domain extraction feature and a time-domain extraction feature; After the time-frequency domain extracted features and the time domain extracted features are input into a feature fusion submodule, they are sequentially passed through a fully connected layer and ReLU, a fully connected layer and Softmax to obtain the shock classification result corresponding to the target available heart sound data segment.

6. The method according to claim 5, characterized in that The time-frequency domain Transformer feature extraction sub-network includes: a cascaded short-time Fourier transform module, a multi-head self-attention module and a gated feedforward module; The time-domain Transformer feature extraction subnetwork includes: a cascaded one-dimensional convolution layer, a stacked hole convolution residual module, a multi-head self-attention module and a gated feedforward module.

7. The method according to claim 5, characterized in that The processing process of the feature fusion submodule includes: Performing layer normalization and 1×1 convolution processing on the time-frequency domain extraction features and the time-domain extraction features, respectively, to obtain time-frequency domain intermediate features and time-domain intermediate features; Obtaining the time-frequency domain intermediate features and the attention weight matrix of each of the time-domain intermediate features; Using an attention interaction mechanism, the time-frequency domain intermediate features, the time domain intermediate features, and two attention weight matrices are operated to obtain fused features; The fused features are sequentially passed through a depth-wise separable convolutional layer and ReLU, layer normalization, and the feature fusion result is output.

8. A shock classification device based on heart sound analysis, characterized in that: The device comprises: A data preprocessing unit, used to obtain preprocessed heart sound data, wherein the heart sound data includes a plurality of heart sound data segments; A data quality control unit is used to perform quality assessment on the heart sound data through a heart sound data quality assessment module, and divide the multiple heart sound data segments into interference-free heart sound data segments, recoverable heart sound data segments and irrecoverable heart sound data segments, wherein the heart sound data quality assessment module includes a complex-valued coding submodule; the quality assessment of the heart sound data through the heart sound data quality assessment module, and the multiple heart sound data segments are divided into interference-free heart sound data segments, recoverable heart sound data segments and irrecoverable heart sound data segments, including: for any target heart sound data segment in the heart sound data, transform the target heart sound data segment into the time-frequency domain to obtain a corresponding target complex time-frequency spectrum matrix; use the target complex time-frequency spectrum matrix as the input of the complex-valued coding submodule, and output the corresponding target complex-valued coding feature, wherein the complex-valued coding submodule sequentially includes: a complex two-dimensional convolutional layer, a complex Transformer of layers and downsampling, and a complex Transformer; use the absolute value of the target complex-valued coding feature as input, and sequentially pass through a fully connected layer and R eLU, a fully connected layer and Softmax, divide the target heart sound data segment into one of a non-interference heart sound data segment, a recoverable heart sound data segment and an irrecoverable heart sound data segment; perform signal recovery on the recoverable heart sound data segment through a heart sound signal recovery module that shares the complex-valued coding submodule with the heart sound data quality assessment module, and output a recovered heart sound data segment, wherein the non-interference heart sound data segment and the recovered heart sound data segment constitute a usable heart sound data segment; perform signal recovery on the recoverable heart sound data segment through a heart sound signal recovery module that shares the complex-valued coding submodule with the heart sound data quality assessment module, and output a recovered heart sound data segment, including: when the target heart sound data segment is divided into the recoverable heart sound data segment, reuse the target complex-valued coding feature output by the complex-valued coding submodule of the target heart sound data segment, use the target complex-valued coding feature as the input of the complex-valued decoding submodule, and output the corresponding target complex-valued decoding feature; transform the target complex-valued decoding feature to the time domain to obtain the recovered heart sound data segment; A shock classification unit, used to take each of the available heart sound data segments as input to obtain a corresponding shock classification result, wherein the shock classification unit includes a parallel processing time-frequency domain Transformer feature extraction subnetwork and a time-domain Transformer feature extraction subnetwork; The fusion output unit is used to perform majority voting on the plurality of shock classification results to obtain a final shock automatic classification result.

9. A computer device, characterized in that: The computer device comprises a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by the processor to implement the shock classification method based on heart sound analysis according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, at least one program, a code set or an instruction set is loaded and executed by a processor to implement the shock classification method based on heart sound analysis according to any one of claims 1 to 7.

Images