Snore classification method and system based on multi-dimensional sleep sign signal feature map

Abstract

The invention provides a snore classification method and system based on a multi-dimensional sleep sign signal feature graph, and relates to the technical field of data processing, the method comprises the following steps: preprocessing an all-night sleep acoustic signal, an oxyhemoglobin saturation signal and a pulse rate signal which are synchronously collected; based on the preprocessed acoustic signal, a plurality of independent snore event segments are positioned and cut by detecting the envelope peak value of the acoustic signal and combining a double-threshold endpoint detection method; based on each snore event segment, Mel frequency cepstral coefficient features, time domain features and frequency domain features are extracted from the corresponding acoustic signals to form an acoustic feature set, and fluctuation features and extreme value features are synchronously extracted from the corresponding oxyhemoglobin saturation signals and pulse rate signals to form a physiological feature set. According to the method, high-precision automatic classification of pathological snore and normal snore is realized.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a snoring classification method and system based on multidimensional sleep symptom signal feature maps. Background Technology

[0002] Snoring is an important indicator of sleep quality and a key characteristic of sleep-disordered breathing. Long-term severe snoring can easily lead to sleep apnea-hypopnea syndrome. In existing snoring classification technologies, methods dominated by single acoustic features cannot associate snoring with key pathophysiological features of OSAHS patients, such as decreased blood oxygenation and pulse rate fluctuations. Some multimodal fusion attempts only achieve simple signal splicing without efficient feature selection. Image-based feature methods do not incorporate the image features of physiological signals, and traditional convolutional neural network models have poor adaptability to specific sleep snoring data.

[0003] Taking an OSAHS patient in the research dataset as an example, the patient was a middle-aged male with a BMI exceeding the normal range for healthy adults. He was clinically diagnosed with moderate OSAHS. During sleep, his snoring amplitude was significantly higher than that of healthy individuals, and he exhibited obvious sleep apnea. During sleep apnea, his blood oxygen saturation dropped significantly, while his pulse rate increased significantly at the end of the apnea, with fluctuations far exceeding the normal range of pulse rate fluctuations in healthy individuals at night. When classifying the patient's snoring segments using a single acoustic feature combined with a traditional classifier, some pathological snoring segments accompanied by decreased blood oxygen were misclassified as normal snoring, while some normal snoring segments were also misclassified. In cases where snoring is misdiagnosed as pathological, the classification results fail to meet the requirements for accurate identification. The method has technical flaws. On the one hand, it relies solely on a single acoustic feature without integrating physiological signal features such as blood oxygen saturation and pulse rate, which are directly related to the pathological mechanism of OSAHS. This makes it impossible to capture the pathological correlation pattern between snoring and physiological indicators, resulting in insufficient characterization of the essential features of pathological snoring. On the other hand, the extracted raw features are not effectively screened, and redundant features can easily amplify irrelevant differences, interfering with the classifier's judgment. In addition, the traditional classifier used lacks complex feature mapping and nonlinear fitting capabilities, ultimately leading to missed diagnoses and misdiagnoses of pathological snoring. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a snoring classification method and system based on multidimensional sleep sign signal feature maps, so as to achieve high-precision automatic classification of pathological snoring and normal snoring.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a snoring classification method based on multidimensional sleep symptom signal feature maps, the method comprising: The acoustic signals, blood oxygen saturation signals, and pulse rate signals collected throughout the night were preprocessed. Based on the preprocessed acoustic signals, multiple independent snoring event segments were located and segmented by detecting the peak value of the acoustic signal envelope and combining it with the dual-threshold endpoint detection method. Based on each snoring event segment, Mel frequency cepstral coefficient features, time domain features, and frequency domain features are extracted from the corresponding acoustic signals to form an acoustic feature set. Simultaneously, fluctuation features and extreme value features are extracted from the corresponding blood oxygen saturation signals and pulse rate signals to form a physiological feature set. The obtained acoustic feature set and physiological feature set are fused to obtain an initial fused feature set; the Relief-F algorithm is used to calculate and filter the feature weights of the initial fused feature set to obtain the final feature subset. The values ​​in the final feature subset are normalized and mapped to grayscale image pixels to obtain the initial grayscale feature map. The horizontal and vertical gradients of the initial grayscale feature map are calculated and the initial grayscale feature map and the gradient maps in the two directions are assigned to the red, green and blue channels respectively to obtain a three-channel feature texture image. The feature texture image is input into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model ultimately determines whether the snoring event is normal or pathological snoring.

[0006] Furthermore, the synchronously acquired acoustic signals, blood oxygen saturation signals, and pulse rate signals from the entire night's sleep were preprocessed. Based on the preprocessed acoustic signals, multiple independent snoring event segments were located and segmented by detecting the peak value of the acoustic signal envelope and combining it with a dual-threshold endpoint detection method, including: Acoustic signals, blood oxygen saturation signals, and pulse rate signals from the entire night's sleep were collected and uploaded, and the signals were parsed into actionable time-series data. The acoustic signal in the time series data is downsampled to obtain a downsampled acoustic signal with reduced computational complexity; Based on a preset physiological range threshold, outlier filtering is performed on the blood oxygen saturation signal and pulse rate signal in the time series data to obtain clean physiological signal data. Based on downsampled acoustic signals and clean physiological signal data, according to their respective start and end timestamps, the same valid data segment where all signals completely overlap in time is extracted; Based on the time-aligned downsampled acoustic signal, the envelope of the downsampled acoustic signal is calculated and the peak point is detected. Combined with the dual-threshold endpoint detection method, the start and end times of each snoring event are determined, and multiple independent snoring event segments are synchronously cut out from the same valid data segment accordingly.

[0007] Furthermore, based on each snoring event segment, Mel-frequency cepstral coefficient features, time-domain features, and frequency-domain features are extracted from the corresponding acoustic signal to form an acoustic feature set. Simultaneously, fluctuation features and extreme value features are extracted from the corresponding blood oxygen saturation and pulse rate signals to form a physiological feature set, including: Based on the acoustic signals in each independent snoring event segment obtained by segmentation, pre-emphasis, framing and windowing preprocessing operations are performed to obtain standardized acoustic frame signals; The standardized acoustic frame signal is subjected to a fast Fourier transform and converted to the frequency domain. The energy of the filter bank is extracted through the Mel filter bank, and then the Mel frequency cepstral coefficient characteristics are calculated through logarithmic operation and discrete cosine transform. Based on the standardized acoustic frame signal, the statistical characteristics of short-time energy and short-time zero-crossing rate are calculated to form a time-domain feature subset, and the spectral centroid, spectral entropy and spectral slope are calculated to form a frequency-domain feature subset. The obtained Mel frequency cepstral coefficient features are combined with the obtained time-domain feature subset and frequency-domain feature subset to form the acoustic feature set of snoring event segments. Based on the blood oxygen saturation signal and pulse rate signal corresponding to the same snoring event segment obtained by segmentation, the numerical fluctuation in the three time intervals of the first, middle and last time intervals is calculated as the fluctuation feature, and the minimum and maximum values ​​in the signal interval are extracted as the extreme value features to form the physiological feature set of the snoring event segment.

[0008] Furthermore, the obtained acoustic feature set and physiological feature set are fused to obtain an initial fused feature set; the Relief-F algorithm is then used to calculate and filter feature weights from the initial fused feature set to obtain the final feature subset, including: The obtained acoustic feature set and the obtained physiological feature set are concatenated and combined to obtain the initial fusion feature set; Based on the initial fused feature set, initialize the feature weight array of the Relief-F algorithm, and set the sampling number and the number of nearest neighbors parameters; Under the control of the set parameters, multiple rounds of random sampling are performed on the initial fused feature set. In each round of sampling, the feature difference between the nearest neighbor sample within each category and the nearest neighbor sample between different categories is calculated to obtain the feature difference result. Based on the obtained feature difference results, the weight of each feature in the feature weight array is dynamically updated, and features with weights higher than the preset threshold are selected to form the final feature subset.

[0009] Furthermore, the values ​​in the final feature subset are normalized, and the normalized values ​​are mapped to grayscale image pixels to obtain an initial grayscale feature map. The horizontal and vertical gradients of the initial grayscale feature map are calculated, and the initial grayscale feature map and the gradient maps in the two directions are assigned to the red, green, and blue channels, respectively, to obtain a three-channel feature texture image, including: Normalize all eigenvalues ​​in the final feature subset to map the linear values ​​to the integer range, and obtain the normalized feature matrix. Each feature value in the obtained normalized feature matrix is ​​used as the gray value of the pixel and arranged according to a preset arrangement rule to obtain the initial gray-scale feature map. Based on the initial grayscale feature map, the gradient matrix in the horizontal direction and the gradient matrix in the vertical direction are calculated respectively. The initial grayscale feature map, the horizontal gradient matrix, and the vertical gradient matrix are mapped to pixel values ​​in the red, green, and blue channels, respectively, to obtain three single-channel images. The three single-channel images are merged to generate the final three-channel feature texture image.

[0010] Furthermore, the feature texture image is input into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model ultimately determines whether the snoring event is normal or pathological, including: The generated three-channel feature texture image is input into the input layer of a pre-trained convolutional neural network model; By using the convolutional layers connected sequentially in the convolutional neural network model, convolution operations are performed on the image to extract multi-layered abstract features; The feature map output from the convolutional layer is input into the pooling layer in the convolutional neural network model for downsampling to obtain high-level features with translation invariance. The high-level features are flattened into a one-dimensional feature vector, which is then input into the fully connected layer of the convolutional neural network model for nonlinear transformation to obtain the feature vector output by the fully connected layer. The feature vectors output by the fully connected layer are used by the Softmax classifier to calculate the final classification result of whether the snoring event belongs to normal snoring or pathological snoring.

[0011] Furthermore, the generated three-channel feature texture image is input into the input layer of a pre-trained convolutional neural network model, including: Construct the basic network architecture and establish a network topology that includes an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer; Based on the constructed basic network architecture, the initial convolutional layer is configured and then connected sequentially to the normalization processing layer and the non-linear activation layer to obtain the feature extraction backbone network. Based on the constructed feature extraction backbone network, a parallel convolution branch is added, and the branch output is fused with the backbone network output to construct a residual connection device. Based on the output of the constructed residual connection device, a feature downsampling layer is configured to reduce the size of the feature map and obtain the processed feature map. Based on the processed feature maps, a multi-layer fully connected network is constructed and a classifier is connected at the end to obtain the complete network structure. Based on the complete network structure, a labeled feature texture image training set is used, and the network parameters are iteratively optimized through the backpropagation algorithm to obtain the final usable convolutional neural network model.

[0012] Secondly, a snoring classification system based on multidimensional sleep symptom signal feature maps includes: The acquisition module is used to preprocess the synchronously acquired acoustic signals of the entire night's sleep, blood oxygen saturation signals, and pulse rate signals; based on the preprocessed acoustic signals, multiple independent snoring event segments are located and segmented by detecting the peak value of the acoustic signal envelope and combining it with the dual-threshold endpoint detection method. The extraction module is used to extract Mel frequency cepstral coefficient features, time domain features and frequency domain features from the corresponding acoustic signals based on each snoring event segment to form an acoustic feature set, and simultaneously extract fluctuation features and extreme value features from the corresponding blood oxygen saturation signal and pulse rate signal to form a physiological feature set. The fusion module is used to fuse the obtained acoustic feature set and physiological feature set to obtain an initial fused feature set; the Relief-F algorithm is used to calculate and filter the feature weights of the initial fused feature set to obtain the final feature subset. The calculation module is used to normalize the values ​​in the final feature subset and map the normalized values ​​to grayscale image pixels to obtain the initial grayscale feature map; calculate the horizontal and vertical gradients of the initial grayscale feature map and assign the initial grayscale feature map and the gradient maps in the two directions to the red, green and blue channels respectively to obtain a three-channel feature texture image. The determination module is used to input the feature texture image into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model ultimately determines whether the snoring event is normal snoring or pathological snoring.

[0013] Thirdly, a computing device, comprising: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0014] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0015] The above-described solution of the present invention has at least the following beneficial effects: This method employs a technique that simultaneously acquires acoustic signals, blood oxygen saturation signals, and pulse rate signals throughout the night, preprocesses them to locate and segment independent snoring event segments, and extracts acoustic feature sets composed of Mel frequency cepstral coefficient features, time domain features, and frequency domain features from the snoring event segments, as well as physiological feature sets composed of fluctuation features and extreme value features. After fusing the two types of features, the Relief-F algorithm is used to calculate weights and filter to obtain the final feature subset. The final feature subset is then normalized and mapped to a grayscale feature map, and the horizontal and vertical gradients are calculated and assigned to red, green, and blue channels to generate a three-channel feature texture image. Finally, the image is input into a pre-trained convolutional neural network model with residual connections for classification. This technique overcomes the technical problems of insufficient single-modal feature representation, inadequate multimodal feature fusion, high-dimensional redundant features interfering with classification accuracy, and the lack of adaptability of traditional convolutional neural networks, which make it difficult to achieve high-precision automatic classification of pathological snoring and normal snoring in existing snoring classification methods. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the snoring classification method based on multidimensional sleep sign signal feature maps provided in an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a snoring classification system based on multidimensional sleep sign signal feature maps provided in an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0019] like Figure 1 As shown, embodiments of the present invention propose a snoring classification method based on multidimensional sleep symptom signal feature maps, the method comprising the following steps: Step 1: Preprocess the acoustic signals, blood oxygen saturation signals, and pulse rate signals collected synchronously throughout the night; based on the preprocessed acoustic signals, locate and segment multiple independent snoring event segments by detecting the peak value of the acoustic signal envelope and combining it with the dual-threshold endpoint detection method. Step 2: Based on each snoring event segment, extract the Mel frequency cepstral coefficient features, time domain features and frequency domain features from the corresponding acoustic signal to form an acoustic feature set, and simultaneously extract the fluctuation features and extreme value features from the corresponding blood oxygen saturation signal and pulse rate signal to form a physiological feature set. Step 3: Fuse the obtained acoustic feature set with the physiological feature set to obtain an initial fused feature set; use the Relief-F algorithm to calculate and filter the feature weights of the initial fused feature set to obtain the final feature subset; Step 4: Normalize the values ​​in the final feature subset and map the normalized values ​​to grayscale image pixels to obtain the initial grayscale feature map; calculate the horizontal and vertical gradients of the initial grayscale feature map and assign the initial grayscale feature map and the gradient maps in the two directions to the red, green and blue channels respectively to obtain a three-channel feature texture image. Step 5: Input the feature texture image into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model finally determines whether the snoring event is normal snoring or pathological snoring.

[0020] In this embodiment of the invention, acoustic signals, blood oxygen saturation signals, and pulse rate signals from a full night of sleep are simultaneously acquired and preprocessed. Independent snoring event segments are located and segmented by detecting the envelope peak of the acoustic signals and using a dual-threshold endpoint detection method. From each snoring event segment, an acoustic feature set containing Mel-frequency cepstral coefficients, time-domain features, and frequency-domain features is extracted. Simultaneously, a physiological feature set containing blood oxygen saturation and pulse rate fluctuation features, as well as extreme value features, is extracted. The two feature sets are fused into an initial fused feature set, and the Relief-F algorithm is used to calculate feature weights and filter to obtain a final feature subset. The final feature subset is then normalized and mapped to grayscale values. This technique involves identifying and calculating the horizontal and vertical gradients of the grayscale feature map and the two gradient maps, assigning them to the red, green, and blue channels respectively to generate a three-channel feature texture image. This feature texture image is then input into a pre-trained convolutional neural network model for forward propagation calculation. Therefore, it overcomes the technical problems of existing snoring classification methods, such as incomplete feature representation due to reliance on a single acoustic signal, ineffective fusion of multimodal features, high-dimensional redundant features interfering with classification accuracy, lack of feature differentiation through image processing, and insufficient adaptability of the classification model, making it difficult to accurately distinguish between normal and pathological snoring. This technique achieves the technical effect of automatically classifying normal and pathological snoring.

[0021] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1 involves collecting and uploading the acoustic signals, blood oxygen saturation signals, and pulse rate signals from the entire night's sleep, and then parsing these signals into operable time-series data. Specifically, this includes: first, preparing the signal acquisition equipment. The acoustic signals are acquired using a wireless Bluetooth headset, which is fixed one centimeter to the side of the user's nose to obtain acoustic signals throughout the night's sleep. The blood oxygen saturation and pulse rate signals are acquired using a portable pulse oximeter, worn on the user's fingertip, simultaneously collecting these signals throughout the night's sleep. During the acquisition process, the wireless Bluetooth headset and portable pulse oximeter establish communication connections with the mobile terminal, uploading the acquired acoustic signals, blood oxygen saturation signals, and pulse rate signals to the mobile terminal in real time. Subsequently, on the mobile terminal or a connected computer, a signal parsing tool is used to process the uploaded raw signals, converting the raw acoustic signals, blood oxygen saturation signals, and pulse rate signals into operable data with time-series attributes, i.e., time-series data. Each time-series data point corresponds to a unique acquisition time information.

[0022] Step 1.2 involves downsampling the acoustic signal in the time-series data to obtain a downsampled acoustic signal with reduced computational complexity. Specifically, this includes: first, determining the original sampling frequency of the acoustic signal; then, selecting a suitable downsampling target frequency based on the original sampling frequency. The target frequency must preserve the key frequency characteristics of the snoring signal while reducing the computational complexity of subsequent data processing. Next, a professional signal processing tool is used to downsample the obtained acoustic signal time-series data. During processing, a signal resampling algorithm is used to adjust the sampling frequency of the acoustic signal from its original value to the target frequency. After processing, it is ensured that the key characteristic information of the snoring is not lost during the downsampling process, ultimately obtaining a downsampled acoustic signal with reduced computational complexity that meets the requirements of subsequent analysis.

[0023] Step 1.3: Based on preset physiological range thresholds, outlier filtering is performed on the blood oxygen saturation and pulse rate signals in the time series data to obtain clean physiological signal data. Specifically, this includes: First, referring to the normal range of physiological indicators for healthy adults and combining clinical research data related to sleep-disordered breathing, setting physiological range thresholds for blood oxygen saturation and pulse rate signals. The threshold for blood oxygen saturation is set to not exceed the normal upper limit of blood oxygen saturation for healthy adults, and the threshold for pulse rate is set to not exceed the reasonable upper limit of pulse rate during sleep in healthy adults. Then, the obtained time series data of blood oxygen saturation and pulse rate signals are iterated, and each data point is judged to see if it exceeds the preset physiological range threshold. Data points that exceed the threshold are judged as outliers and removed from the time series data. Data points that do not exceed the threshold are retained in the time series data. After completing the filtering of all data points, the remaining time series data of blood oxygen saturation and pulse rate signals are processed to obtain clean physiological signal data without outlier interference.

[0024] Step 1.4: Based on the downsampled acoustic signal and clean physiological signal data, according to their respective start and end timestamps, extract the same effective data segment where all signals completely overlap in time. Specifically, this includes: extracting the start and end timestamps of the acquisition process from the obtained downsampled acoustic signal time series data; and simultaneously extracting the corresponding start and end timestamps from the obtained clean oxygen saturation signal time series data and clean pulse rate signal time series data. Comparing the start timestamps of the three sets of signals, the latest start timestamp is selected as the unified start time of the effective data segment; comparing the end timestamps of the three sets of signals, the earliest end timestamp is selected as the unified end time of the effective data segment; based on the determined start and end times of the effective data segment, the downsampled acoustic signal, clean oxygen saturation signal, and clean pulse rate signal are extracted respectively, retaining the data within the time interval of the three sets of signals, ultimately obtaining the same effective data segment where all signals completely overlap in time.

[0025] Step 1.5: Based on the time-aligned downsampled acoustic signal, calculate the envelope of the downsampled acoustic signal and detect peak points; combine the dual-threshold endpoint detection method to determine the start and end times of each snoring event, and accordingly, synchronously cut out multiple independent snoring event segments from the same valid data segment. Specifically, this includes: using a signal envelope extraction algorithm to process the obtained time-aligned downsampled acoustic signal, generating an envelope curve of the acoustic signal by calculating the amplitude envelope of the signal; based on the generated envelope curve, using a peak detection algorithm to identify peak points on the envelope curve, recording the time information corresponding to each peak point, and then using the dual-threshold endpoint detection method to process the downsampled acoustic signal. When the signal-to-noise ratio of the signal is at a low level, the start and end of the snoring event are determined by calculating the short-time zero-crossing rate of the signal; when the signal-to-noise ratio of the signal is at a high level, the start and end of the snoring event are determined by calculating the short-time energy of the signal. By combining the time information of the peak point of the envelope curve with the judgment results of the dual-threshold endpoint detection method, the accurate start and end times of each snoring event are determined. Based on the start and end times of each snoring event, the corresponding downsampled acoustic signal segment, blood oxygen saturation signal segment, and pulse rate signal segment are simultaneously cut from the same valid data segment. The three signal segments are combined into an independent snoring event segment. This process is repeated until all snoring events are cut, resulting in multiple independent snoring event segments.

[0026] In this embodiment of the invention, the acquisition and analysis of acoustic signals, blood oxygen saturation signals, and pulse rate signals throughout the night are used as operable time-series data. The acoustic signals are downsampled, and outliers in the blood oxygen saturation and pulse rate signals are filtered based on preset physiological range thresholds. Effective data segments with completely overlapping times are extracted according to the timestamps of each signal. Then, the start and end times of snoring events are determined by calculating the acoustic signal envelope, detecting peak points, and combining a dual-threshold endpoint detection method to synchronously cut independent snoring event segments. Therefore, this method overcomes the technical problems of high computational complexity of the original acoustic signals, interference of physiological signals by outliers, inability to accurately correlate different types of signals due to time asynchrony, and difficulty in separating independent segments due to ambiguous snoring event positioning. This achieves the technical effects of reducing signal processing computational costs, obtaining clean and time-aligned sleep sign data, and accurately locating and cutting out multiple independent snoring event segments.

[0027] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1: Based on the acoustic signals in each independent snoring event segment obtained from the cut-out segment, pre-emphasis, framing, and windowing preprocessing operations are performed to obtain standardized acoustic frame signals. Specifically, this includes: First, pre-emphasis processing is performed on the acoustic signals in each independent snoring event segment obtained from the cut-out segment. Since high-frequency components of acoustic signals are easily attenuated during propagation, each data point of the signal is processed by linear filtering to increase the amplitude of the high-frequency part, making the signal characteristics of different frequency components more balanced and avoiding the masking of snoring details in the high-frequency band. Next, framing processing is performed. According to the time length of the snoring signal, the acoustic signal is divided into multiple continuous time segments. An appropriate frame length and frame shift are set. The frame shift is less than the frame length to ensure the continuity between frames and ensure that the signal characteristics of adjacent frames can be connected without feature breaks. Finally, windowing processing is performed. Each segmented frame is multiplied by a smooth window function to reduce the spectral leakage problem caused by signal abrupt changes at frame boundaries, making the signal amplitude change within each frame smoother. After these three preprocessing steps, a standardized acoustic frame signal that conforms to a unified standard is obtained.

[0028] Step 2.2 involves performing a Fast Fourier Transform on the standardized acoustic frame signal, converting it to the frequency domain, extracting the filter bank energy through a Mel filter bank, and then calculating the Mel frequency cepstral coefficient characteristics through logarithmic operations and discrete cosine transform. Specifically, this includes processing each of the obtained standardized acoustic frame signals individually. First, a Fast Fourier Transform (FFT) is performed on each standardized acoustic frame signal to convert the time-domain signal into a frequency-domain signal, obtaining the spectral amplitude distribution information corresponding to each frame and clearly presenting the energy distribution of the signal at different frequencies. Then, the converted frequency-domain signal is input into a pre-constructed Mel filter bank. The frequency interval of this filter bank is set to conform to the human ear's perception of sound frequency, focusing on capturing the snoring characteristics within the frequency range that the human ear is sensitive to. The signal energy within the frequency interval corresponding to each filter is extracted through the filter bank, resulting in a set of filter bank energy values ​​reflecting the energy of different frequency intervals. Subsequently, logarithmic operations are performed on the extracted filter bank energy values ​​to simulate the logarithmic perception characteristics of sound intensity by the human ear, converting the energy values ​​into logarithmic energy, making the energy characteristics more consistent with the human auditory perception logic. Finally, a Discrete Cosine Transform (DCT) is performed on the logarithmic energy results to compress the high-dimensional logarithmic energy data into a low-dimensional feature vector. The low-dimensional feature vector is the Mel frequency cepstral coefficient feature that can accurately reflect the frequency perception characteristics of the snoring signal.

[0029] Step 2.3: Based on the standardized acoustic frame signal, calculate the statistical characteristics of short-time energy and short-time zero-crossing rate to form a time-domain feature subset, and calculate the spectral centroid, spectral entropy, and spectral slope to form a frequency-domain feature subset. Specifically, in terms of time-domain feature calculation, first calculate the short-time energy of each acoustic frame, that is, the sum of the squares of the amplitudes of all signal data points in each frame, and then calculate the statistical characteristics such as the mean, variance, maximum value, and minimum value of the short-time energy of all frames; at the same time, calculate the short-time zero-crossing rate of each acoustic frame, that is, the number of times the signal data points in each frame change from positive amplitude to negative amplitude or from negative amplitude to positive amplitude, and similarly calculate the statistical characteristics such as the mean, variance, maximum value, and minimum value of the short-time zero-crossing rate of all frames. Integrate these statistical characteristics of short-time energy and short-time zero-crossing rate to form a time-domain feature subset that reflects the time-domain amplitude variation law of the snoring signal.

[0030] In terms of frequency domain feature calculation, a fast Fourier transform is first performed on each standardized acoustic frame signal to obtain the frequency domain signal. Then, the spectral centroid is calculated. Based on each frequency point and corresponding amplitude of the frequency domain signal, the frequency position of the concentrated spectral energy is calculated by weighted averaging, reflecting the concentration trend of snoring energy in the frequency domain. The spectral entropy is calculated, and the amplitude ratio of each frequency component of the frequency domain signal is statistically analyzed. The entropy calculation method is used to measure the complexity of the spectral distribution. The higher the complexity, the more diverse the snoring frequency components. The spectral slope is calculated. Based on the amplitude distribution of the frequency domain signal, the asymmetry of the spectral distribution relative to the mean is measured, reflecting the skewness of the spectral shape. The spectral centroid, spectral entropy, spectral slope and their corresponding statistical features are integrated to form a frequency domain feature subset that reflects the frequency domain distribution characteristics of the snoring signal.

[0031] Step 2.4 involves combining the obtained Mel frequency cepstral coefficient features with the obtained time-domain and frequency-domain feature subsets to form an acoustic feature set for snoring event segments. Specifically, this includes: first, organizing the obtained Mel frequency cepstral coefficient features, which capture the frequency characteristics of snoring from the perspective of human ear perception; then, organizing the obtained time-domain and frequency-domain feature subsets, where the time-domain feature subset reflects the amplitude changes and energy fluctuations of snoring in the time dimension, and the frequency-domain feature subset reflects the energy concentration location and component complexity characteristics of snoring in the frequency dimension. The three types of features are then concatenated according to the feature dimensions to ensure that the Mel frequency cepstral coefficient features, time-domain feature subsets, and frequency-domain feature subsets corresponding to each snoring event segment are completely combined to form an acoustic feature set that comprehensively describes the acoustic properties of snoring from three dimensions: time, frequency, and human ear perception.

[0032] Step 2.5: Based on the blood oxygen saturation signal and pulse rate signal corresponding to the same snoring event segment obtained by segmentation, calculate the numerical fluctuation amount in the three time intervals of the front, middle and back as fluctuation features, and extract the minimum and maximum values ​​in the signal interval as extreme value features to form the physiological feature set of the snoring event segment. Specifically, for the blood oxygen saturation signal and pulse rate signal that are completely synchronized with the snoring event segment in time, first divide the total time length of each signal into three consecutive time intervals of the front, middle and back according to equal time proportions. The time length of each interval is the same, ensuring that the three intervals can completely cover the physiological signal time period corresponding to the entire snoring event without missing any physiological change information.

[0033] For the blood oxygen saturation signal, the numerical variation ranges of all data points in the first, middle, and last time intervals are calculated respectively. These three ranges reflect the fluctuation characteristics of blood oxygen saturation at different time periods. At the same time, the minimum and maximum values ​​within the entire blood oxygen saturation signal interval are extracted as extreme value features reflecting extreme changes in blood oxygen saturation. For the pulse rate signal, the same processing method as for the blood oxygen saturation signal is used. The numerical variation ranges within the first, middle, and last time intervals are calculated as fluctuation features, and the minimum and maximum values ​​within the entire signal interval are extracted as extreme value features. Finally, the fluctuation and extreme value features of the blood oxygen saturation signal are integrated with the fluctuation and extreme value features of the pulse rate signal to form a set of physiological features reflecting the changes in physiological indicators during snoring events.

[0034] In this embodiment of the invention, the acoustic signal of each independent snoring event segment is pre-emphasized, framed, and windowed sequentially to obtain a standardized acoustic frame signal. The standardized acoustic frame signal is then subjected to Fast Fourier Transform, Mel filter bank energy extraction, logarithmic operation, and Discrete Cosine Transform to obtain Mel frequency cepstral coefficient features. The short-time energy and short-time zero-crossing rate statistical features of the standardized acoustic frame signal are calculated in parallel to form a time-domain feature subset. The spectral centroid, spectral entropy, and spectral slope are calculated to form a frequency-domain feature subset, which is then combined with the Mel frequency cepstral coefficient features to form an acoustic feature set. Simultaneously, the blood oxygen saturation of the same snoring event segment is also considered. The technique of calculating the numerical fluctuations of the sum and pulse signals in the three time intervals (before, during, and after) as fluctuation features and extracting the minimum and maximum values ​​of the signal intervals as extreme value features to form a set of physiological features overcomes the technical problems in existing technologies, such as poor feature consistency due to non-standard acoustic signal preprocessing, single dimension of acoustic feature extraction which cannot fully reflect the acoustic attributes of snoring, and incomplete feature representation due to the failure to simultaneously extract blood oxygen saturation and pulse rate features associated with pathological snoring, making it difficult to capture the synergistic pathological information of snoring and physiological indicators. Thus, it achieves the goal of obtaining a standardized, multi-dimensional, and comprehensive set of acoustic and physiological features.

[0035] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1: The obtained acoustic feature set and the obtained physiological feature set are concatenated and combined to obtain the initial fusion feature set. Specifically, this includes: collecting the obtained acoustic feature set and the obtained physiological feature set, wherein the acoustic feature set contains multi-dimensional features reflecting the acoustic attributes of snoring, and the physiological feature set contains features reflecting changes in blood oxygen saturation and pulse rate; according to the order of feature dimensions, all features in the acoustic feature set and all features in the physiological feature set are concatenated in sequence to form a high-dimensional feature set containing the two types of features, i.e., the initial fusion feature set.

[0036] Step 3.2: Based on the initial fused feature set, initialize the feature weight array for the Relief-F algorithm and set the sampling number and nearest neighbor count parameters. Specifically, this includes: determining the total number of features based on the obtained initial fused feature set, and creating a feature weight array accordingly. The length of the array is the same as the number of features in the initial fused feature set, and initially setting the value of each element in the array to zero; determining the sampling number for the Relief-F algorithm based on the sample size and feature dimensions of the initial fused feature set. The sampling number needs to be sufficient to cover the feature distribution of different samples, usually determined according to a certain proportion of the total sample size; simultaneously setting the nearest neighbor count, i.e., the number of similar and dissimilar nearest neighbor samples to be found for each sample. This number needs to balance computational efficiency and the accuracy of feature difference assessment, generally selecting 3 to 5 to ensure effective reflection of feature associations between samples. The Relief-F algorithm is an algorithm used to evaluate feature importance; its core purpose is to analyze the ability of features to distinguish between different categories of samples, thus providing a basis for screening. To select effective features, in the application scenario of this invention, the basic idea of ​​the algorithm is to randomly sample multiple times from the initial fusion feature set, i.e., feature samples corresponding to snoring events. For each sample, the algorithm finds the most similar samples in the same category and the most similar samples in different categories. Then, it compares the differences of the sample with these nearest neighbor samples of the same and different categories on each feature. If a feature has small differences among samples of the same category and large differences among samples of different categories, it means that the feature can effectively distinguish different categories and has high importance; otherwise, it has low importance. To achieve this process, the algorithm first creates a weight array corresponding to the number of features, with all initial values ​​of 0, to record the importance of each feature. Then, it sets the number of sampling times and the number of nearest neighbors according to the total number of samples and the number of features, generally 3 to 5, to balance computational efficiency and difference assessment accuracy. Through multiple rounds of sampling, comparison of differences, and dynamic updating of the weight value of each feature in the weight array, the feature with the highest weight is finally considered as the key feature for distinguishing different categories of snoring.

[0037] Step 3.3: Under the control of the set parameters, multiple rounds of random sampling are performed on the initial fusion feature set. In each round of sampling, the feature differences between the nearest neighbor samples within each category and the nearest neighbor samples between different categories are calculated to obtain the feature difference results. Specifically, this includes: under the control of the set sampling number and the number of nearest neighbors parameters, multiple rounds of random sampling are performed on the samples in the initial fusion feature set. Each time, one sample is randomly selected as the target sample. For each target sample, firstly, a set number of nearest neighbor samples are found in the same category, such as all being feature samples corresponding to pathological snoring. By calculating the numerical difference between the target sample and these same-category nearest neighbor samples on each feature, the same-category feature difference is obtained. Then, in different categories, such as the target sample being a pathological snoring feature sample, a set number of nearest neighbor samples are found in the normal snoring feature sample. The numerical difference between the target sample and these different-category nearest neighbor samples on each feature is calculated to obtain the different-category feature difference. After each round of sampling, the same-category feature differences and different-category feature differences of all features are summarized to form the feature difference results. After multiple rounds of sampling, multiple sets of feature difference results are accumulated to comprehensively reflect the ability of different features to distinguish samples of different categories.

[0038] Step 3.4: Based on the obtained feature difference results, dynamically update the weights of each feature in the feature weight array, and select features with weights higher than a preset threshold to form the final feature subset. Specifically, this includes: dynamically adjusting the weight value of each feature in the feature weight array according to the weight update rules of the Relief-F algorithm based on the obtained multi-round feature difference results; increasing the weight of features that can effectively distinguish different categories (i.e., features with large differences between dissimilar features and small differences between features of the same category); and decreasing the weight of features with weak discriminative ability (i.e., features with small differences between dissimilar features and large differences between features of the same category). After multiple rounds of updates, the final weight value of each feature is obtained. Referring to the weight distribution of all features, a weight threshold is set. This threshold should be able to filter out redundant features with low weights while retaining effective features with high weights. Select all features with weights higher than the threshold from the initial fusion feature set and combine these features to form a concise final feature subset with strong discriminative ability.

[0039] In this embodiment of the invention, an initial fusion feature set is formed by concatenating acoustic feature sets and physiological feature sets. Based on the initial fusion feature set, the feature weight array of the Relief-F algorithm is initialized, and the sampling number and nearest neighbor number parameters are set. Under parameter control, multiple rounds of random sampling are performed on the initial fusion feature set, and the feature differences between nearest neighbor samples within the same category and nearest neighbor samples between different categories are calculated in each round. Then, the feature weights are dynamically updated according to the feature difference results, and features with weights higher than a preset threshold are selected to form the final feature subset. Therefore, this method overcomes the technical problems in existing multimodal fusion technologies, which simply concatenate acoustic and physiological signal features without efficiently screening the high-dimensional features after fusion. This results in redundant features amplifying irrelevant differences and effective features being masked, making it difficult to accurately capture the core features related to snoring and pathology. Thus, it achieves deep fusion of acoustic and physiological features, eliminates redundant interference features, retains high-discrimination core features, and obtains a concise final feature subset with strong representational ability.

[0040] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1 involves normalizing all feature values ​​in the final feature subset, mapping them linearly to an integer range to obtain a normalized feature matrix. This process includes: first, traversing all feature values ​​in the final feature subset, counting the maximum and minimum values, and defining the overall numerical range of the feature values; then, selecting a suitable integer range based on the image processing requirements, ensuring the range matches the standard values ​​for image pixels to directly correspond to pixel attributes; and finally, mapping each feature value from its original numerical range to the selected integer range using a linear transformation. During the mapping process, the relative magnitudes of different feature values ​​are maintained to avoid disrupting the original distribution patterns of the features. After mapping all feature values, the mapped integer feature values ​​are organized into a two-dimensional matrix according to the dimensional order of the features in the final feature subset and the arrangement order of the snoring event samples. This matrix is ​​the normalized feature matrix containing key information about the fused features.

[0041] Step 4.2: Each feature value in the obtained normalized feature matrix is ​​used as the gray value of a pixel and arranged according to a preset arrangement rule to obtain an initial grayscale feature map. Specifically, this includes: first, determining the preset arrangement rule, which must ensure that the correlation between acoustic features and physiological features in the final feature subset is preserved in the image. For example, all features of the same snoring event sample are arranged in a row with acoustic features first and physiological features last, or similar features are arranged in a column according to feature type. Then, based on the normalized feature matrix, each element in the matrix (i.e., the normalized feature value) is directly used as the gray value of the image pixel and sequentially filled into the pixel position of the image according to the preset arrangement rule. During the filling process, it is ensured that each feature value corresponds one-to-one with the pixel position, without misalignment or omission. Finally, an initial grayscale feature map is formed that intuitively reflects the distribution of fused features. The image can initially demonstrate the difference between normal snoring and pathological snoring in terms of fused features.

[0042] Step 4.3: Based on the initial grayscale feature map, calculate the gradient matrix in the horizontal direction and the gradient matrix in the vertical direction. Specifically, for the obtained initial grayscale feature map, perform gradient calculation pixel by pixel. In the horizontal gradient calculation, take each pixel as the center, obtain the grayscale value of the pixel and its horizontally adjacent pixels, and calculate the difference between the grayscale values ​​of adjacent pixels to obtain the change amplitude of the pixel in the horizontal direction. Arrange all the horizontal change amplitudes of all pixels in the original image pixel arrangement order to form the horizontal gradient matrix. In the vertical gradient calculation, similarly take each pixel as the center, obtain the grayscale value of the pixel and its vertically adjacent pixels, such as the grayscale value of the pixel below and the pixel above, calculate the difference between the grayscale values ​​of adjacent pixels to obtain the change amplitude of the pixel in the vertical direction, and then arrange all the vertical change amplitudes in the original image pixel arrangement order to form the vertical gradient matrix.

[0043] Step 4.4 involves mapping the initial grayscale feature map, horizontal gradient matrix, and vertical gradient matrix to pixel values ​​for the red, green, and blue channels, respectively, to obtain three single-channel images. Specifically, this includes: first, processing the initial grayscale feature map by directly using the grayscale value of each pixel as the pixel value for the red channel image, generating a single-channel image containing only red channel information, thus preserving the basic distribution information of the fused features; next, processing the horizontal gradient matrix by statistically analyzing the maximum and minimum values ​​of all gradient values ​​in the matrix, linearly mapping the gradient values ​​to an integer range consistent with the standard range of pixel values, and using these mapped values ​​as pixel values ​​for the green channel image, generating a single-channel image containing only green channel information, reflecting the horizontal variation of the fused features; and finally, processing the vertical gradient matrix using the same mapping method as the horizontal gradient matrix, mapping the vertical gradient values ​​to the same integer range, using these as pixel values ​​for the blue channel image, generating a single-channel image containing only blue channel information, reflecting the vertical variation of the fused features.

[0044] Step 4.5 involves merging the three single-channel images to generate the final three-channel feature texture image. This includes: preparing an image channel merging tool; importing the obtained red, green, and blue single-channel images into the tool; combining pixel values ​​at the same coordinate positions in the three single-channel images according to the principle of one-to-one pixel correspondence; ensuring that each pixel contains red, green, and blue color components; ensuring complete alignment of pixel positions in the three channels during the merging process; and generating a complete three-channel image, which is the final three-channel feature texture image. The image simultaneously contains the basic distribution information of the fused features and the variation information in the horizontal and vertical directions. Normal snoring and pathological snoring show differences in color distribution and texture complexity in the image.

[0045] In this embodiment of the invention, all feature values ​​in the final feature subset are normalized to the integer range to obtain a normalized feature matrix. Each feature value in the normalized feature matrix is ​​arranged as a pixel gray value according to a preset rule to obtain an initial gray-scale feature map. Based on the initial gray-scale feature map, the horizontal and vertical gradient matrices are calculated respectively. Then, the initial gray-scale feature map, the horizontal gradient matrix, and the vertical gradient matrix are mapped to red, green, and blue channel pixel values ​​respectively to obtain three single-channel images. Finally, the three single-channel images are merged to generate a three-channel feature texture image. Therefore, this method overcomes the limitations of existing image feature methods that rely solely on acoustic signal imagery without incorporating physiological signal imagery. Furthermore, the single gray-scale image features lack sufficient distinguishability, making it difficult to fully reflect the feature differences between normal and pathological snoring. This method transforms acoustic and physiological fusion features into a three-channel image rich in texture differences, enhancing the feature recognition of different types of snoring through gradient information.

[0046] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1 involves inputting the generated three-channel feature texture image into the input layer of a pre-trained convolutional neural network model. Specifically, this includes: first, confirming the input requirements of the pre-trained convolutional neural network model, including parameters such as image size and pixel value range; adjusting the generated three-channel feature texture image to meet the specifications required by the model's input layer, ensuring that the image's width, height, and the pixel value distribution of the three color channels are consistent with the input standards during model training; the convolutional neural network model is trained using a large number of labeled feature texture images, and the model parameters are iteratively optimized during training to enable the model to adapt to sleep snoring-related feature texture data, avoiding the problem of poor adaptability of traditional convolutional neural networks to this type of specific data; then, inputting the adjusted three-channel feature texture image into the input layer of the convolutional neural network model to complete the image input preparation before classification.

[0047] Step 5.2 involves performing convolution operations on the image through sequentially connected convolutional layers in the convolutional neural network model to extract multi-layer abstract features. Specifically, the convolutional layers of the convolutional neural network model consist of multiple sequentially connected convolutional units. Each convolutional unit contains a convolutional kernel of a preset size and an activation function. When the three-channel feature texture image enters the first convolutional layer, the convolutional kernel slides across the image at a set stride, performing convolution calculations on the image region covered during the sliding process. That is, by operating the convolutional kernel on the pixel values ​​of the image region, local features of that region are extracted, such as texture edges and color distribution differences. These features are then processed by the activation function to enhance the non-linear expressive power of the features. After the processing of the first convolutional layer is completed, the output feature map is passed to the next convolutional layer. The convolutional layer further extracts more abstract global features based on the local features extracted by the previous layer, such as the smooth texture combination unique to normal snoring and the irregular texture pattern unique to pathological snoring. Through the progressive multi-layer convolutional operations, multi-layer feature information from low-level local features to high-level abstract features is mined from the three-channel feature texture image.

[0048] Step 5.3 involves inputting the feature map output from the convolutional layer into the pooling layer of the convolutional neural network model for downsampling to obtain high-level features with translation invariance. Specifically, after each convolutional layer is processed, the output feature map is directly input into the corresponding pooling layer. The pooling layer uses a preset pooling window and slides on the feature map with a set stride to aggregate and calculate the feature values ​​covered by each pooling window. For example, the maximum value within the window is taken, i.e., max pooling. The most representative feature value within the window is retained, while some redundant details are discarded, thus achieving downsampling of the feature map. Downsampling not only reduces the dimensionality and computation of the feature data but also enables the extracted features to have translation invariance. That is, no matter where the key feature is in the feature map, it can still be effectively identified after pooling, avoiding the problem of model misjudgment due to slight changes in feature position. After processing by all pooling layers, the high-level features reflecting the core distinguishing information of the three-channel feature texture image are finally obtained.

[0049] Step 5.4 flattens the high-level features into a one-dimensional feature vector and inputs it into the fully connected layer of the convolutional neural network model for nonlinear transformation to obtain the feature vector output by the fully connected layer. Specifically, this involves: flattening the obtained high-level features, usually multi-channel two-dimensional feature maps, according to the pixel arrangement order, that is, arranging the feature values ​​of each channel and each pixel of the two-dimensional feature map into a one-dimensional vector in sequence, ensuring that all information in the high-level features is completely preserved without misalignment; then inputting the flattened one-dimensional feature vector into the fully connected layer of the convolutional neural network model, where each neuron of the fully connected layer establishes a connection with all feature values ​​in the one-dimensional feature vector, linearly combining the feature vector through preset weight parameters, and then performing a nonlinear transformation through an activation function, such as the ReLU function; the nonlinear transformation can capture the complex correlations in the high-level features, such as the synergistic change patterns of acoustic feature texture and physiological feature texture, integrating the scattered high-level features into more discriminative feature information, and finally outputting a feature vector with lower dimension and stronger expressive power. Step 5.5: Using the feature vector output from the fully connected layer, the Softmax classifier calculates the final classification result of whether the snoring event belongs to normal snoring or pathological snoring. Specifically, this includes: inputting the feature vector output from the fully connected layer into the Softmax classifier; the classifier calculates the probability values ​​of the feature vector corresponding to the two categories of normal snoring and pathological snoring; during the calculation, the Softmax classifier converts the value of the feature vector into two probability values ​​between 0 and 1, and the sum of the two probability values ​​is 1. One probability value represents the probability that the current snoring event belongs to normal snoring, and the other represents the probability that it belongs to pathological snoring; then, the two probability values ​​are compared, and the category with the larger probability value is taken as the final classification result of the snoring event. If the probability value corresponding to normal snoring is larger, the snoring event is determined to be normal snoring; if the probability value corresponding to pathological snoring is larger, it is determined to be pathological snoring. The core function of the Softmax classifier is to convert the raw values ​​output from the last fully connected layer of the convolutional neural network into probability values ​​of the corresponding categories, and then determine the category to which the sample belongs based on the probability magnitude.

[0050] In this embodiment of the invention, the generated three-channel feature texture image is input into the input layer of a pre-trained convolutional neural network model. The image is convolved through the convolutional layers connected in sequence in the model to extract multi-layer abstract features. The feature map output by the convolutional layer is input into the pooling layer for downsampling to obtain high-level features with translation invariance. The high-level features are flattened into a one-dimensional feature vector and then input into a fully connected layer for nonlinear transformation. The feature vector output by the fully connected layer is then used by a Softmax classifier to calculate the classification result. Therefore, this method overcomes the technical problems in the prior art, such as the lack of complex feature mapping and nonlinear fitting capabilities of traditional classifiers, the poor adaptability of traditional convolutional neural networks to specific data of sleep snoring, and the difficulty in deeply mining the key distinguishing information between normal and pathological snoring from fused features. This method achieves the technical effect of extracting the acoustic and physiological fusion feature differences contained in the three-channel feature texture image, efficiently and accurately determining whether a snoring event belongs to normal or pathological snoring, and providing a reliable classification basis for the auxiliary diagnosis of sleep-disordered breathing.

[0051] In a preferred embodiment of the present invention, step 5.1 above may include: Step 5.11: Construct the basic network architecture, establishing a network topology consisting of an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer. Specifically, this includes: designing the hierarchical composition and connections of the basic network to distinguish between normal and pathological snoring based on the dimensions of the three-channel feature texture image (e.g., width, height, number of channels) and classification requirements; setting the input layer to a structure matching the size of the feature texture image to receive image data; sequentially connecting multiple convolutional layers after the input layer, each layer configured with a different number of convolutional kernels to extract features at different levels; connecting pooling layers after each convolutional layer to reduce feature dimensionality; connecting fully connected layers after the pooling layers to integrate feature information; and finally, ending with an output layer to output two classification results, clarifying the order of layers and data flow, forming a complete network topology consisting of an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer.

[0052] Step 5.12: Based on the constructed basic network architecture, configure the initial convolutional layer and connect it sequentially to the normalization processing layer and the non-linear activation layer to obtain the feature extraction backbone network. Specifically, this includes: Based on the established basic network architecture, first configure the initial convolutional layer, selecting a 3x3 convolutional kernel, suitable for capturing local texture details of the image, such as the edges and color changes of the snoring feature texture. Set an appropriate number of convolutional kernels, such as 64, and a stride, such as stride 1, to ensure that the basic local features of the feature texture image can be initially extracted. After the initial convolutional layer, connect the normalization processing layer to normalize the feature data output by the convolution, so that the mean of each feature channel is equal to the mean of the convolutional layer. To maintain stable variance and avoid the impact of data distribution fluctuations on feature extraction, thus enhancing the stability of network training, a non-linear activation layer, such as the ReLU activation function, is connected after the standardization layer. This non-linear transformation enhances the network's ability to express complex features, such as the textural differences between normal and pathological snoring, solving the problem that traditional linear models cannot capture non-linear relationships between features. Ultimately, a stable backbone network for extracting basic features is formed. The ReLU activation function is calculated as follows: for an input value greater than 0, the output is equal to the input value itself; when the input value is less than or equal to 0, the output is 0.

[0053] Step 5.13: Based on the constructed feature extraction backbone network, a parallel convolutional branch is added, and the branch output is fused with the backbone network output to construct a residual connection device. Specifically, this includes: adding a parallel convolutional branch on the basis of the constructed feature extraction backbone network. The branch adopts a structure of 1x1 convolutional kernel plus 3x3 convolutional kernel. The 1x1 convolutional kernel reduces the number of feature channels to reduce the amount of computation, and the 3x3 convolutional kernel further extracts feature details that are complementary to the backbone network. The output of the parallel convolutional branch is fused with the output of the backbone network: if the number of channels is different, the number of channels of the backbone network output is adjusted by the 1x1 convolutional kernel to make the dimensions of the branch and the backbone output consistent. Then, the features of the two are fused by element-wise addition to form a residual connection device.

[0054] Step 5.14: Based on the output of the constructed residual connection device, configure a feature downsampling layer to reduce the dimensionality of the feature map and obtain a processed feature map. Specifically, this includes configuring a feature downsampling layer for the feature map output by the residual connection device, using a 2x2 max pooling window, and setting a sliding step size of 2 to reduce the width and height of the feature map to half of their original size, such as reducing a 64x64 feature map to 32x32. Through downsampling, the most representative feature values ​​in the feature map are retained, while reducing the dimensionality and redundant information of the feature data, reducing the computational complexity of the network layers. During the downsampling process, key discriminative features are ensured not to be lost, such as the texture changes corresponding to blood oxygen and pulse rate fluctuations unique to pathological snoring. Finally, a processed feature map with a simplified size and retained core information is obtained.

[0055] Step 5.15: Based on the processed feature map, a multi-layer fully connected network is constructed and a classifier is connected at the end to obtain the complete network structure. Specifically, this includes: based on the processed feature map, it is first flattened into a one-dimensional feature vector, and then a multi-layer fully connected network is constructed. The first fully connected layer has a large number of neurons to initially integrate high-level features; the second fully connected layer reduces the number of neurons to further compress features and strengthen key information; the last fully connected layer has two neurons, corresponding to the two categories of normal snoring and pathological snoring. A classifier is connected at the end of the fully connected network. A Softmax classifier is selected, and classification is achieved by calculating the probability values ​​of the two categories, solving the problem that traditional classifiers cannot handle high-dimensional complex features. The combination of the multi-layer fully connected network and the classifier can deeply explore the correlation patterns between features, forming a complete network structure.

[0056] Step 5.16: Based on the complete network structure, using an annotated feature texture image training set, the network parameters are iteratively optimized through the error backpropagation algorithm to obtain the final usable convolutional neural network model. Specifically, this includes: collecting a large number of annotated feature texture images as the training set, where each image is clearly labeled as either normal or pathological snoring, and includes diverse samples from OSAHS patients and healthy individuals to ensure the training data covers the feature differences between the two types of snoring; inputting the training set into the constructed complete network structure, and training it using the error backpropagation algorithm. First, the predicted classification result is calculated through forward propagation and compared with the actual annotations, using the cross-entropy loss function to calculate the prediction error; then, the error is backpropagated from the output layer to the input layer, and the weight parameters of each layer are adjusted according to the error magnitude; repeating the forward and backward propagation process, iteratively optimizing the parameters, such as iterating 100 times, until the loss function value tends to stabilize and the classification accuracy of the validation set reaches the preset standard; through targeted training, the network parameters are adapted to the sleep snoring feature texture data, ultimately obtaining a convolutional neural network model that accurately distinguishes between normal and pathological snoring, solving the problem of poor adaptability of traditional models. The cross-entropy loss function is... , This represents the average cross-entropy loss value of the training set, where N represents the total number of feature texture image samples in the training set, and i is the sample index. This represents the true label of the i-th sample in the training set. This represents the probability that the convolutional neural network model predicts for the i-th sample in the training set.

[0057] In this embodiment of the invention, a basic network topology is constructed, comprising an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer. Based on this architecture, an initial convolutional layer is configured and sequentially connected to a normalization processing layer and a nonlinear activation layer to form a feature extraction backbone network. Parallel convolutional branches are added to the backbone network, and the branch outputs are fused with the backbone output to construct a residual connection device. Based on the output of the residual connection device, a feature downsampling layer is configured to achieve feature map dimensionality reduction. The processed feature map is combined to construct a multi-layer fully connected network, with a classifier connected at the end to form a complete network structure. Then, a labeled feature texture image is used as the training set to pass through the error correction process. The backpropagation algorithm iteratively optimizes network parameters, thus overcoming the technical problems of existing traditional convolutional neural networks, such as poor adaptability to specific sleep snoring data, lack of residual connections leading to gradient vanishing resulting in insufficient extraction of deep features, failure to enhance feature expression through the synergy of standardization and activation layers, unreasonable configuration of downsampling layers leading to feature redundancy or dimensionality out-of-control, and insufficient classification accuracy due to the lack of targeted iterative optimization of network parameters. In this way, it achieves the construction of a convolutional neural network model that adapts to sleep snoring feature texture data, can efficiently extract deep abstract features, avoids gradient vanishing, reduces feature redundancy, and optimizes parameters.

[0058] like Figure 2 As shown, embodiments of the present invention also provide a snoring classification system based on multidimensional sleep symptom signal feature maps, including: The acquisition module is used to preprocess the synchronously acquired acoustic signals of the entire night's sleep, blood oxygen saturation signals, and pulse rate signals; based on the preprocessed acoustic signals, multiple independent snoring event segments are located and segmented by detecting the peak value of the acoustic signal envelope and combining it with the dual-threshold endpoint detection method. The extraction module is used to extract Mel frequency cepstral coefficient features, time domain features and frequency domain features from the corresponding acoustic signals based on each snoring event segment to form an acoustic feature set, and simultaneously extract fluctuation features and extreme value features from the corresponding blood oxygen saturation signal and pulse rate signal to form a physiological feature set. The fusion module is used to fuse the obtained acoustic feature set and physiological feature set to obtain an initial fused feature set; the Relief-F algorithm is used to calculate and filter the feature weights of the initial fused feature set to obtain the final feature subset. The calculation module is used to normalize the values ​​in the final feature subset and map the normalized values ​​to grayscale image pixels to obtain the initial grayscale feature map; calculate the horizontal and vertical gradients of the initial grayscale feature map and assign the initial grayscale feature map and the gradient maps in the two directions to the red, green and blue channels respectively to obtain a three-channel feature texture image. The determination module is used to input the feature texture image into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model ultimately determines whether the snoring event is normal snoring or pathological snoring.

[0059] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A snoring classification method based on multidimensional sleep sign signal feature maps, characterized in that, The method includes: The acoustic signals, blood oxygen saturation signals, and pulse rate signals collected throughout the night were preprocessed. Based on the preprocessed acoustic signals, multiple independent snoring event segments were located and segmented by detecting the peak value of the acoustic signal envelope and combining it with the dual-threshold endpoint detection method. Based on each snoring event segment, Mel frequency cepstral coefficient features, time domain features, and frequency domain features are extracted from the corresponding acoustic signals to form an acoustic feature set. Simultaneously, fluctuation features and extreme value features are extracted from the corresponding blood oxygen saturation signals and pulse rate signals to form a physiological feature set. The obtained acoustic feature set and physiological feature set are fused to obtain an initial fused feature set; the Relief-F algorithm is used to calculate and filter the feature weights of the initial fused feature set to obtain the final feature subset. The values ​​in the final feature subset are normalized and mapped to grayscale image pixels to obtain the initial grayscale feature map. The horizontal and vertical gradients of the initial grayscale feature map are calculated and the initial grayscale feature map and the gradient maps in the two directions are assigned to the red, green and blue channels respectively to obtain a three-channel feature texture image. The feature texture image is input into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model ultimately determines whether the snoring event is normal or pathological snoring.

2. The snoring classification method based on multidimensional sleep sign signal feature maps according to claim 1, characterized in that, Preprocessing is performed on the synchronously acquired acoustic signals, blood oxygen saturation signals, and pulse rate signals from the entire night's sleep. Based on the preprocessed acoustic signal, by detecting the peak value of the acoustic signal envelope and combining it with a dual-threshold endpoint detection method, multiple independent snoring event segments are located and segmented, including: Acoustic signals, blood oxygen saturation signals, and pulse rate signals from the entire night's sleep were collected and uploaded, and the signals were parsed into actionable time-series data. The acoustic signal in the time series data is downsampled to obtain a downsampled acoustic signal with reduced computational complexity; Based on a preset physiological range threshold, outlier filtering is performed on the blood oxygen saturation signal and pulse rate signal in the time series data to obtain clean physiological signal data. Based on downsampled acoustic signals and clean physiological signal data, according to their respective start and end timestamps, the same valid data segment where all signals completely overlap in time is extracted; Based on the time-aligned downsampled acoustic signal, the envelope of the downsampled acoustic signal is calculated and the peak point is detected. Combined with the dual-threshold endpoint detection method, the start and end times of each snoring event are determined, and multiple independent snoring event segments are synchronously cut out from the same valid data segment accordingly.

3. The snoring classification method based on multidimensional sleep sign signal feature maps according to claim 2, characterized in that, Based on each snoring event segment, Mel-frequency cepstral coefficient features, time-domain features, and frequency-domain features are extracted from the corresponding acoustic signal to form an acoustic feature set. Simultaneously, fluctuation features and extreme value features are extracted from the corresponding blood oxygen saturation and pulse rate signals to form a physiological feature set, including: Based on the acoustic signals in each independent snoring event segment obtained by segmentation, pre-emphasis, framing and windowing preprocessing operations are performed to obtain standardized acoustic frame signals; The standardized acoustic frame signal is subjected to a fast Fourier transform and converted to the frequency domain. The energy of the filter bank is extracted through the Mel filter bank, and then the Mel frequency cepstral coefficient characteristics are calculated through logarithmic operation and discrete cosine transform. Based on the standardized acoustic frame signal, the statistical characteristics of short-time energy and short-time zero-crossing rate are calculated to form a time-domain feature subset, and the spectral centroid, spectral entropy and spectral slope are calculated to form a frequency-domain feature subset. The obtained Mel frequency cepstral coefficient features are combined with the obtained time-domain feature subset and frequency-domain feature subset to form the acoustic feature set of snoring event segments. Based on the blood oxygen saturation signal and pulse rate signal corresponding to the same snoring event segment obtained by segmentation, the numerical fluctuation in the three time intervals of the first, middle and last time intervals is calculated as the fluctuation feature, and the minimum and maximum values ​​in the signal interval are extracted as the extreme value features to form the physiological feature set of the snoring event segment.

4. The snoring classification method based on multidimensional sleep sign signal feature maps according to claim 3, characterized in that, The obtained acoustic feature set is fused with the physiological feature set to obtain an initial fused feature set; The Relief-F algorithm is used to calculate and filter feature weights in the initial fused feature set to obtain the final feature subset, which includes: The obtained acoustic feature set and the obtained physiological feature set are concatenated and combined to obtain the initial fusion feature set; Based on the initial fused feature set, initialize the feature weight array of the Relief-F algorithm, and set the parameters for the number of samplings and the number of nearest neighbors; Under the control of the set parameters, multiple rounds of random sampling are performed on the initial fused feature set. In each round of sampling, the feature difference between the nearest neighbor sample within each category and the nearest neighbor sample between different categories is calculated to obtain the feature difference result. Based on the obtained feature difference results, the weight of each feature in the feature weight array is dynamically updated, and features with weights higher than the preset threshold are selected to form the final feature subset.

5. The snoring classification method based on multidimensional sleep sign signal feature maps according to claim 4, characterized in that, The values ​​in the final feature subset are normalized, and the normalized values ​​are mapped to grayscale image pixels to obtain the initial grayscale feature map. Calculate the horizontal and vertical gradients of the initial grayscale feature map, and assign the initial grayscale feature map and the gradient maps in the two directions to the red, green, and blue channels respectively, to obtain a three-channel feature texture image, including: Normalize all eigenvalues ​​in the final feature subset to map the linear values ​​to the integer range, and obtain the normalized feature matrix. Each feature value in the obtained normalized feature matrix is ​​used as the gray value of the pixel and arranged according to a preset arrangement rule to obtain the initial gray-scale feature map. Based on the initial grayscale feature map, the gradient matrix in the horizontal direction and the gradient matrix in the vertical direction are calculated respectively. The initial grayscale feature map, the horizontal gradient matrix, and the vertical gradient matrix are mapped to pixel values ​​in the red, green, and blue channels, respectively, to obtain three single-channel images. The three single-channel images are merged to generate the final three-channel feature texture image.

6. The snoring classification method based on multidimensional sleep sign signal feature maps according to claim 5, characterized in that, The feature texture image is input into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model ultimately determines whether the snoring event is normal or pathological, including: The generated three-channel feature texture image is input into the input layer of a pre-trained convolutional neural network model; By using the convolutional layers connected sequentially in the convolutional neural network model, convolution operations are performed on the image to extract multi-layered abstract features; The feature map output from the convolutional layer is input into the pooling layer in the convolutional neural network model for downsampling to obtain high-level features with translation invariance. The high-level features are flattened into a one-dimensional feature vector, which is then input into the fully connected layer of the convolutional neural network model for nonlinear transformation to obtain the feature vector output by the fully connected layer. The feature vectors output by the fully connected layer are used by the Softmax classifier to calculate the final classification result of whether the snoring event belongs to normal snoring or pathological snoring.

7. The snoring classification method based on multidimensional sleep sign signal feature maps according to claim 6, characterized in that, The generated three-channel feature texture image is input into the input layer of a pre-trained convolutional neural network model, including: Construct the basic network architecture and establish a network topology that includes an input layer, multiple convolutional layers, pooling layers, fully connected layers, and an output layer; Based on the constructed basic network architecture, the initial convolutional layer is configured and then connected sequentially to the normalization processing layer and the non-linear activation layer to obtain the feature extraction backbone network. Based on the constructed feature extraction backbone network, a parallel convolution branch is added, and the branch output is fused with the backbone network output to construct a residual connection device. Based on the output of the constructed residual connection device, a feature downsampling layer is configured to reduce the size of the feature map and obtain the processed feature map. Based on the processed feature maps, a multi-layer fully connected network is constructed and a classifier is connected at the end to obtain the complete network structure. Based on the complete network structure, a labeled feature texture image training set is used, and the network parameters are iteratively optimized through the backpropagation algorithm to obtain the final usable convolutional neural network model.

8. A snoring classification system based on multidimensional sleep symptom signal feature maps, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to preprocess the synchronously acquired acoustic signals, blood oxygen saturation signals, and pulse rate signals from the entire night's sleep. Based on the preprocessed acoustic signal, multiple independent snoring event segments are located and segmented by detecting the peak value of the acoustic signal envelope and combining it with the dual-threshold endpoint detection method. The extraction module is used to extract Mel frequency cepstral coefficient features, time domain features and frequency domain features from the corresponding acoustic signals based on each snoring event segment to form an acoustic feature set, and simultaneously extract fluctuation features and extreme value features from the corresponding blood oxygen saturation signal and pulse rate signal to form a physiological feature set. The fusion module is used to fuse the obtained acoustic feature set and physiological feature set to obtain an initial fused feature set; the Relief-F algorithm is used to calculate and filter the feature weights of the initial fused feature set to obtain the final feature subset. The calculation module is used to normalize the values ​​in the final feature subset and map the normalized values ​​to grayscale image pixels, thereby obtaining the initial grayscale feature map. Calculate the horizontal and vertical gradients of the initial grayscale feature map, and assign the initial grayscale feature map and the gradient maps in the two directions to the red, green and blue channels respectively, to obtain a three-channel feature texture image; The determination module is used to input the feature texture image into a pre-trained convolutional neural network model for forward propagation calculation. The output layer of the convolutional neural network model ultimately determines whether the snoring event is normal snoring or pathological snoring.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images