A sleep apnea detection method and device, an electronic device, and a storage medium
By collecting multi-source physiological signals in a waking state and using frequency and respiratory feature dimensionality reduction models for feature extraction and fusion, the problem of high cost and low efficiency in existing sleep apnea detection technologies has been solved, achieving rapid and accurate sleep apnea detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PEKING UNIVERSITY FIRST HOSPITAL (PEKING UNIVERSITY FIRST CLINICAL MEDICAL COLLEGE)
- Filing Date
- 2026-04-30
- Publication Date
- 2026-07-03
Smart Images
Figure CN122320481A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of medical testing technology, and more specifically, to a method, device, electronic device, and storage medium for detecting sleep apnea. Background Technology
[0002] Sleep apnea syndrome is a common sleep disorder characterized by repeated pauses in breathing and insufficient ventilation during sleep. It can cause decreased blood oxygen saturation at night, fragmented sleep, and symptoms such as daytime sleepiness and decreased attention. If left untreated for a long time, it can easily induce complications such as hypertension, arrhythmia, diabetes and cognitive impairment, seriously endangering the health and life of patients.
[0003] Polysomnography (PSG) is currently the gold standard for clinical diagnosis. However, it requires overnight monitoring in specialized institutions, which presents challenges such as expensive equipment, cumbersome procedures, and difficulties in widespread adoption at the grassroots level. Furthermore, diagnostic results are easily influenced by physician experience, hindering large-scale early screening. Existing diagnostic algorithms based on physiological signals largely rely on overnight sleep data, requiring patients to be asleep to collect effective data. They can only perform fragmented identification of sleep apnea events, lacking a complete and efficient automated detection process, resulting in high overall detection costs, low efficiency, and poor accessibility. Summary of the Invention
[0004] In view of the above, the purpose of this application is to provide a method, device, electronic device and storage medium for detecting sleep apnea, so as to overcome at least one of the above-mentioned defects.
[0005] In a first aspect, embodiments of this application provide a method for detecting sleep apnea, including: Multi-source physiological signals were collected from conscious patients, including multi-channel electroencephalogram (EEG) and electromyogram (EMG) signals from the cardiopulmonary system. Multi-source physiological signals are filtered to obtain time-frequency diagrams of brain-eye electroencephalograms and respiratory physiological signals corresponding to cardiopulmonary electromyography signals for each channel. Using frequency feature dimensionality reduction models and respiratory feature dimensionality reduction models, feature dimensionality reduction processing was performed on time-frequency diagrams and respiratory physiological signals to obtain frequency features and respiratory features, respectively. Frequency features and breathing features are input into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results.
[0006] In an optional implementation, the steps of using a frequency feature dimensionality reduction model and a respiratory feature dimensionality reduction model to perform feature dimensionality reduction processing on the time-frequency graph and respiratory physiological signal respectively to obtain frequency features and respiratory features include: constructing a respiratory feature dimensionality reduction model and a frequency feature dimensionality reduction model corresponding to each channel in the electroencephalogram and electrooculogram; inputting the time-frequency graph of different frequency bands under each channel into the corresponding frequency feature dimensionality reduction model to obtain the frequency features under that channel; and inputting the respiratory physiological signal into the respiratory feature dimensionality reduction model to obtain respiratory features.
[0007] In an optional implementation, the frequency feature dimensionality reduction model includes a first upper branch, a first middle branch, and a first lower branch. Different frequency bands include low-frequency bands, high-frequency bands, and full-frequency bands. The step of inputting the time-frequency plots of different frequency bands under each channel into the corresponding frequency feature dimensionality reduction model to obtain the frequency features of that channel includes: in the first upper branch, performing feature extraction, feature mapping, and nonlinear transformation processing on the time-frequency plot of the low-frequency band under that channel in sequence to obtain low-frequency gating weights; in the first middle branch, performing feature extraction on the time-frequency plot of the full-frequency band under that channel to obtain full-frequency local features; in the first lower branch, performing feature extraction on the time-frequency plot of the high-frequency band under that channel to obtain high-frequency local features; and multiplying the full-frequency local features element-wise with the low-frequency gating weights and the high-frequency local features respectively, and determining the frequency features based on the multiplication results.
[0008] In an optional implementation, the respiratory physiological signals include electrocardiogram (ECG) and electromyogram (EMG) signals, and respiratory signals. The respiratory feature dimensionality reduction model includes a second upper branch and a second lower branch. The step of inputting the respiratory physiological signals into the respiratory feature dimensionality reduction model to obtain respiratory features includes: in the second upper branch, performing temporal feature extraction and dimensionality reduction processing on the ECG and EMG signals to obtain myocardial temporal features; in the second lower branch, capturing the temporal features used to characterize the respiratory pattern and the key time periods used to characterize the apnea rhythm in the respiratory signal to obtain respiratory temporal features; and concatenating the myocardial temporal features and respiratory temporal features to obtain respiratory features.
[0009] In an optional implementation, the step of multiplying the full-frequency local features element-wise with the low-frequency gating weights and the high-frequency local features respectively, and determining the frequency features based on the multiplication results, includes: multiplying the full-frequency local features element-wise with the low-frequency gating weights to obtain the low-frequency modulated full-frequency features; multiplying the full-frequency local features element-wise with the high-frequency local features to obtain the cross-band fused features; and compressing and fusing the low-frequency modulated full-frequency features and the high-frequency local features to obtain the frequency features.
[0010] In an optional implementation, the sleep apnea detection model includes a feature splicing layer, a third upper branch, and a third lower branch. The steps of inputting frequency features and respiratory features into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results include: in the feature splicing layer, concatenating frequency features and respiratory features to obtain spliced features; in the third upper branch, performing feature dimensionality reduction and global feature extraction on the spliced features to obtain a first detection feature representing the global trend; in the third lower branch, performing temporal feature extraction and max pooling on the spliced features to extract a second detection feature representing local peaks; and multiplying the first and second detection features element-wise, and sequentially performing bidirectional temporal dependency modeling and feature compression to obtain the sleep apnea detection results.
[0011] In an optional implementation, the step of filtering the multi-source physiological signals to obtain the time-frequency diagram corresponding to the electroencephalogram of brain and eye signals for each channel and the respiratory physiological signal corresponding to the electromyography of the heart and lungs includes: filtering and performing short-time Fourier transform processing on the multi-channel electroencephalogram of brain and eye signals to obtain the time-frequency diagram of different frequency bands under each channel; and filtering the electromyography of the heart and lungs to obtain the preprocessed respiratory physiological signal.
[0012] Secondly, embodiments of this application also provide a sleep apnea detection device, the device comprising: The signal acquisition module is used to acquire multi-source physiological signals from patients who are awake. These multi-source physiological signals include multi-channel electroencephalogram (EEG) signals and electrocardiomyogram (ECG) signals. The signal preprocessing module is used to filter multi-source physiological signals to obtain the time-frequency diagram corresponding to the electroencephalogram of the brain and eye signals and the respiratory physiological signal corresponding to the electromyography of the heart and lungs. The feature dimensionality reduction module is used to perform feature dimensionality reduction processing on the time-frequency graph and respiratory physiological signal using the frequency feature dimensionality reduction model and the respiratory feature dimensionality reduction model, respectively, to obtain frequency features and respiratory features. The sleep apnea detection module is used to input frequency features and breathing features into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results.
[0013] Thirdly, embodiments of this application also provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the sleep apnea detection method described above are performed.
[0014] Fourthly, embodiments of this application also provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the sleep apnea detection method described above.
[0015] The embodiments of this application bring the following beneficial effects: This application provides a method, device, electronic device, and storage medium for detecting sleep apnea. It only requires collecting multi-source physiological signals from a patient while awake to diagnose sleep apnea, eliminating reliance on overnight monitoring, significantly reducing detection costs, and facilitating widespread adoption and early screening at the grassroots level. Furthermore, by employing multi-branch deep learning and cross-frequency band feature fusion, it fully leverages the correlation features between EEG, ECG, and respiratory signals, improving detection accuracy and providing stable and reliable technical support for non-invasive, rapid, and automated diagnosis. Compared to existing sleep apnea detection methods, it solves the problems of high detection costs, low efficiency, and poor accessibility.
[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0018] Figure 1 A flowchart of the sleep apnea detection method provided in the embodiments of this application is shown; Figure 2 This paper shows a schematic diagram of the structure of the frequency feature dimensionality reduction model provided in the embodiments of this application; Figure 3 This paper shows a schematic diagram of the structure of the respiratory feature dimensionality reduction model provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of the sleep apnea detection model provided in the embodiments of this application is shown; Figure 5 A schematic diagram of the sleep apnea detection device provided in the embodiments of this application is shown; Figure 6 A schematic diagram of the structure of the electronic device provided in the embodiments of this application is shown. Detailed Implementation
[0019] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.
[0020] To facilitate understanding of this embodiment, the following description uses the sleep apnea detection method provided in this application embodiment applied to a terminal device as an example to illustrate the above exemplary steps provided in this application embodiment.
[0021] Please see Figure 1 , Figure 1 This is a flowchart illustrating a sleep apnea detection method provided in an embodiment of this application. Figure 1 As shown in the embodiments of this application, the sleep apnea detection method includes: Step S101: Collect multi-source physiological signals from a patient who is awake.
[0022] Multi-source physiological signals include multi-channel brain-eye electroencephalogram (EEG) signals and cardiopulmonary electromyography (CM) signals.
[0023] While the patient is awake, a 30-second segment of multi-source physiological signals without a trail is collected. The multi-source physiological signals include multi-channel electroencephalogram (EEG) signals and electrocardiomyogram (ECG) signals.
[0024] For example, the multi-channel electroencephalogram (EEG) signal includes six channels of EEG signals and two channels of EEG signals. The six channels of EEG signals can be EEG signals from six channels: F3, F4, C3, C4, O1, and O2. The two channels of EEG signals can be EEG signals from two channels: left eyeball electroencephalogram (EOG-L) and right eyeball electroencephalogram (EOG-R).
[0025] Cardiopulmonary electromyography (CEMG) is a collective term for electrocardiogram (ECG), respiratory signals, and electromyography (EMG). Among them, respiratory signals include thoracic respiratory signals, abdominal respiratory signals, thermal signals, and nasal airflow signals.
[0026] Compared with existing technologies, the multi-source physiological signal acquisition method of this application has a shorter data acquisition time and does not require the acquisition of physiological data after the patient falls asleep, which greatly improves the detection efficiency and convenience of sleep apnea detection methods.
[0027] Step S102: Filter the multi-source physiological signals to obtain the time-frequency diagram corresponding to the electroencephalogram of brain and eye and the respiratory physiological signal corresponding to the electromyography of the heart and lungs for each channel.
[0028] Filtering of multi-source physiological signals includes filtering of multi-channel electroencephalogram (EEG) signals and filtering of electromyogram (EMG) signals from the heart and lungs.
[0029] In one embodiment, when filtering multi-channel electroencephalogram (EEG) signals, the multi-channel EEG signals can be filtered and processed by short-time Fourier transform to obtain time-frequency diagrams of different frequency bands in each channel.
[0030] For example, the acquired six-channel EEG signals and two-channel EEG signals are filtered separately. The electrical signals of each channel retain data within three preset frequency bands: full frequency (0.3~35Hz), low frequency (0.3~7Hz), and high frequency (8~30Hz). Then, the short-time Fourier transform (STFT) method is used to convert the filtered one-dimensional data within the three preset frequency bands into two-dimensional time-frequency maps. Each time-frequency map can be recognized as a 2D image (frequency × time), which is equivalent to converting the electrical signal data of each channel into three time-frequency maps of different frequency bands.
[0031] In one embodiment, when filtering the cardiopulmonary electromyography (CPM) signal, the CPM signal is filtered and then notch filtering is applied to the filtered CPM signal to obtain a preprocessed respiratory physiological signal.
[0032] For example, ECG signals are filtered to retain those within the 0.3–70 Hz frequency range; electromyography (EMG) signals are filtered to retain those within the 10–100 Hz frequency range; and respiratory signals are filtered to retain those within the 0.1–15 Hz frequency range. Furthermore, the filtered cardiopulmonary EMG signals from all channels can be subjected to a 50 Hz power frequency notch filter to remove power frequency interference.
[0033] Step S103: Using the frequency feature dimensionality reduction model and the respiratory feature dimensionality reduction model, the time-frequency diagram and the respiratory physiological signal are subjected to feature dimensionality reduction processing to obtain the frequency features and respiratory features, respectively.
[0034] In this step, firstly, a dimensionality reduction model of respiratory features and a dimensionality reduction model of frequency features corresponding to each channel in the electroencephalogram of brain and eye are constructed. Then, the frequency features of time-frequency maps of different frequency bands are extracted by the frequency feature dimensionality reduction model, and the respiratory features of respiratory physiological signal related channels are extracted by the respiratory feature dimensionality reduction model.
[0035] The following reference Figure 2 This section introduces the frequency feature dimensionality reduction model.
[0036] Figure 2 The diagram shows a schematic representation of the frequency feature dimensionality reduction model provided in this application embodiment. Figure 2 As shown, the frequency feature dimensionality reduction model includes a first upper branch 210, a first middle branch 220, and a first lower branch 230. The first upper branch 210 includes a first convolutional layer 211, a first fully connected layer 212, a first activation function 213, a second convolutional layer 214, and a first pooling layer 215; the first middle branch 220 includes a third convolutional layer 221; and the first lower branch 230 includes a fourth convolutional layer 231, a second activation function 232, a fifth convolutional layer 233, and a second pooling layer 234.
[0037] When performing feature dimensionality reduction on the time-frequency maps, each EEG or EOS channel corresponds to a frequency feature dimensionality reduction model. Furthermore, each channel's frequency feature dimensionality reduction model employs a three-branch deep network with identical structure and independent parameters. Therefore, the eight channels of EEG / EOS signals correspond to eight independent frequency feature dimensionality reduction models. In this way, the time-frequency maps of different frequency bands under each channel can be input into the corresponding frequency feature dimensionality reduction model to obtain the frequency features of that channel.
[0038] For example, when performing feature dimensionality reduction processing on the EEG signal of the F3 channel, the low-frequency band time-frequency map corresponding to the EEG signal of 0.3~7HZ in the F3 channel is input into the first upper branch, the full-frequency band time-frequency map corresponding to the EEG signal of 0.3~35HZ in the F3 channel is input into the first middle branch, and the high-frequency band time-frequency map corresponding to the EEG signal of 8~30HZ in the F3 channel is input into the first lower branch to obtain the frequency characteristics of the EEG signal in the F3 channel.
[0039] In one embodiment, in the first upper branch of the frequency feature dimensionality reduction model, feature extraction, feature mapping, and nonlinear transformation are sequentially performed on the time-frequency graph of the low-frequency band under the channel to obtain low-frequency gating weights; in the first middle branch, feature extraction is performed on the time-frequency graph of the full-frequency band under the channel to obtain full-frequency local features; in the first lower branch, feature extraction is performed on the time-frequency graph of the high-frequency band under the channel to obtain high-frequency local features; the full-frequency local features are multiplied element-wise with the low-frequency gating weights and the high-frequency local features respectively, and the frequency features are determined based on the multiplication results.
[0040] For example: A first convolutional layer extracts local features from the low-frequency time-frequency map; a first fully connected layer performs global feature mapping and dimension reconstruction on the local features extracted by the first convolutional layer; and a first activation function performs non-linear normalization on the features mapped by the first fully connected layer to obtain low-frequency gating weights for weighted modulation. Simultaneously, a third convolutional layer extracts local features from the full-frequency time-frequency map to obtain full-frequency local features, and a fourth convolutional layer extracts local features from the high-frequency time-frequency map to obtain high-frequency local features.
[0041] The full-frequency local features are multiplied element-wise with the low-frequency gated weights to obtain the low-frequency modulated full-frequency features. The full-frequency local features are then multiplied element-wise with the high-frequency local features to obtain the cross-frequency band fused features. The low-frequency modulated full-frequency features and the high-frequency local features are then compressed and fused to obtain the frequency features.
[0042] In one embodiment, when compressing and fusing the full-frequency features after low-frequency modulation and the local features of high frequencies, the second convolutional layer can be used in the first upper branch to extract local features and perform cross-modal fusion on the full-frequency features after low-frequency modulation, so as to capture the local patterns and correlations of signals in different frequency bands in the time-frequency dimension, enhance effective features, and suppress noise interference. The upper branch compressed features are obtained by performing spatial dimension compression and feature invariance enhancement on the feature map output by the second convolutional layer through the first pooling layer, which reduces the number of parameters and computational load, while retaining key features and avoiding model overfitting.
[0043] In the first lower branch, the cross-band fusion features are nonlinearly transformed using the second activation function to suppress invalid or noisy features and select effective features. The fifth convolutional layer captures the local correlation and subtle patterns of the cross-band fusion features output by the second activation function in the time-frequency dimension, and abstracts the output cross-band fusion features into higher-level semantic features to enhance the interaction information between different frequency bands. The second pooling layer downsamples the feature map output by the fifth convolutional layer to obtain the compressed features of the lower branch, which greatly reduces the feature dimension and computational cost, while retaining key global features and mitigating the risk of model overfitting.
[0044] Finally, the upper branch compression feature output from the first upper branch is multiplied element-wise with the lower branch compression feature output from the first lower branch to obtain the frequency feature.
[0045] Among them, the first, second, third, fourth, and fifth convolutional layers are all two-dimensional convolutional layers (Conv2d); the first fully connected layer is a fully connected layer (FC); the first and second activation functions are both hyperbolic tangent activation functions (Tanh); and the first and second pooling layers are both max pooling layers or average pooling layers.
[0046] By performing the above processing on each channel, a total of eight frequency characteristics can be obtained for the eight channels.
[0047] The following reference Figure 3 Let's introduce the dimensionality reduction model of respiratory features.
[0048] Figure 3 The diagram shows a schematic representation of the respiratory feature dimensionality reduction model provided in this application embodiment. Figure 3 As shown, the respiratory feature dimensionality reduction model includes a second upper branch 310, a second lower branch 320, and a first pooling layer 301. The second upper branch 310 includes a sixth convolutional layer 311, a seventh convolutional layer 312, and a third pooling layer 313; the second lower branch 320 includes a first bidirectional long short-term memory network 321, an attention mechanism module 322, and a long short-term memory network 323.
[0049] When performing feature dimensionality reduction processing on respiratory physiological signals, the respiratory physiological signals are input into the respiratory feature dimensionality reduction model to obtain respiratory features, which include electrocardiogram and electromyography signals and respiratory signals.
[0050] For example, ECG and EMG signals can be input into the second upper branch, and respiratory signals can be input into the second lower branch. The processing results of the second upper branch and the second lower branch can be spliced together to obtain respiratory characteristics.
[0051] In one embodiment, in the second upper branch of the respiratory feature dimensionality reduction model, temporal features are extracted and dimensionality reduced from electrocardiogram and electromyography signals to obtain myocardial temporal features; in the second lower branch, temporal features used to characterize respiratory patterns and key time periods used to characterize apnea rhythms in respiratory signals are captured to obtain respiratory temporal features; the myocardial temporal features and respiratory temporal features are concatenated to obtain respiratory features.
[0052] For example, in the second upper branch, local features of ECG and EMG signals are extracted through the sixth and seventh convolutional layers. The local features extracted by the seventh convolutional layer are then dimensionality-reduced through the third pooling layer, while retaining key features to obtain the first respiratory feature. In this way, the second upper branch can compress the sequence length, reduce computation, highlight the most significant features (such as peaks and abnormal fluctuations) in each local time window, and enhance the model's robustness to time shifts.
[0053] In the second branch, the respiratory signal is modeled bidirectionally using a first bidirectional long short-term memory network (LSTM). This network captures long-term contextual information of the respiratory rhythm from both past and future time directions, avoiding the limitation of a unidirectional LSTM network that can only utilize historical information. This allows for a more complete preservation of the temporal features of the respiratory signal. An attention mechanism module adaptively assigns weights to the temporal features output by the first bidirectional LSTM network, automatically focusing on key time segments in the respiratory signal related to apnea or abnormal rhythms. This strengthens effective information, suppresses redundant noise, and improves the specificity and discriminative power of the features. Based on the attention-weighted features, the unidirectional temporal features are further refined using the LSTM network to obtain second respiratory features. These features focus on the temporal evolution of the respiratory signal and compress the attention-enhanced key temporal information into a higher-level respiratory feature vector, preparing for subsequent multimodal fusion.
[0054] Using the first splicing layer, the first breathing feature output from the second upper branch and the second breathing feature output from the second lower branch are spliced together to generate a breathing feature.
[0055] Both the sixth and seventh convolutional layers are one-dimensional convolutions (Conv1d), and their activation functions are both Exponential Linear Units (ELUs). The first bidirectional long short-term memory network can refer to Bi-LSTM (Bidirectional Long Short-Term Memory), and the long short-term memory network can refer to LSTM (Long Short-Term Memory). The activation functions of the first bidirectional long short-term memory network and the long short-term memory network are Rectified Linear Units (ReLUs).
[0056] Step S104: Input the frequency features and breathing features into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results.
[0057] Because EEG signals differ across different brain regions, sleep apnea patients exhibit characteristic abnormal respiratory rhythms, and different physiological signal channels contribute differently to apnea detection, an apnea detection model can be constructed based on the differences in EEG signals across different brain regions, the respiratory rhythms of sleep apnea patients, and the roles of different channels in apnea detection. This model can then achieve deep fusion and discrimination of multi-channel and multi-modal features, enabling rapid detection of sleep apnea during wakefulness.
[0058] The following reference Figure 4 Let's introduce the sleep apnea detection model.
[0059] Figure 4A schematic diagram of the structure of the sleep apnea detection model provided in the embodiments of this application is shown, as follows: Figure 4 As shown, the sleep apnea detection model includes a feature splicing layer 401, a third upper branch 410, a third lower branch 420, a flattening layer 431, a second bidirectional long short-term memory network 432, a second fully connected layer 433, and a third fully connected layer 434. The third upper branch 410 includes an eighth convolutional layer 411 and a fourth pooling layer 412, and the third lower branch 420 includes a ninth convolutional layer 421 and a fifth pooling layer 422.
[0060] When using the sleep apnea detection model to determine the sleep apnea detection result, frequency features and breathing features can be concatenated in the feature concatenation layer to obtain concatenated features. In the third upper branch, feature dimensionality reduction and global feature extraction are performed on the concatenated features to obtain the first detection feature used to characterize the global trend. In the third lower branch, temporal feature extraction and max pooling are performed on the concatenated features to extract the second detection feature used to characterize the local peak. The first detection feature and the second detection feature are multiplied element-wise, and bidirectional temporal dependency modeling and feature compression are performed sequentially to obtain the sleep apnea detection result.
[0061] For example, frequency features and breathing features are concatenated in the feature concatenation layer, and the concatenated features are input into the third upper branch and the third lower branch respectively.
[0062] In the third branch, the eighth convolutional layer is used to extract features from the spliced features to extract the local average pattern and smoothing trend of the temporal features and preserve the overall rhythmic information of the signal. The fourth pooling layer is used to perform average pooling on the features extracted by the eighth convolutional layer to obtain the first detection features, so as to compress the sequence length, smooth noise, stabilize the feature distribution, and thus extract features to characterize the global trend.
[0063] In the third lower branch, the ninth convolutional layer is used to extract features from the spliced features; the fifth pooling layer is used to perform max pooling on the features extracted by the ninth convolutional layer to obtain the second detection features, so as to compress the sequence length, smooth noise, stabilize feature distribution, and thus extract key features to characterize local peaks.
[0064] The first and second detection features are multiplied element-wise, and the resulting detection feature is input into a flattening layer. The flattening layer converts the multiplied detection feature into a one-dimensional feature vector to eliminate the spatial-temporal dimensional structure. A second bidirectional long short-term memory network is used to model the bidirectional temporal dependency of the flattened one-dimensional feature vector, capturing long-term contextual associations between features from both historical and future time directions, fully preserving temporal dependencies and improving the ability to model temporal patterns. The temporal dependency features output by the second bidirectional long short-term memory network are input into a second fully connected layer and a third fully connected layer. The second fully connected layer performs global feature mapping and dimensional transformation on the temporal dependency features output by the second bidirectional long short-term memory network to compress the temporal dependency features into high-level semantic features, realizing the conversion of temporal information into discriminative features. The third fully connected layer maps the high-level semantic features output by the second fully connected layer into sleep apnea detection results.
[0065] Among them, the eighth and ninth convolutional layers are one-dimensional convolutional layers (Conv1d); the fourth pooling layer is an average pooling layer; the fifth pooling layer is a max pooling layer; the second bidirectional long short-term memory network is Bi-LSTM; the activation function of the eighth and ninth convolutional layers is ELU; and the activation function of the second fully connected layer, the third fully connected layer, and the second bidirectional long short-term memory network is ReLU.
[0066] The sleep apnea detection method provided in this application has the following effects: First, it only requires collecting 30 seconds of short-range multi-source physiological signals from the patient (subject) while they are awake to complete the detection of sleep apnea syndrome. This eliminates the limitations of traditional polysomnography (PSG), which requires monitoring all night and relies on professional equipment and facilities. It significantly reduces equipment investment, time costs, and labor costs, making it suitable for large-scale use in primary healthcare institutions.
[0067] Secondly, it breaks through the limitation of existing technologies that can only identify sleep apnea events during sleep. It can capture the differences in brain waves, abnormal breathing rhythms and latent characteristics of multi-channel physiological signals in patients with sleep apnea while they are awake, so as to achieve early screening, risk warning and personalized intervention.
[0068] Third, it integrates multi-channel signals from EEG, EEG, ECG, respiration, and EMG, and uses a three-level deep learning architecture of frequency feature dimensionality reduction, respiratory feature dimensionality reduction, and channel feature integration to comprehensively extract time-frequency, rhythm, temporal, and cross-frequency interaction features. Its stability and discrimination ability are superior to traditional single-signal and single-feature methods.
[0069] Fourth, it employs convolution, pooling, and attention mechanisms with Bi... The lightweight structure combined with LSTM is highly robust to signal noise and time-series shifts; the feature dimensionality reduction and efficient fusion strategy significantly reduces computational load, supports real-time and rapid diagnosis, and is easy to deploy in portable devices and mobile terminals.
[0070] Fifth, by fully utilizing the differences in EEG signals across different brain regions and the diagnostic contributions of different physiological channels, channel-specific modeling and deep cross-channel fusion have been achieved, which not only improves detection accuracy but also preserves clear pathophysiological significance, making it easier for clinical understanding and application.
[0071] Based on the same inventive concept, this application also provides a sleep apnea detection device corresponding to the sleep apnea detection method. Since the principle of the device in this application is similar to the sleep apnea detection method described above, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.
[0072] Please see Figure 5 , Figure 5 This is a schematic diagram of a sleep apnea detection device provided in an embodiment of this application. Figure 5 As shown, the sleep apnea detection device 500 includes: The signal acquisition module 501 is used to acquire multi-source physiological signals from a patient in an conscious state. The multi-source physiological signals include multi-channel electroencephalogram (EEG) and electrocardiomyogram (ECG) signals. The signal preprocessing module 502 is used to filter multi-source physiological signals to obtain the time-frequency diagram corresponding to the electroencephalogram of brain and eye signals and the respiratory physiological signal corresponding to the electromyography of the heart and lungs. The feature dimensionality reduction module 503 is used to perform feature dimensionality reduction processing on the time-frequency diagram and respiratory physiological signal respectively using the frequency feature dimensionality reduction model and the respiratory feature dimensionality reduction model to obtain frequency features and respiratory features. The sleep apnea detection module 504 is used to input frequency features and breathing features into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results.
[0073] Please see Figure 6 , Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 6 As shown, the electronic device 600 includes a processor 610, a memory 620, and a bus 630.
[0074] The memory 620 stores machine-readable instructions executable by the processor 610. When the electronic device 600 is running, the processor 610 and the memory 620 communicate via the bus 630. When the machine-readable instructions are executed by the processor 610, they can perform the operations described above. Figure 1 The specific implementation of the sleep apnea detection method in the illustrated method embodiment can be found in the method embodiment, and will not be repeated here.
[0075] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The specific implementation of the sleep apnea detection method in the illustrated method embodiment can be found in the method embodiment, and will not be repeated here.
[0076] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.
[0077] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.
[0078] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0079] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0080] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0081] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A sleep apnea detection method, characterized by, include: Collect multi-source physiological signals from a patient in a conscious state, including multi-channel electroencephalogram (EEG) and electromyogram (EMG) signals from the cardiopulmonary system. The multi-source physiological signals are filtered to obtain the time-frequency diagrams corresponding to the electroencephalograms of the brain and eyes for each channel and the respiratory physiological signals corresponding to the electromyograms of the heart and lungs. Using frequency feature dimensionality reduction models and respiratory feature dimensionality reduction models, feature dimensionality reduction processing is performed on the time-frequency diagram and the respiratory physiological signal, respectively, to obtain frequency features and respiratory features. The frequency features and breathing features are input into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results.
2. The method according to claim 1, characterized in that, The step of using a frequency feature dimensionality reduction model and a respiratory feature dimensionality reduction model to perform feature dimensionality reduction processing on the time-frequency graph and the respiratory physiological signal, respectively, to obtain frequency features and respiratory features, includes: Construct a dimensionality reduction model of respiratory features and a dimensionality reduction model of frequency features corresponding to each channel in electroencephalogram (EEG) signals. Input the time-frequency plots of different frequency bands under each channel into the corresponding frequency feature dimensionality reduction model to obtain the frequency features under that channel; The respiratory physiological signals are input into the respiratory feature dimensionality reduction model to obtain respiratory features.
3. The method according to claim 2, characterized in that, The frequency feature dimensionality reduction model includes a first upper branch, a first middle branch, and a first lower branch. The different frequency bands include low-frequency bands, high-frequency bands, and full-frequency bands. The step of inputting the time-frequency plots of different frequency bands under each channel into the corresponding frequency feature dimensionality reduction model to obtain the frequency features under that channel includes: In the first upper branch, the time-frequency graph of the low-frequency band under the channel is sequentially subjected to feature extraction, feature mapping and nonlinear transformation to obtain the low-frequency gating weight; In the first branch, feature extraction is performed on the time-frequency map of the full frequency band under the channel to obtain full-frequency local features; In the first lower branch, feature extraction is performed on the time-frequency map of the high-frequency band under the channel to obtain high-frequency local features; The full-frequency local features are multiplied element-wise with the low-frequency gating weights and the high-frequency local features, and the frequency features are determined based on the multiplication results.
4. The method according to claim 2, characterized in that, The respiratory physiological signals include electrocardiogram and electromyogram signals, and respiratory signals. The respiratory feature dimensionality reduction model includes a second upper branch and a second lower branch. The step of inputting the respiratory physiological signals into the respiratory feature dimensionality reduction model to obtain respiratory features includes: In the second upper branch, the electrocardiogram and electromyography signals are subjected to temporal feature extraction and dimensionality reduction processing to obtain myocardial temporal features; In the second lower branch, the temporal features used to characterize the breathing pattern and the key time periods used to characterize the apnea rhythm in the breathing signal are captured to obtain the breathing temporal features; The respiratory features are obtained by concatenating the myocardial time sequence features with the respiratory time sequence features.
5. The method according to claim 3, characterized in that, The step of multiplying the full-frequency local features element-wise with the low-frequency gating weights and the high-frequency local features respectively, and determining the frequency features based on the multiplication results, includes: The full-frequency local features are multiplied element-wise with the low-frequency gate weights to obtain the low-frequency modulated full-frequency features. The full-frequency local features and the high-frequency local features are multiplied element-wise to obtain cross-frequency band fusion features; The full-frequency features after low-frequency modulation and the local features of high frequency are compressed and fused to obtain frequency features.
6. The method according to claim 1, characterized in that, The sleep apnea detection model includes a feature splicing layer, a third upper branch, and a third lower branch. The step of inputting the frequency features and the breathing features into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results includes: In the feature splicing layer, the frequency feature and the breathing feature are spliced together to obtain the spliced feature; In the third upper branch, feature dimensionality reduction and global feature extraction are performed on the spliced features to obtain the first detection feature used to characterize the global trend; In the third lower branch, temporal feature extraction and max pooling are performed on the spliced features to extract a second detection feature for characterizing local peaks; The first detection feature and the second detection feature are multiplied element-wise, and bidirectional temporal dependency modeling and feature compression are performed sequentially to obtain the sleep apnea detection result.
7. The method according to claim 1, characterized in that, The step of filtering the multi-source physiological signals to obtain the time-frequency diagram corresponding to the electroencephalogram of each channel and the respiratory physiological signal corresponding to the electromyography of the heart and lungs includes: The multi-channel electroencephalogram (EEG) signals are filtered and processed using short-time Fourier transform to obtain time-frequency diagrams for different frequency bands in each channel. The cardiopulmonary electromyography signals are filtered to obtain preprocessed respiratory physiological signals.
8. A sleep apnea detection device, characterized in that, include: The signal acquisition module is used to acquire multi-source physiological signals from a patient in a conscious state, including multi-channel electroencephalogram (EEG) and electrocardiomyogram (ECG) signals. The signal preprocessing module is used to filter the multi-source physiological signals to obtain the time-frequency diagram corresponding to the electroencephalogram of the brain and eye signals and the respiratory physiological signal corresponding to the electromyography of the heart and lungs. The feature dimensionality reduction module is used to perform feature dimensionality reduction processing on the time-frequency graph and the respiratory physiological signal using a frequency feature dimensionality reduction model and a respiratory feature dimensionality reduction model, respectively, to obtain frequency features and respiratory features. The sleep apnea detection module is used to input the frequency features and the breathing features into the sleep apnea detection model for feature fusion and temporal dependency capture to obtain sleep apnea detection results.
9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the sleep apnea detection 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 computer program that, when executed by a processor, performs the steps of the sleep apnea detection method as described in any one of claims 1 to 7.