A sleep apnea detection method based on multi-channel respiratory signals

Abstract

The application discloses a sleep apnea detection method based on multi-channel respiratory signals and belongs to the technical field of biomedical signal processing. The method first collects multi-channel respiratory signals such as abdominal signals, chest signals and nasal airflow signals of a human body, and after filtering, normalization, segmentation and other preprocessing steps, the processed signal segments are input into a deep learning encoder to learn the human respiratory pattern and obtain corresponding respiratory features. Based on the respiratory features, a classification model based on a fully connected feedback neural network is further established to realize the classification of three states of normal breathing, insufficient breathing and apnea. The application provides a simple, efficient and low-cost sleep apnea detection technology, which utilizes multi-scale time series analysis and supervised contrast learning methods to more effectively process complex physiological signals, and provides a more efficient and accurate technical means for the screening and diagnosis of sleep apnea.

Description

Technical Field

[0001] This invention relates to the field of biomedical signal processing technology, and in particular to a method for detecting sleep apnea based on multi-channel respiratory signals. Background Technology

[0002] Sleep apnea syndrome is a common sleep disorder in which patients experience pauses or significantly reduced breathing during sleep, leading to symptoms such as poor sleep quality at night and excessive daytime sleepiness. Long-term untreated sleep apnea can cause a variety of health problems, including cardiovascular disease, metabolic syndrome, and neuropsychiatric disorders.

[0003] Traditional sleep apnea diagnosis relies on polysomnography (PSG) monitoring, which requires specialized sleep centers, resulting in high costs and significant inconvenience for patients. Existing detection methods (Zhu Weiwei, Zhang Han, Yu Baoxian, et al. A method, device, equipment, and storage medium for sleep apnea detection [P]. Guangdong Province: CN117017233B, 2024-01-09.) analyze changes in small-scale signals relative to large-scale signals, perform statistical comparisons to obtain relative features characterizing respiratory signal changes, and use feature selection to extract target relative features to construct a training feature set for model training. The relative features of the user's respiratory signal are then input into the target sleep detection model for detection, obtaining the user's sleep apnea detection result. The steps include: inputting the relative features of the user's respiratory signal into the target sleep detection model to obtain the user's sleep breathing probability vector; and obtaining the user's sleep apnea detection result based on the sleep breathing probability vector and a preset probability vector threshold. However, existing detection methods generally only use single-channel respiratory signals, resulting in poor detection performance for hypopnea events. Therefore, developing an efficient, low-cost, and stable automatic detection method for sleep apnea has significant clinical application value. Summary of the Invention

[0004] The purpose of this invention is to provide an automatic sleep apnea detection method based on multi-channel respiratory signals. By combining advanced signal processing technology and deep learning algorithms, it achieves efficient and accurate diagnosis of sleep apnea, overcoming the shortcomings of existing technologies.

[0005] The present invention is achieved by at least one of the following technical solutions.

[0006] A method for detecting sleep apnea based on multi-channel respiratory signals includes the following steps:

[0007] S1. Collect multi-channel respiratory signals from the human body;

[0008] S2. Preprocess the collected multi-channel respiratory signals to obtain several multi-channel respiratory signal segments;

[0009] S3. Input the preprocessed multi-channel respiratory signal segments into a deep learning encoder to learn human breathing patterns and obtain the corresponding respiratory features.

[0010] S4. Based on the breathing features, establish and train a classification model based on a fully connected feedback neural network, and use the trained model to classify the breathing features.

[0011] Furthermore, the preprocessing includes filtering, normalization, and segmentation, comprising the following steps:

[0012] d. Process the multi-channel breathing signal using a low-pass filter to remove high-frequency noise from the signal.

[0013] e. Normalize the filtered respiratory signal to scale the signal amplitude to a uniform range. The normalization formula used is:

[0014]

[0015] Where, x norm σ is the normalized signal, x is the filtered signal, μ is the mean of the filtered signal, and σ is the standard deviation of the filtered signal.

[0016] f. Divide the normalized respiratory signal into segments, with each segment representing a respiratory pattern of a fixed duration, to facilitate subsequent feature extraction and classification.

[0017] Furthermore, the deep learning encoder includes a multi-scale temporal feature extraction module and a residual module.

[0018] Furthermore, the multi-scale temporal feature extraction module includes:

[0019] Frequency analysis of the respiratory signal segment in each channel was performed using Fast Fourier Transform (FFT) to identify and select k key frequency components and their corresponding cycles.

[0020] The original respiratory signal is reconstructed into k two-dimensional tensors in the time domain according to the identified period;

[0021] A set of parameter-shared two-dimensional convolution kernels are applied to the two-dimensional tensor at each time scale to extract local spatiotemporal features at each time scale in parallel, and the extracted two-dimensional features are reconstructed into k one-dimensional sequences.

[0022] Adaptively aggregate k one-dimensional sequences to form a comprehensive multi-scale temporal feature representation;

[0023] The multi-scale temporal feature representations extracted from the respiratory signals of the three channels are stacked to obtain a new two-dimensional tensor output.

[0024] Furthermore, the residual module is used to integrate the multi-scale temporal features extracted by the multi-scale temporal feature extraction module structure, including the following steps:

[0025] The two-dimensional tensor output by the multi-scale temporal feature extraction module is used as input, and the features are further processed through a series of residual learning units. Each residual learning unit contains two main parts: a set of convolutional layers and skip connections. In each residual learning unit, the convolutional layers are responsible for deep feature extraction of the input features, and the non-linear processing capability is increased through activation functions. The skip connections directly add the unit input to the output of the convolutional layers to achieve rapid feature transfer. The formula for the residual module is as follows:

[0026] x out =ReLU(BN(W2(ReLU(BN(W1x)) in )))))+W3x in

[0027] Where, x out For the output characteristics of the residual module, x in W1 represents the input features of the residual module, W2 represents the parameters of the first convolutional kernel in the module, W3 represents the parameters of the second convolutional kernel, ReLU represents the activation function, and BN represents batch normalization.

[0028] After processing by three residual modules, the multi-scale temporal feature representation with enhanced depth and complexity is integrated to obtain robust breathing pattern features.

[0029] Furthermore, the classification model based on a fully connected feedback neural network includes:

[0030] The input layer is used to receive breathing features from the deep learning encoder;

[0031] At least one hidden layer is used to perform nonlinear transformation and feature combination on the input breathing features to capture complex relationships between breathing patterns;

[0032] The output layer uses the softmax activation function to map the output of the hidden layer to the probability distribution of three categories: normal, insufficient breathing, and apnea, in order to classify respiratory events.

[0033] Furthermore, the classification model of the fully connected feedback neural network is trained using a supervised learning method, with the loss function as follows:

[0034]

[0035] Where n is the total number of samples, m is a single sample, and y h It is a Boolean indicator (0 or 1) indicating whether the sample belongs to the h-th category, a h It is the probability that the model predicts a sample belongs to the h-th class, where h∈(1, 2, 3).

[0036] Furthermore, the training of the deep learning encoder employs a supervised contrastive learning strategy, including the following steps:

[0037] First, an anchor sample is randomly selected from the training dataset, and a positive sample is selected from the same class and multiple negative samples are selected from different classes;

[0038] Then, the anchor samples, positive samples, and negative samples are processed by a deep learning encoder to extract their respective feature representations.

[0039] Next, the similarity between the anchor feature representation and the positive sample feature representation, as well as the similarity between the anchor feature representation and each negative sample feature representation, is calculated using the cosine similarity formula:

[0040]

[0041] Where u and v represent two eigenvectors, (·) represents the dot product of vectors, and ||u|| and ||v|| represent the magnitudes of the two eigenvectors, respectively;

[0042] Based on the similarity calculated above, a contrastive loss function is used to optimize the model parameters. The contrastive loss function aims to minimize the distance between the anchor point and the positive sample, while maximizing the distance between the anchor point and the negative sample.

[0043] Furthermore, the comparison loss function is:

[0044]

[0045] Among them, v positive V represents the feature representation of a positive sample. i This represents the feature representation of all selected samples, including positive and negative samples. N is the number of negative samples, and τ is a temperature parameter used to adjust the scaling of similarity.

[0046] Furthermore, the multi-channel respiratory signal includes abdominal signal, chest signal, and nasal airflow signal.

[0047] Compared with existing technologies, the beneficial effects of the present invention are as follows:

[0048] This invention significantly improves the accuracy and efficiency of sleep apnea detection by comprehensively applying multi-scale temporal feature extraction, residual learning mechanisms, and supervised contrastive learning strategies. Compared to traditional methods, this invention can more effectively process complex and nonlinear biomedical signals, providing a more efficient and accurate technical means for the screening and diagnosis of sleep apnea.

[0049] Furthermore, this invention employs a supervised contrastive learning strategy to train the deep learning encoder. By minimizing the distance between similar samples and maximizing the distance between samples of different categories, this invention significantly enhances the model's ability to distinguish breathing features.

[0050] This invention significantly improves the accuracy and efficiency of sleep apnea detection by comprehensively applying multi-scale temporal feature extraction, residual learning mechanisms, and supervised contrastive learning strategies. Compared to traditional methods, this invention can more effectively process complex and nonlinear biomedical signals, providing a more efficient and accurate technical means for the screening and diagnosis of sleep apnea. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating a sleep apnea detection method based on multi-channel respiratory signals, as an example.

[0052] Figure 2 This is a diagram of a deep learning encoder model for a sleep apnea detection method based on multi-channel respiratory signals, as shown in the embodiment.

[0053] Figure 3 This is a flowchart illustrating the supervised contrastive learning strategy for a sleep apnea detection method based on multi-channel respiratory signals, as an example. Detailed Implementation

[0054] To facilitate a more accurate understanding of the technical solutions of this invention by those skilled in the art, the technical solutions proposed by this invention will be described in detail below with reference to specific embodiments and corresponding drawings. It should be noted that the specific embodiments described are only some implementations of this invention, not all of them. Based on these embodiments of this invention, those skilled in the art can derive many other modified implementations without departing from the inventive concept of this invention, and these modifications should also be considered within the scope of protection of this invention. Furthermore, to make the description of this invention more concise and clear, basic and conventional technologies already familiar to those skilled in the art will not be elaborated upon here.

[0055] like Figure 1As shown in the figure, this embodiment of a sleep apnea detection method based on multi-channel respiratory signals first collects multi-channel respiratory signals from the human body; then, the collected multi-channel respiratory signals undergo preprocessing steps such as filtering, normalization, and segmentation to obtain several multi-channel respiratory signal segments; next, the multi-channel respiratory signal segments are input into a deep learning encoder to obtain corresponding respiratory features; then, a classification model based on a fully connected feedback neural network is established for classification, specifically including the following steps:

[0056] S1. Collect the user's multi-channel respiratory signals, including abdominal signals, chest signals, and nasal airflow signals.

[0057] S2. Preprocess the collected multi-channel respiratory signals.

[0058] Specifically, the multi-channel breathing signal is first processed using a low-pass filter to remove high-frequency noise from the signal.

[0059] It should be noted that the low-pass filter used is a fourth-order Butterworth low-pass filter with a cutoff frequency of 0.7Hz. The Butterworth low-pass filter has a flat response characteristic, providing a smooth frequency response at the cutoff frequency and maximizing the consistency of signal amplitude within the passband.

[0060] The filtered respiratory signal is then normalized to scale its amplitude to a uniform range. The normalization formula used is as follows:

[0061]

[0062] Where, x norm Here, x is the normalized signal, and μ is the filtered signal.

[0063] σ is the standard deviation of the filtered signal;

[0064] Finally, the normalized respiratory signal is segmented, with each segment representing a respiratory pattern of a fixed duration.

[0065] S3. Input the multi-channel respiratory signal segments into a deep learning encoder to learn human breathing patterns and obtain corresponding respiratory features; such as... Figure 2 As shown, the deep learning encoder includes a multi-scale temporal feature extraction module (multi-scale temporal analysis module) and a residual module.

[0066] Specifically, the multi-scale temporal feature extraction module works as follows: First, frequency analysis is performed on the respiratory signal segment of each channel using Fast Fourier Transform (FFT) to identify and select k key frequency components and their corresponding periods. The original respiratory signal is then reconstructed into k two-dimensional tensors in the time domain according to the identified periods. A set of parameter-shared two-dimensional convolutional kernels are applied to each two-dimensional tensor at each time scale to extract local spatiotemporal features at each time scale in parallel. The extracted two-dimensional features are then reconstructed into k one-dimensional sequences. These k one-dimensional sequences are adaptively aggregated to form a comprehensive multi-scale temporal feature representation. The multi-scale temporal feature representations extracted from the respiratory signals of the three channels are stacked to obtain a new two-dimensional tensor output. The two-dimensional tensor output from the multi-scale temporal feature extraction module is used as input, and the features are further processed through a series of residual learning units.

[0067] The specific formula for the residual module is:

[0068] x out =ReLU(BN(W2(ReLU(BN(W1x)) in )))))+W3x in

[0069] Where, x out For the output characteristics of the module, x in W1 represents the input features of the module, W2 represents the parameters of the first convolutional kernel in the module, W3 represents the parameters of the second convolutional kernel, ReLU represents the activation function, and BN represents batch normalization.

[0070] As one example, the processing of three residual modules integrates a multi-scale temporal feature representation with enhanced depth and complexity, resulting in the model extracting respiratory pattern features.

[0071] like Figure 3 As shown, the deep learning encoder in this embodiment is trained using a supervised contrastive learning strategy.

[0072] It should be noted that supervised contrastive learning is a training method for deep neural networks. This method selects anchor points, positive samples, and negative samples from samples of multiple categories, optimizes model parameters using a contrastive loss function, reduces the similarity difference between anchor points and positive samples, and expands the difference with negative samples, thereby improving the model's ability to distinguish between samples of different categories. The specific steps are as follows:

[0073] In some optional embodiments, the respiratory signal dataset used for training includes, but is not limited to, publicly available datasets such as MESA and SHHS, as well as synchronously acquired human respiratory data.

[0074] An anchor sample is randomly selected from the training dataset, along with a positive sample from the same category and multiple negative samples from different categories. Features are extracted from the anchor sample, positive samples, and negative samples using a deep learning encoder to obtain their respective feature representations. The similarity between the anchor feature representation and the positive sample feature representation, and between the anchor feature representation and each negative sample feature representation, is calculated. The similarity is calculated using the cosine similarity formula:

[0075]

[0076] Where u and v represent two eigenvectors, (·) represents the dot product of vectors, and ||u|| and ||v|| represent the magnitudes of the vectors, respectively.

[0077] Based on the similarity calculated above, a contrastive loss function is used to optimize the model parameters. The contrastive loss function aims to minimize the distance between the anchor point and positive samples, while maximizing the distance between the anchor point and negative samples. The contrastive loss function used is:

[0078]

[0079] Among them, v positive V represents the feature representation of a positive sample. i This represents the feature representation of all selected samples, including positive and negative samples. N is the number of negative samples, and τ is a temperature parameter used to adjust the scaling of similarity.

[0080] S4. Based on the described breathing features, establish a classification model based on a fully connected feedback neural network, classifying the features into three categories: normal, insufficient breathing, and apnea. Specifically, the fully connected feedback neural network classification model includes: an input layer for receiving breathing features from a deep learning encoder; two hidden layers for performing nonlinear transformations and feature combinations on the input breathing features; and an output layer using a softmax activation function to map the output of the hidden layers onto the probability distributions of the three categories (normal, insufficient breathing, and apnea) to achieve classification of breathing events. The fully connected feedback neural network classification model is trained using a supervised learning method, and the specific loss function is as follows:

[0081]

[0082] Where n is the total number of samples, m is a single sample, and y h It is a Boolean indicator (0 or 1) indicating whether the sample belongs to the h-th category, a h It is the probability that the model predicts a sample belongs to the h-th class, where h∈(1, 2, 3).

[0083] The embodiments of the present invention have been described in detail through the accompanying drawings and specific examples. However, it should be understood that these embodiments represent only some applications of the present invention and do not limit the scope of implementation. Based on the knowledge possessed by those skilled in the art, various modifications or variations can be made to these embodiments according to specific needs without departing from the basic principles and spirit of the present invention. These corresponding changes and adjustments should also be considered to fall within the protection scope of the present invention.

Claims

1. A method for detecting sleep apnea based on multi-channel respiratory signals, characterized in that, Includes the following steps: S1. Collect multi-channel respiratory signals from the human body; S2. Preprocess the acquired multi-channel respiratory signals to obtain several multi-channel respiratory signal segments, including the following steps: a. The multi-channel breathing signal is processed using a low-pass filter to remove high-frequency noise. The low-pass filter used is a fourth-order Butterworth low-pass filter with a cutoff frequency of 0.7Hz. b. Normalize the filtered respiratory signal to scale the signal amplitude to a uniform range. The normalization formula used is: in, It is the normalized signal. It is the filtered signal. It is the mean of the filtered signal. It is the standard deviation of the filtered signal; c. Divide the normalized respiratory signal into segments, with each segment representing a respiratory pattern of a fixed duration, to facilitate subsequent feature extraction and classification. S3. Input the preprocessed multi-channel respiratory signal segments into a deep learning encoder to learn human breathing patterns and obtain the corresponding respiratory features. S4. Based on the breathing features, establish and train a classification model based on a fully connected feedback neural network, and use the trained model to classify the breathing features; The deep learning encoder includes a multi-scale temporal feature extraction module and a residual module; The multi-scale temporal feature extraction module includes: Frequency analysis of the respiratory signal segments in each channel was performed using Fast Fourier Transform (FFT) to identify and select... Key frequency components and their corresponding periods; The original respiratory signal was reconstructed in the time domain according to the identified period. A two-dimensional tensor; A set of parameter-shared two-dimensional convolutional kernels is applied to the two-dimensional tensor at each time scale to extract local spatiotemporal features at each time scale in parallel. The extracted two-dimensional features are then reconstructed into... A one-dimensional sequence; Will Adaptive aggregation of one-dimensional sequences is performed to form a comprehensive multi-scale temporal feature representation; The multi-scale temporal feature representations extracted from the respiratory signals of the three channels are stacked to obtain a new two-dimensional tensor output; The residual module is used to integrate the multi-scale temporal features extracted by the multi-scale temporal feature extraction module, including the following steps: The two-dimensional tensor output by the multi-scale temporal feature extraction module is used as input, and the features are further processed through a series of residual learning units. Each residual learning unit contains two main parts: a set of convolutional layers and skip connections. In each residual learning unit, the convolutional layers are responsible for deep feature extraction of the input features, and the non-linear processing capability is increased through activation functions. The skip connections directly add the unit input to the output of the convolutional layers to achieve rapid feature transfer. The formula for the residual module is as follows: )) + in, The output characteristics of the residual module, These are the input features for the residual module. The parameters of the first convolution kernel in the module. For the parameters of the second convolution kernel, For the third convolution kernel parameters, For activation function, For batch normalization; After processing by three residual modules, the multi-scale temporal feature representation with enhanced depth and complexity is integrated to obtain robust breathing pattern features. The classification model based on a fully connected feedback neural network includes: The input layer is used to receive breathing features from the deep learning encoder; At least one hidden layer is used to perform nonlinear transformation and feature combination on the input breathing features to capture complex relationships between breathing patterns; The output layer uses the softmax activation function to map the output of the hidden layer to the probability distribution of three categories: normal, insufficient breathing, and apnea, in order to classify respiratory events.

2. The sleep apnea detection method based on multi-channel respiratory signals according to claim 1, characterized in that, The classification model of the fully connected feedback neural network is trained using a supervised learning method, and the specific loss function is as follows: in, The total number of samples, For a single sample, It is a Boolean indicator (0 or 1) indicating whether the sample belongs to the first... Categories The model predicts that the sample belongs to the first... The probability of each category, ∈ (1, 2, 3).

3. The sleep apnea detection method based on multi-channel respiratory signals according to claim 1, characterized in that, The deep learning encoder is trained using a supervised contrastive learning strategy, including the following steps: First, an anchor sample is randomly selected from the training dataset, and a positive sample is selected from the same class and multiple negative samples are selected from different classes; Then, the anchor samples, positive samples, and negative samples are processed by a deep learning encoder to extract their respective feature representations. Next, the similarity between the anchor feature representation and the positive sample feature representation, as well as the similarity between the anchor feature representation and each negative sample feature representation, is calculated using the cosine similarity formula: in, and These represent two eigenvectors, This represents the dot product of vectors. and These represent the magnitudes of the two eigenvectors, respectively. Based on the similarity calculated above, a contrastive loss function is used to optimize the model parameters. The contrastive loss function aims to minimize the distance between the anchor point and the positive sample, while maximizing the distance between the anchor point and the negative sample.

4. The sleep apnea detection method based on multi-channel respiratory signals according to claim 3, characterized in that, The contrastive loss function is: in, The feature representation of positive samples, This represents the feature representation of all selected samples, including both positive and negative samples. It is the number of negative samples. It is a temperature parameter used to adjust the scaling of similarity.

5. A method for detecting sleep apnea based on multi-channel respiratory signals according to any one of claims 1 to 4, characterized in that, The multi-channel respiratory signal includes abdominal signal, chest signal, and nasal airflow signal.

Images