Apnea hypopnea detection system and method

Abstract

The invention provides an apnea hypopnea detection system and method, and relates to the technical field of medical detection, and the system comprises a preprocessing module which is used for preprocessing an original ballistocardiogram signal to obtain an optimized ballistocardiogram signal; the change point detection module is used for positioning suspected event signal segments from the optimized ballistocardiogram signals; the feature extraction module is used for extracting time-frequency features and time-domain sequence features from the event signal segments; the classification decision module performs key feature enhancement and multi-scale time sequence mode extraction on the time-frequency features through a time-frequency analysis branch, outputs time-frequency morphological features, captures bidirectional time sequence dependence and key time point information from time-domain sequence features through a sequence feature branch, outputs complexity features, and performs classification on the time-frequency features; and fusing the time-frequency morphological features with the complexity features to obtain fused features, and determining a respiratory event classification result through a classifier according to the fused features. The breathing event can be conveniently and accurately detected through the system.

Description

Technical Field

[0001] This application relates to the field of medical testing technology, and in particular to a sleep apnea-hypopnea detection system and method. Background Technology

[0002] Sleep apnea-hypopnea syndrome (SAHS) is a common sleep disorder characterized by recurrent episodes of apnea and hypopnea during sleep.

[0003] Traditional diagnostic methods primarily rely on polysomnography (PSG), which requires connecting multiple sensors to the patient's body, including sensors for electroencephalography (EEG), electrooculography (EOG), electromyography (EMG), chest and abdominal respiratory movements, nasal and oral airflow, and blood oxygen saturation. This complex connection can severely disrupt the patient's normal sleep, leading to the "first night effect" and affecting diagnostic accuracy. Furthermore, PSG requires specialized sleep laboratories, where the equipment is expensive and requires professional operation, making it difficult to implement in home settings. Summary of the Invention

[0004] In view of this, this application proposes a sleep apnea-hypopnea detection system and method.

[0005] In a first aspect, this application provides an apnea-hypopnea detection system, which includes: a preprocessing module, a change point detection module, a feature extraction module, and a classification decision module; The preprocessing module is used to acquire the raw cardiac impaction signal obtained from detecting the user's respiratory events, and to preprocess the raw cardiac impaction signal to obtain an optimized cardiac impaction signal; The change point detection module is used to locate the suspected event signal segment corresponding to the suspected apnea event from the optimized cardiac impact map signal using a preset change point detection algorithm; The feature extraction module is used to extract time-frequency features and time-domain sequence features from each of the suspected event signal fragments; The classification decision module enhances key features and extracts multi-scale temporal patterns from the time-frequency features through the time-frequency analysis branch, and outputs time-frequency morphological features; it captures bidirectional temporal dependence and key time point information from the time-domain sequence features through the sequence feature branch, and outputs complexity features; it fuses the time-frequency morphological features and the complexity features to obtain fused features; and it determines the classification result of respiratory events through a classifier based on the fused features.

[0006] In one embodiment, the preprocessing module includes: a 4th-order Butterworth bandpass filter and a quality assessment unit; The fourth-order Butterworth bandpass filter is used to filter the original cardiac impaction signal to obtain a filtered signal; The quality assessment unit is used to calculate the signal quality index of each time segment in the filtered signal based on a preset assessment algorithm, mark the time segments with a signal quality index lower than a preset quality threshold as low-quality segments, and remove each low-quality segment to obtain the optimized cardiac impaction signal.

[0007] In one embodiment, the change point detection module includes: a segmentation unit, a detection unit, and a fragment extraction unit; The segmentation unit is used to segment the optimized cardiac impaction signal into multiple non-overlapping time-series signal segments of a predetermined time period, and calculate the mean, variance, and sample entropy of each time-series signal segment as variation features. The detection unit is used to detect abrupt changes in the changing features using a Bayesian online change point detection algorithm; The segment extraction unit is used to extract signal segments at preset times before and after the mutation point to form the suspected event signal segment.

[0008] In one embodiment, the classification decision module includes: a dual-branch neural network, which includes: the time-frequency analysis branch, the sequence feature branch, the feature fusion unit, and the classifier; The time-frequency analysis branch is used to perform feature enhancement and multi-scale temporal pattern extraction on the time-frequency features using a lightweight CNN network, and outputs the time-frequency morphological features; The sequence feature branch is used to capture bidirectional temporal dependencies from the temporal sequence features through a bidirectional gated recurrent unit, capture key time point information from the temporal sequence features through a temporal attention mechanism, and output the complexity feature; The feature fusion unit is used to fuse the time-frequency morphological features with the complexity features to obtain fused features; The classifier is used to determine the classification result of respiratory events based on the fusion features.

[0009] In one embodiment, the time-frequency analysis branch includes: a first input layer, a standard convolutional layer, a first depthwise separable convolutional block, a second depthwise separable convolutional block, a multi-scale residual block, and a global average pooling layer; The first input layer is used to receive the time-frequency features and transmit them to the standard convolutional layer; The standard convolutional layer is used to perform fast downsampling and primary feature detection on the time-frequency features, and output primary features; The first depthwise separable convolutional block is used to extract channel-level features from the primary features and, in conjunction with the SE attention mechanism, calibrate the feature response to output the first depthwise features; The second depthwise separable convolutional block is used to perform secondary downsampling on the first depth features to expand the receptive field and capture intermediate abstract time-frequency semantics, outputting the second depth features; The multi-scale residual block is used to simultaneously capture time-frequency patterns at different scales through multi-size convolutional paths; The global average pooling layer is used to aggregate spatial feature information based on the output features of the multi-scale residual block and output the time-frequency morphological features.

[0010] In one embodiment, the sequence feature branch includes: a second input layer, a bidirectional GRU layer, an attention layer, and a first fully connected layer; The second input layer is used to receive the time-domain sequence features and transmit them to the bidirectional GRU layer; The bidirectional GRU layer is used to process the time-domain sequence features bidirectionally, capture forward and backward temporal dependencies, and output multi-dimensional temporal features. The attention layer is used to calculate the attention weight of each time step in the multidimensional temporal feature, perform a weighted summation of the attention weights of each time step, determine the key time point information of the multidimensional temporal feature, and output the attention-weighted temporal feature. The first fully connected layer is used to map the attention-weighted temporal features to a semantic space aligned with the time-frequency analysis branch, and output the complexity features.

[0011] In one embodiment, the classifier includes: a second fully connected layer and an output layer; The second fully connected layer is used to perform advanced nonlinear interactive learning on the fused features and suppress overfitting through Dropout to output classification features that fit the classification. The output layer is used to map the classification features to a probability distribution of each classification result using a Softmax activation function, and to determine the classification result of the respiratory event based on the probability of each classification result.

[0012] In one embodiment, the apnea-hypopnea detection system further includes: a diagnostic module; The diagnostic module is used to obtain the user's total sleep time, calculate the AHI index based on the respiratory event classification results obtained within the sleep cycle and the total sleep time, and determine the respiratory abnormality grading result based on the AHI index.

[0013] Secondly, this application also provides a method for detecting apnea-hypopnea, comprising: Obtain the training sample set; Based on the training sample set and the weighted cross-entropy loss function, the apnea-hypopnea detection system as described in the first aspect is trained using a domain adversarial training strategy to obtain the target apnea-hypopnea detection model. The original cardiac impaction signal is input into the target apnea-hypopnea detection model to obtain respiratory event classification results.

[0014] Thirdly, this application also provides a method for detecting apnea-hypopnea, comprising: The raw cardiac impactogram signal obtained from detecting user breathing events is acquired, and the raw cardiac impactogram signal is preprocessed to obtain an optimized cardiac impactogram signal; A preset change point detection algorithm is used to locate the suspected event signal segment corresponding to the suspected apnea event from the optimized cardiac impact map signal; Time-frequency features and time-domain sequence features are extracted from each of the suspected event signal fragments; The time-frequency analysis branch enhances key features and extracts multi-scale temporal patterns from the time-frequency features, outputting time-frequency morphological features. The sequence feature branch captures bidirectional temporal dependence and key time point information from the time-domain sequence features, outputting complexity features. The time-frequency morphological features and the complexity features are fused to obtain fused features. The respiratory event classification result is determined by a classifier based on the fused features.

[0015] The apnea-hypopnea detection system of this application has the following advantages over related technologies: 1. The apnea-hypopnea detection system of this application uses BCG signals as the core detection carrier, naturally possessing the core advantages of being contactless and comfortable. It completely eliminates the constraints of traditional PSG systems that rely on multiple sensors connected to the body, avoiding interference with the patient's sleep caused by complex connections, fundamentally reducing diagnostic errors caused by the "first night effect," and requiring no professional environment or operation of a dedicated sleep laboratory. It perfectly adapts to the long-term monitoring needs of home settings, effectively solving the core pain point that makes PSG systems difficult to popularize.

[0016] 2. The preprocessing module optimizes the original BCG signal, filtering out noise and improving signal quality, providing a reliable data foundation for subsequent accurate detection and ensuring the accuracy of feature extraction and classification decisions. The change point detection module uses a preset algorithm to accurately locate suspected apnea event segments, performing in-depth analysis only on suspected segments, significantly reducing the amount of invalid data processed and significantly reducing the overall computational load of the system, laying the foundation for real-time detection. After the feature extraction module extracts time-frequency features and time-domain sequence features from each suspected event signal segment, the classification decision module enhances key features and extracts multi-scale time-series patterns through the time-frequency analysis branch, while capturing the bidirectional time-series dependency and key time point information of the time-domain sequence through the sequence feature branch. Then, feature fusion is used to achieve deep integration of the two types of core features, completely solving the problems of insufficient feature extraction and inadequate adaptation to the non-stationary characteristics of BCG signals in traditional algorithms, greatly improving the accuracy of respiratory event recognition. Finally, based on the fused features, the classifier outputs accurate respiratory event classification results, further ensuring the reliability of the diagnosis.

[0017] 3. Through the synergistic effect of its various modules, the entire system inherits the contactless advantages of BCG technology and is adapted to home scenarios. It ensures high detection accuracy through a refined feature extraction and fusion mechanism, and significantly reduces computational complexity through a suspected fragment focusing design. It successfully achieves the dual goals of high accuracy and low computing power. It can be flexibly deployed on embedded devices to achieve real-time operation, effectively making up for the shortcomings of traditional PSG devices, such as high cost, complex operation, and poor accessibility. It provides a feasible solution for convenient and accurate diagnosis of sleep apnea-hypopnea syndrome. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the structure of an apnea-hypopnea detection system in one embodiment of this application; Figure 2 This is a schematic diagram of the processing flow of the apnea-hypopnea detection system in one embodiment of this application; Figure 3 This is a schematic diagram of the structure of a dual-branch neural network in one embodiment of this application; Figure 4 This is a schematic diagram of the time-frequency analysis branch in one embodiment of this application; Figure 5This is a schematic diagram of the structure of a sequence feature branch in one embodiment of this application; Figure 6 This is a flowchart illustrating the method for detecting apnea-hypopnea in one embodiment of this application. Figure 7 This is a flowchart illustrating the method for detecting apnea-hypopnea in another embodiment of this application.

[0020] Explanation of reference numerals in the attached figures: 10-Sleep apnea-hypopnea detection system, 11-Preprocessing module, 12-Change point detection module, 13-Feature extraction network, 14-Classification decision module. Detailed Implementation

[0021] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0022] In some embodiments, such as Figure 1 As shown, this application provides an apnea-hypopnea detection system 10, which includes: a preprocessing module 11, a change point detection module 12, a feature extraction module 13, and a classification decision module 14. In one example, the processing flow of the apnea-hypopnea detection system can be as follows: Figure 2 As shown.

[0023] The preprocessing module 11 is used to acquire the raw cardiac impact map (BCG) signal obtained from detecting the user's respiratory events, and to preprocess the raw cardiac impact map signal to improve the signal quality, provide reliable input for subsequent analysis, and obtain an optimized cardiac impact map signal.

[0024] In this application, a monitoring pad equipped with a fiber optic micro-vibration sensor can be used to collect BCG signals. The monitoring pad is installed under the mattress; the vibrations from the monitored person's heartbeat and respiration cause slight bending of the optical fiber, resulting in changes in the light intensity passing through the fiber. The monitoring pad collects the changing light intensity signal, thereby reconstructing the waveforms of the heartbeat and respiration. This raw BCG signal directly reflects the micro-movement characteristics of human physiology, but it inevitably contains redundant information such as environmental interference, sensor noise, and low-frequency baseline drift, which may affect the accuracy of subsequent feature extraction and event recognition. Therefore, preprocessing can include noise reduction, and other processing steps can be added as needed.

[0025] The change point detection module 12 is used to locate the suspected event signal segment corresponding to the suspected apnea event from the optimized cardiac impact map signal using a preset change point detection algorithm.

[0026] Among them, the preset change point detection algorithm can be designed based on technical approaches such as statistically significant mutation identification or feature difference comparison of sliding windows.

[0027] It is understandable that when a person experiences apnea-hypopnea, the amplitude of chest movement and the state of cardiac mechanical conduction will undergo characteristic changes. These changes will be directly reflected in key dimensions such as amplitude fluctuations, frequency distribution, and waveform morphology of the cardiac impactogram signal. The preset change point detection algorithm can keenly capture abrupt change nodes or continuous abnormal intervals in these dimensions that deviate from the normal physiological fluctuation pattern. Based on this, the corresponding signal segments can be delineated, which can achieve preliminary screening of massive continuous optimized cardiac impactogram signals, filter out normal signal segments without clinical significance, significantly reduce the computational redundancy of the subsequent feature extraction module 13 for feature extraction, and at the same time ensure that potential abnormal signals of apnea-hypopnea are not missed.

[0028] The feature extraction module 13 is used to extract time-frequency features and time-domain sequence features from each suspected event signal segment. The feature extraction module 13 simultaneously extracts two types of core basic features from each signal segment: one type is time-frequency features that can reflect the frequency distribution and time correlation characteristics of the signal, and the other type is time-domain sequence features that reflect the continuous change law of the signal over time.

[0029] The classification decision module 14 uses a dual-branch neural network (DBNN) as its core. It enhances key features and extracts multi-scale temporal patterns from time-frequency features through the time-frequency analysis branch, and outputs time-frequency morphological features. It captures bidirectional temporal dependence and key time point information from time-domain sequence features through the sequence feature branch, and outputs complexity features. It fuses the time-frequency morphological features and complexity features to obtain fused features. Based on the fused features, it determines the classification result of respiratory events through a classifier.

[0030] The time-frequency analysis branch performs in-depth processing on the extracted time-frequency features. Attention mechanisms can be used to amplify key time-frequency information highly correlated with sleep apnea events and suppress interference from irrelevant noise features. Multi-scale temporal pattern extraction technology can be employed to mine the time-frequency evolution patterns of signals across different time windows, ensuring the network can simultaneously perceive anomalies at different time scales. The final integrated output provides time-frequency morphological features that accurately characterize the waveform shape and frequency composition of the signal. The sequence feature branch, running parallel to the time-frequency analysis branch, focuses on time-domain sequence features. Utilizing a network structure with bidirectional information capture capabilities, it deeply mines the temporal dependencies of the signal and accurately locates key time points such as abrupt amplitude changes and trend reversals, outputting complexity features that reflect the complexity of the signal's temporal evolution. Finally, the classification decision module 14 uses a pre-defined feature fusion strategy to organically integrate the time-frequency morphological features and complexity features extracted from different dimensions, generating high-dimensional fusion features that simultaneously encompass both time-frequency morphological information and time-domain sequence patterns, providing comprehensive and highly recognizable feature support for subsequent classification.

[0031] The classification decision module 14 can match the currently input fusion features with the feature templates of various respiratory events based on the feature distribution patterns of different respiratory events (such as normal breathing, obstructive apnea, central apnea, hypoventilation, etc.) learned during the training phase. Finally, based on the preset judgment threshold and classification criteria, it outputs a clear respiratory event classification result, providing a direct and reliable decision basis for subsequent clinical diagnosis, health risk assessment or home monitoring and early warning.

[0032] The apnea-hypopnea detection system 10 of this application uses cardiac impulse map (BCG) signals as the core detection carrier, naturally possessing the core advantages of being contactless and comfortable. It completely breaks free from the constraints of traditional PSG systems that rely on multiple sensors connected close to the body, perfectly adapting to the long-term monitoring needs of home settings. Based on this, the preprocessing module 11 optimizes the raw BCG signal, filtering out noise and improving signal quality, providing a reliable data foundation for subsequent accurate detection and ensuring the accuracy of feature extraction and classification decisions. The change point detection module 12 uses a preset algorithm to accurately locate suspected apnea event segments, performing in-depth analysis only on suspected segments, significantly reducing the amount of invalid data processed and significantly lowering the overall computational load of the system, laying the foundation for real-time detection. After the feature extraction module 13 extracts time-frequency features and time-domain sequence features from each of the suspected event signal fragments, the classification decision module 14 enhances key features and extracts multi-scale time-series patterns through the time-frequency analysis branch, and captures the bidirectional time-series dependency and key time point information of the time-domain sequence through the sequence feature branch. Then, through feature fusion, the two types of core features are deeply integrated, which completely solves the problems of insufficient feature extraction and insufficient adaptation to the non-stationary characteristics of BCG signals in traditional algorithms, and greatly improves the accuracy of respiratory event recognition. Based on the fused features, accurate respiratory event classification results are output, further ensuring the reliability of diagnosis. Through the synergistic effect of each module, the entire system inherits the contactless advantages of BCG technology and is adapted to home scenarios. At the same time, it ensures high detection accuracy with its refined feature extraction and fusion mechanism, and greatly reduces computational complexity through suspected fragment focusing design. It successfully achieves the dual goals of high accuracy and low computing power, and can be flexibly deployed on embedded devices to achieve real-time operation.

[0033] In some embodiments, the preprocessing module 11 includes a 4th-order Butterworth bandpass filter and a quality evaluation unit.

[0034] A fourth-order Butterworth bandpass filter is used to filter the raw cardiac impaction signal to obtain a filtered signal. The cutoff frequencies of the fourth-order Butterworth bandpass filter are 0.5Hz and 20Hz. It can remove high-frequency noise and low-frequency baseline drift from the raw cardiac impaction signal.

[0035] The quality assessment unit is used to calculate the signal quality index of each time segment in the filtered signal based on a preset assessment algorithm. Time segments with a signal quality index lower than a preset quality threshold are marked as low-quality segments, and each low-quality segment is removed to obtain the optimized cardiac impaction signal.

[0036] Taking a 30-second time segment as an example, the formula for calculating the Signal Quality Index (SQI) for each 30-second segment can be: SQI=0.5×SNR+0.3×Corr+0.2×Stability Here, SNR is the signal-to-noise ratio, Corr is the correlation between adjacent segments, and Stability is the signal stability. Segments with an SQI below a preset quality threshold (e.g., 0.6) are marked as low quality and given a lower weight during analysis, thus avoiding impact on subsequent processing.

[0037] Understandably, the quality assessment process involves systematically quantifying and analyzing each time segment of the filtered signal, divided into preset time lengths, to accurately calculate the signal quality index for each segment. The signal quality index for each time segment is then compared to a preset quality threshold. Time segments with indices below this threshold are marked as low-quality segments. These segments often exhibit waveform distortion, blurred features, or missing information; directly including them in subsequent analysis would severely interfere with the accuracy of change point detection and the effectiveness of feature extraction. Finally, by precisely removing all marked low-quality segments and systematically stitching together the remaining high-quality signal segments, an optimized cardiac impaction signal with clear waveform features and complete effective physiological information is obtained, thereby further improving signal quality.

[0038] In some embodiments, the change point detection module 12 includes: a segmentation unit, a detection unit, and a segment extraction unit.

[0039] The segmentation unit is used to divide the optimized cardiac impaction signal into multiple non-overlapping time-series signal segments of a predetermined time period, and calculates the mean, variance, and sample entropy of each time-series signal segment as variation features. For example, the optimized cardiac impaction signal can be segmented into 5-second non-overlapping segments.

[0040] It is understandable that dividing a continuous time-series signal into multiple independent and non-overlapping time-series signal segments ensures the independence of each signal segment, avoids data redundancy and redundant feature calculations caused by segment overlap, and allows each segment to accurately correspond to a specific time interval, facilitating subsequent analysis of signal characteristics at different times. After signal segmentation, the segmentation unit calculates three core quantitative indicators as change features for each time-series signal segment. The mean reflects the overall amplitude level of the internal impulse map signal in the corresponding time period, the variance measures the degree of fluctuation of the signal amplitude within that time period, and the sample entropy quantifies the temporal complexity and irregularity of the signal. These three features comprehensively characterize the basic characteristics of the signal at each time period from the three dimensions of signal strength, stability, and complexity, providing accurate and reliable quantitative analysis basis for the subsequent change point detection module 12 to identify signal abrupt changes and locate suspected sleep apnea events.

[0041] The detection unit is used to detect abrupt changes in changing features using the Bayesian Online Change Point Detection (BOCPD) algorithm. The BOCPD algorithm uses Bayesian posterior probability to evaluate the distribution difference between the current feature and historical features. When the difference exceeds a threshold, the corresponding position is determined to be a change point, thereby achieving rapid response and detection of changes.

[0042] The segment extraction unit is used to extract signal segments at preset time points before and after the mutation point to form suspected event signal segments. For example, 15 seconds before and after the mutation point can be extracted to form a 30-second suspected event segment. After extracting the suspected event signal segments, only these suspected segments can be analyzed in detail, while non-suspected segments are directly marked as normal breathing. This strategy can reduce computation by approximately 60-70%, enabling the algorithm to run in real-time on resource-constrained embedded devices.

[0043] In some embodiments, the feature extraction module 13 includes: a time-frequency feature extraction unit and a sequence feature extraction unit.

[0044] The time-frequency feature extraction unit is used to convert the optimized cardiac impaction signal into time-frequency features through continuous wavelet transform. The wavelet transform formula for converting the optimized cardiac impaction signal into time-frequency features is as follows:

[0045] in, To optimize the cardiac impaction signal, The wavelet is the mother wavelet. 'a' is the scaling parameter, controlling the wavelet's width / frequency, and 'b' is the translation parameter, controlling the wavelet's temporal position. Since the Morlet wavelet approaches the lower bound of the Heisenberg uncertainty principle, achieving optimal resolution in both the time and frequency domains, it is suitable for non-stationary BCG signals; therefore, the Morlet wavelet is used as the mother wavelet. The scale can be linearly distributed between 8 and 128, corresponding to a frequency range of 0.5-20 Hz, generating a 128×128 pixel grayscale time-frequency feature map.

[0046] The sequence feature extraction unit is used to extract multi-scale entropy features and waveform features from the optimized cardiac impaction signal, and arrange the extracted features in chronological order to form time-domain sequence features.

[0047] Multidimensional features can be extracted from optimized cardiac impaction signals to form a sequence (one feature vector per second) over a predetermined time period (e.g., 30 seconds). The time-domain sequence features can include multi-scale entropy features and waveform features, and further, signal quality indicators can also be included. For example, the time-domain sequence features can be 11-dimensional features, including multi-scale entropy features (6-dimensional), waveform features (4-dimensional), and signal quality indicators (1-dimensional).

[0048] The multi-scale entropy feature includes the sample entropy at scales 1, 2, and 3, as well as the fuzzy entropy at scales 1, 2, and 3. The calculation method for the multi-scale entropy feature can be found in the following description.

[0049] For length of time series Given parameters (Template length) and (Tolerance threshold), the sample entropy calculation steps include the following steps 1 to 5.

[0050] Step 1, construct an m-dimensional vector, the corresponding formula is as follows: Step 2, define the distance between vectors, the corresponding formula is as follows: Where i and j are indices of the m-dimensional vector, and their values ​​range from... .

[0051] Step 3: Count the number of similar vectors, using the following formula: in, This means that for each group (i, j) ( If satisfied Increment the result by 1. i and j are indices of the m-dimensional vector, with values ​​ranging from... .

[0052] Step 4, calculate the average similarity: Step 5: Increase the dimension to m+1, and repeat steps 1-4 to obtain the average similarity over the m+1 dimension. ,in: Step 6, the sample entropy is defined as: in, It is the natural logarithm. It is a time series. For template length, This is the tolerance threshold.

[0053] For multiscale sample entropy (MSE), a coarse-graining process (scale factor s) is first performed. The original sequence is divided into non-overlapping windows and averaged. The coarse-grained sequence is as follows:

[0054] in, This indicates rounding down to the nearest integer.

[0055] The sample entropy at scale s is defined as: The scale is defined as follows: Scale 1 is the sample entropy of the original sequence. Scale 2 is the coarse-grained sequence averaged over every 2 points; Scale 3 is the coarse-grained sequence averaged over every 3 points.

[0056] FuzzyEntropy improves upon binary similarity judgment using sample entropy by providing a smoother similarity measure using a fuzzy membership function.

[0057] For length of time series Given parameters (Template length) and (Tolerance threshold), the fuzzy entropy calculation steps include the following steps 10 to 60.

[0058] Step 10, construct an m-dimensional vector and zero-mean: for The mean of the sequence.

[0059] Step 20, define the distance between vectors: Step 30: Calculate similarity using the fuzzy membership function: in, It is the gradient parameter (usually taken as 2).

[0060] Step 40, calculate the average similarity: Step 50: Increase the dimension to m+1, and repeat steps 10-40 to obtain the average similarity over the m+1 dimension. .

[0061] Step 60, the fuzzy entropy is defined as: Multiscale fuzzy entropy (MFE) is defined similarly to multiscale sample entropy, but it is first coarsened. The coarsening sequence is as follows: The fuzzy entropy of scale s is: The parameter settings can be as follows: Waveform characteristics include respiratory amplitude index, root mean square value, amplitude drop ratio, and coefficient of variation of respiratory interval. The formula for calculating the respiratory amplitude index is: RA = max(S) - min(S) (within the sliding window, S is the respiratory signal amplitude). The respiratory amplitude index directly reflects the respiratory depth; RA decreases during hypoventilation and approaches zero during pause.

[0062] The formula for calculating the root mean square (RMS) value is: RMS = sqrt(mean(S^2)). RMS reflects the average power of the signal and is sensitive to sustained amplitude reduction.

[0063] The formula for calculating the amplitude drop ratio is: ADR = RA_current / RA_baseline (the baseline is the average normal breathing rate in the previous few minutes). ADR is a key diagnostic feature; ADR < 0.7 indicates hypoventilation; ADR < 0.1 indicates pause.

[0064] The formula for calculating the coefficient of variation of respiratory intervals is: CV = std(Intervals) / mean(Intervals). The coefficient of variation of respiratory intervals is used to quantify the stability of respiratory rhythm. The CV may increase before and after central apnea. Obstructive apnea may be accompanied by periodic changes.

[0065] The formula for calculating signal quality metrics is: Signal-to-noise ratio (SNR) = 10 * log 10 (Ps / Pn), where Ps is the power of the signal and Pn is the power of the noise.

[0066] In some embodiments, the classification decision module 14 includes a two-branch neural network, which includes a time-frequency analysis branch, a sequence feature branch, a feature fusion unit, and a classifier. In one example, the framework structure of the two-branch neural network can be as follows: Figure 3 As shown.

[0067] In the application, the time-frequency analysis branch is used to perform feature enhancement and multi-scale temporal pattern extraction on time-frequency features using a lightweight CNN network, outputting time-frequency morphological features. Lightweight CNNs can significantly reduce the number of parameters and computational load through designs such as depthwise separable convolutions, small-sized convolutional kernels (e.g., 1×3, 3×3), and simplified network layers, making them suitable for wearable and portable detection devices with limited computing power, balancing computational efficiency and feature extraction effectiveness. Specifically, the lightweight CNN first performs feature enhancement processing on the input time-frequency feature map. Through the local perception and weight sharing characteristics of the convolutional kernel, it accurately captures key information highly correlated with apnea-hypopnea events in the time-frequency features, and amplifies the signal strength of these effective features through convolutional operations, while suppressing the influence of irrelevant features such as environmental noise and device interference, thus improving the recognizability of core features. Subsequently, multi-scale temporal pattern extraction is achieved by setting convolutional layers with multi-scale receptive fields, that is, mining the hidden temporal patterns in the time-frequency features at different time scales. Ultimately, the lightweight CNN integrates the time-frequency features extracted through feature enhancement and multi-scale temporal patterns, and outputs time-frequency morphological features that can accurately characterize the waveform shape, frequency composition, and spatiotemporal distribution of the cardiac impact signal in the time-frequency domain.

[0068] The sequence feature branch is used to capture bidirectional temporal dependencies from temporal sequence features through bidirectional gated recurrent units, capture key time point information from temporal sequence features through a temporal attention mechanism, and output complexity features.

[0069] The bidirectional gated recurrent unit (BRN) overcomes the limitation of traditional unidirectional recurrent neural networks, which can only capture forward temporal dependencies. Through a parallel architecture of forward and backward networks, it can capture the forward evolution of signals from the past to the present along the timeline, and also uncover the reverse correlation features from the future to the present along the timeline. This allows for a complete capture of the hidden bidirectional temporal dependencies within the time-domain sequence features, thus comprehensively characterizing the temporal evolution logic of the signal. Simultaneously, the temporal attention mechanism plays a feature-focusing role. By calculating the contribution weights of features at different time points in the time-domain sequence to the determination of apnea events, it automatically filters and amplifies key time-point information crucial for event identification, effectively filtering out redundant information in the time-domain sequence. Finally, this branch organically integrates the global bidirectional temporal dependency features captured by the BRN with the local key time-point features enhanced by the temporal attention mechanism, outputting complexity features that accurately quantify the complexity and evolutionary patterns of the time-domain sequence signal.

[0070] The feature fusion unit integrates time-frequency morphological features with complexity features to obtain fused features. Specifically, the feature fusion unit can combine time-frequency morphological features and complexity features at a high-level representation level, allowing subsequent layers to learn the complex interactions between them. For example, when it learns that "when the time-frequency map shows the disappearance of low-frequency energy and a simultaneous sharp decrease in sample entropy," it is considered strong evidence of central apnea. The feature fusion unit can choose from splicing fusion, weighted fusion, or attention-guided fusion strategies according to actual application needs, ultimately outputting fused features that simultaneously encompass the time-frequency morphological characteristics and temporal sequence complexity of the cardiac impaction signal. This significantly improves the feature's discriminative power and representational ability, providing comprehensive and efficient feature support for subsequent classifiers to accurately determine the type of respiratory event.

[0071] In some embodiments, the time-frequency analysis branch includes: a first input layer, a standard convolutional layer, a first depthwise separable convolutional block, a second depthwise separable convolutional block, a multi-scale residual block, and a global average pooling layer. In one example, the framework structure of the time-frequency analysis branch can be as follows: Figure 4 As shown.

[0072] The first input layer is used to receive time-frequency features and transmit them to the standard convolutional layer. For example, the first input layer can receive a 128×128×1 time-frequency map, each of which is generated by continuous wavelet transform of the optimized BCG signal and contains frequency information from 0.5 to 20 Hz.

[0073] Standard convolutional layers are used for rapid downsampling of time-frequency features and primary feature detection, outputting primary features. A standard convolutional layer can achieve rapid downsampling and primary feature detection simultaneously by configuring 16 3×3 convolutional filters with a stride of 2. The stride of 2 efficiently compresses the time-frequency map size from 128×128 to 64×64, significantly reducing the computational cost of subsequent network layers and laying the foundation for lightweight deployment. The 3×3 filters accurately capture basic visual features related to respiration and heartbeat in the time-frequency map, such as the slanted edges of energy bands (corresponding to frequency changes) and local energy patches (corresponding to transient physiological events), among other key primary patterns. Simultaneously, this layer uses the ReLU activation function to introduce a non-linear transformation, enhancing the network's ability to express features. Subsequent batch normalization processing stabilizes the data distribution and accelerates model training, ultimately outputting primary features whose dimensions match those of subsequent depth-separable convolutional blocks.

[0074] The first depthwise separable convolutional block is used to extract channel-level features from primary features and calibrate the feature response using the SE attention mechanism to output the first depth feature. This first depthwise separable convolutional block can employ 32 3×3 depthwise separable convolutional kernels, with a stride of 1 to ensure sufficient preservation of feature information. It incorporates the ReLU activation function to introduce non-linear expressiveness, followed by batch normalization to stabilize feature distribution and accelerate model training. Simultaneously, it integrates the SE attention module, achieving approximate expressiveness at a much lower parameter cost than standard convolution, efficiently extracting channel-level features from primary features and calibrating the feature response, ultimately outputting the first depth feature. This module achieves channel-level spatial filtering through depthwise convolution, allowing each filter to independently scan a single feature channel. For example, it specifically searches for continuous energy ridges corresponding to normal breathing within the "0.2-0.3Hz frequency band," accurately capturing local time-frequency patterns within each channel. Then, through pointwise convolution, it linearly combines the input 16 channel features and expands them to 32 channels, creating more abstract higher-order features such as "regular oscillations" and "energy dispersion," achieving effective feature fusion and dimensionality enhancement. The SE attention mechanism, as the core calibration unit, can automatically learn the importance weights of 32 channels. In the signal stability segment, it suppresses channels sensitive to "transient" changes to reduce interference, and in the suspected event segment, it enhances channels sensitive to "low energy" and "rhythm disorder" to highlight key information. This adaptive filtering design significantly improves the signal-to-noise ratio and ensures that the extracted channel-level features are more targeted and effective.

[0075] The second depthwise separable convolutional block is used to perform secondary downsampling on the first depth features to expand the receptive field and capture intermediate-level abstract time-frequency semantics, outputting the second depth features. The second depthwise separable convolutional block is a key structure in the time-frequency analysis branch for deepening feature extraction and balancing accuracy and efficiency. It can employ 64 3×3 depthwise separable convolutional kernels with a stride of 2 to achieve secondary downsampling of the first depth features. It incorporates a ReLU activation function to introduce non-linear expression, followed by batch normalization to stabilize the feature distribution. Simultaneously, it integrates an SE attention module to ensure the effectiveness of the feature response. Its core function is to accurately expand the receptive field while efficiently capturing intermediate-level abstract time-frequency semantics, ultimately outputting the second depth features.

[0076] The secondary downsampling design of this module represents an optimal trade-off between accuracy and efficiency. The step size of 2 reduces the input feature map size from 64×64 to 32×32, ensuring that the 32×32 resolution is sufficient to fully characterize breathing events that typically last for more than 10 seconds, while reducing the subsequent computational load to 1 / 4 of the original, significantly reducing the computational burden on embedded devices. With the implementation of downsampling, the receptive field of a single feature point corresponds to an area of ​​approximately 4×4 (approximately 0.5Hz×2 seconds) in the original time-frequency map, enabling it to overcome the limitations of local features and accurately capture more macroscopic time-frequency patterns such as "widespread low-frequency energy decay exceeding 10 seconds". At the same time, the 64 convolutional channels can encode more complex feature combinations, such as the typical obstructive pause cue of "stable high-frequency heart rate harmonics but lack of energy in the low-frequency breathing band". Combined with the adaptive calibration of channel features by the SE attention module, the response of key semantic features is further enhanced, and redundant interference is suppressed, ensuring that the extracted intermediate-level abstract time-frequency semantics are more recognizable and effective.

[0077] Multi-scale residual blocks are used to simultaneously capture time-frequency patterns at different scales through multi-size convolutional paths. Addressing the multi-scale characteristics of apnea events, which exhibit both instantaneous amplitude abrupt changes (small scale) and pattern changes lasting tens of seconds (large scale), multi-scale residual blocks can achieve simultaneous capture of time-frequency patterns at different scales by designing three parallel convolutional paths containing filter sizes of 1×1, 3×3, and 5×5. The 1×1 convolutional path focuses on the reorganization of inter-channel relationships, is sensitive to changes in local points, and can accurately capture small-scale instantaneous features. The 3×3 convolutional path balances the receptive field size and can effectively capture medium-scale features such as local time-frequency correlations (e.g., instantaneous changes in frequency). The 5×5 convolutional path has a larger receptive field and can identify long-range large-scale patterns such as the gradual weakening of the entire respiratory cycle. At the same time, by splicing and fusing the output features of the three paths, deep integration of information at different scales is achieved. Combined with residual connections to avoid deep network degradation and ensure effective gradient backpropagation, the network can accurately perceive small-scale instantaneous anomalies and comprehensively capture large-scale long-term pattern changes, thus fully covering the time-frequency features of apnea events at different time scales.

[0078] The global average pooling layer is used to aggregate spatial feature information based on the output features of multi-scale residual blocks, outputting time-frequency morphological features. The global average pooling layer receives multi-scale fused features from the multi-scale residual block output and can efficiently aggregate the spatial information from these features, outputting a 64-dimensional time-frequency morphological feature vector. By compressing the two-dimensional time-frequency pattern extracted from each channel into a globally representative intensity value, accurate aggregation of spatial information is achieved. For example, the pooling value for the channel representing "breathing rhythm" remains high during normal breathing segments, while significantly decreasing when apnea occurs. Simultaneously, this layer forces the network to focus on learning global features, avoiding the overfitting problem to the spatial location of the feature map that easily occurs in fully connected layers. This significantly enhances the model's robustness to small shifts in the temporal location of events. Furthermore, by outputting a fixed-length 64-dimensional feature vector, the number of parameters in subsequent network layers is greatly reduced, lowering the computational load.

[0079] In some embodiments, the sequence feature branch includes: a second input layer, a bidirectional GRU layer, an attention layer, and a first fully connected layer. In one example, the framework structure of the sequence feature branch can be as follows: Figure 5 As shown.

[0080] The second input layer is used to receive time-domain sequence features and transmit them to the bidirectional GRU layer. For example, based on the foregoing embodiment, the second input layer can receive a 30×11 feature sequence (30 time steps, 11-dimensional features per time step).

[0081] The bidirectional GRU layer is used for bidirectional processing of temporal sequence features, capturing forward and backward temporal dependencies and outputting multidimensional temporal features. The bidirectional GRU layer is the core structure for temporal modeling in the sequence feature branch. It can be configured with 32 hidden units and processes the input temporal sequence features bidirectionally through the parallel operation of the forward and backward GRUs, accurately capturing forward and backward temporal dependencies and outputting 30×64-dimensional multidimensional temporal features. The forward GRU interprets the feature sequence frame by frame from the "past to future" direction, learning temporal evolution patterns with cause and effect, such as "a gradual decrease in entropy may indicate a weakening of respiratory drive." The backward GRU analyzes the sequence from the reverse perspective of "future to past," helping to accurately determine key event boundaries such as "a sudden drop in amplitude is the starting point of an event." This bidirectional fusion design makes temporal modeling more comprehensive and boundary judgment more accurate. Meanwhile, GRU's built-in update gate and reset gating mechanism enable it to selectively memorize important temporal information and forget redundant interference, effectively modeling the evolution of breathing patterns spanning tens of seconds. This is crucial for identifying scenarios that require long-term dependency modeling, such as periodic breathing. The final output of 30×64 multi-dimensional temporal features not only retains the key information of each time step but also incorporates bidirectional temporal correlation.

[0082] The attention layer is used to calculate the attention weight of each time step in the multidimensional temporal features. It then performs a weighted sum of the attention weights for each time step to determine the key time point information of the multidimensional temporal features and outputs attention-weighted temporal features. After receiving the multidimensional temporal features output from the bidirectional GRU layer, the attention layer calculates the weight coefficient of each time step through the attention mechanism, accurately quantifying the importance of each time step for respiratory event recognition. Subsequently, it performs a weighted sum of the features of all time steps according to their corresponding attention weights, thereby efficiently filtering and focusing on the most discriminative key time point information in the respiratory event, such as the time step with the most drastic decrease in respiratory amplitude, the most significant entropy change, or the most obvious rhythm disturbance, and its surrounding area. Simultaneously, this layer, by visualizing the attention weight distribution, can clearly present the key temporal nodes on which the model's decisions depend, ultimately outputting attention-weighted temporal features with concise dimensions and prominent key information. This design can meet the core requirement of interpretability in medical AI, making diagnostic evidence more convincing.

[0083] The first fully connected layer maps the attention-weighted temporal features to a semantic space aligned with the time-frequency analysis branch, outputting complexity features. As the final feature mapping structure of the sequence feature branch, this first fully connected layer can be configured with 64 neurons and employs the ReLU activation function to introduce non-linear expressive power. This allows the non-linear transformation of the neural network to accurately map the attention-weighted temporal features to a semantic space perfectly aligned with the time-frequency morphological features output by the time-frequency analysis branch, ensuring that both types of features are not only 64-dimensional but also fusionable at the semantic level.

[0084] In some embodiments, the classifier includes a second fully connected layer and an output layer.

[0085] The second fully connected layer is used for advanced nonlinear interactive learning of the fused features and suppresses overfitting through Dropout, outputting classification features that fit the classification. As the core structure of feature deepening interaction in the classifier, the second fully connected layer can be configured with 128 neurons and adopts the ReLU activation function to introduce strong nonlinear expressive power. At the same time, a Dropout rate of, for example, 0.3 is set. Its core function is to receive the 128-dimensional fused features output by the feature fusion unit, and realize advanced, complex, and highly relevant combination interaction between the two types of branch features through the nonlinear transformation of the neural network. This fully explores the joint discriminative value of the two complementary features. At the same time, the Dropout layer randomly discards 30% of the neurons during model training, effectively avoiding the network's over-reliance on certain specific features, reducing the risk of overfitting, and significantly improving the model's generalization ability to respiratory signals in different patients and scenarios.

[0086] The output layer is used to map the classification features to the probability distribution of each classification result using the Softmax activation function, and the classification result of the respiratory event is determined based on the probability of each classification result.

[0087] The output layer is the final stage of the 10-class classification decision in the entire apnea-hypopnea detection system. It can be configured with four neurons corresponding one-to-one with the four types of respiratory events (normal breathing, obstructive apnea, hypopnea, and central apnea). The Softmax activation function maps the high-dimensional classification features to a probability distribution within the 0-1 range, with a sum of 1. The probability value output by each neuron corresponds to the likelihood of the detected segment belonging to its respective category. Simultaneously, based on the numerical comparison of the probabilities of each category, this layer selects the category with the highest probability as the final respiratory event classification result. This achieves accurate classification of the detected events and visually presents the confidence level of the model's decision through the probability distribution, providing a reference for clinical diagnosis. The characteristics of the Softmax activation function ensure the rationality and interpretability of the probability distribution. Combined with the advantages of pre-sequence bi-branch feature fusion and advanced nonlinear interactive learning, it ultimately achieves efficient and accurate differentiation of the four types of respiratory events, providing reliable classification conclusions for the diagnosis of apnea-hypopnea syndrome.

[0088] In some embodiments, the apnea-hypopnea detection system 10 further includes a diagnostic module.

[0089] The diagnostic module is used to obtain the user's total sleep time, calculate the AHI index based on the respiratory event classification results obtained within the sleep cycle and the total sleep time, and determine the respiratory abnormality grading results based on the AHI index.

[0090] It is understandable that the diagnostic module obtains the user's total sleep time (in hours) for the entire sleep cycle through system monitoring, and simultaneously retrieves the respiratory event classification results within the sleep cycle, which have been categorized into 14 categories by the classification decision module and verified by clinical rules. The judgment rules can be as follows: Apnea requires a neural network output probability > 0.5 and a BCG-derived respiratory amplitude decrease ≥ 90% with a duration ≥ 10 seconds; hypoventilation requires a neural network output probability > 0.5 and a BCG-derived respiratory amplitude decrease ≥ 30% with a duration ≥ 10 seconds, accompanied by a decrease in blood oxygen saturation ≥ 3% (e.g., when connected to a pulse oximeter) or BCG micro-awakening characteristics. By confirming the number of valid apnea and hypoventilation events, the highest probability value for each category output by the neural network is selected to determine the specific type of each valid event (obstructive apnea / central apnea / hypopnea).

[0091] Subsequently, following the relevant standard method for calculating the AHI index, the sum of the total number of apnea events and the total number of hypopnea events is divided by the total sleep time to obtain the AHI (apnea-hypopnea index). The corresponding formula is: AHI = (Total number of apnea events + Total number of hypopnea events) / Total sleep time (hours). After calculating the AHI, the severity of respiratory abnormalities is graded based on the AHI value, where AHI < 5 is normal, 5 ≤ AHI < 15 is mild, 15 ≤ AHI < 30 is moderate, and AHI ≥ 30 is severe. The final output is an accurate respiratory abnormality grading result, providing a standardized and referable core basis for clinical diagnosis of apnea-hypopnea syndrome and assessment of disease severity.

[0092] In some embodiments, such as Figure 6 As shown, this application also provides a method for detecting apnea-hypopnea, including the following steps S601 to S603.

[0093] S601: Obtain the training sample set.

[0094] S602: Based on the training sample set and the weighted cross-entropy loss function, the apnea-hypopnea detection system of any of the above schemes is trained using a domain adversarial training strategy to obtain the target apnea-hypopnea detection model.

[0095] The loss function uses weighted cross-entropy loss. To solve the class imbalance problem, the corresponding formula is as follows: in, For category The weight, This is the true label of the sample, encoded using one-hot encoding. If the sample belongs to the... Class, then Otherwise, it is 0. The model predicts that this sample belongs to the first... The probability of the class (output by the Softmax layer).

[0096] In applications, a domain classifier can be added after the feature fusion layer, and a gradient inversion layer can be used to enable the feature extractor to learn domain-invariant features, achieving domain adversarial training. Domain adversarial training introduces a domain discriminator during feature extraction and constructs an adversarial game between the feature extractor and the discriminator, forcing the feature extractor to learn "domain-invariant features" that can confuse the domain discriminator, thereby improving the model's generalization ability on cross-domain tasks. The optimizer during training can use the Adam optimizer, with an initial learning rate of 0.001, decaying by 0.5 every 20 epochs. An early stopping mechanism can also be set during training, stopping training when the validation set loss does not improve for 10 consecutive epochs.

[0097] S603: Input the original cardiac impaction signal into the target apnea-hypopnea detection model to obtain respiratory event classification results.

[0098] The aforementioned apnea-hypopnea detection method effectively addresses the potential class imbalance between respiratory event samples and normal respiratory samples through a weighted cross-entropy loss function. By assigning higher weights to minority class event samples, it avoids insufficient identification and high false negative rates of minority class events due to sample proportion differences during model training, significantly improving the accuracy of respiratory abnormality event identification. The domain adversarial training strategy addresses the domain distribution differences in BCG signals across different users, monitoring environments, and device states. Through adversarial training, the model learns domain-independent universal features, reducing the impact of environmental interference and individual differences on detection results, significantly improving the model's generalization ability and maintaining stable detection performance for different users in diverse scenarios such as home environments. The trained target apnea-hypopnea detection model inherits the core advantages of the original system—non-contact operation, low computational complexity, and adaptability to embedded devices—and further enhances the effectiveness of feature extraction and the reliability of classification decisions through targeted training. It can quickly output accurate respiratory event classification results with only the original BCG signal as input.

[0099] In other embodiments, such as Figure 7 As shown, this application also provides a method for detecting apnea-hypopnea, including the following steps S701 to S704.

[0100] S701: Acquire the raw cardiac impact map signal obtained from detecting the user's breathing event, and preprocess the raw cardiac impact map signal to obtain an optimized cardiac impact map signal.

[0101] S702: Use a preset change point detection algorithm to locate the suspected event signal segment corresponding to the suspected apnea event from the optimized cardiac impact map signal.

[0102] S703: Extract time-frequency features and time-domain sequence features from each suspected event signal fragment.

[0103] S704: The time-frequency analysis branch enhances key features and extracts multi-scale time-series patterns from the time-frequency features, outputting time-frequency morphological features. The sequence feature branch captures bidirectional temporal dependence and key time point information from the time-domain sequence features, outputting complexity features. The time-frequency morphological features and complexity features are fused to obtain fused features. The classification result of respiratory events is determined by a classifier based on the fused features.

[0104] It should be noted that the apnea-hypopnea detection method and the apnea-hypopnea detection system provided in this application are based on the same inventive concept. Therefore, the specific implementation of this embodiment can refer to the implementation of the aforementioned apnea-hypopnea detection system, and the repeated parts will not be described again.

[0105] In some embodiments, an electronic device provided in this application includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the above-described apnea-hypopnea detection method.

[0106] Specifically, the processor may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include onboard memory for caching purposes. The processor may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0107] Memory can be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory can include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and also random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0108] This application also provides a non-transitory computer storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described apnea-hypopnea detection method. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method as described in the embodiments of this application.

[0109] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0110] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0111] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application. Therefore, the scope of this application should not be limited to the above embodiments. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A sleep apnea-hypopnea detection system, characterized in that, The sleep apnea-hypopnea detection system includes: a preprocessing module, a change point detection module, a feature extraction module, and a classification decision module; The preprocessing module is used to acquire the raw cardiac impaction signal obtained from detecting the user's respiratory events, and to preprocess the raw cardiac impaction signal to obtain an optimized cardiac impaction signal; The change point detection module is used to locate the suspected event signal segment corresponding to the suspected apnea event from the optimized cardiac impact map signal using a preset change point detection algorithm; The feature extraction module is used to extract time-frequency features and time-domain sequence features from each of the suspected event signal fragments; The classification decision module enhances key features and extracts multi-scale temporal patterns from the time-frequency features through the time-frequency analysis branch, and outputs time-frequency morphological features; it captures bidirectional temporal dependence and key time point information from the time-domain sequence features through the sequence feature branch, and outputs complexity features; it fuses the time-frequency morphological features and the complexity features to obtain fused features; and it determines the classification result of respiratory events through a classifier based on the fused features.

2. The sleep apnea-hypopnea detection system as described in claim 1, characterized in that, The preprocessing module includes: a fourth-order Butterworth bandpass filter and a quality assessment unit; The fourth-order Butterworth bandpass filter is used to filter the original cardiac impaction signal to obtain a filtered signal; The quality assessment unit is used to calculate the signal quality index of each time segment in the filtered signal based on a preset assessment algorithm, mark the time segments with a signal quality index lower than a preset quality threshold as low-quality segments, and remove each low-quality segment to obtain the optimized cardiac impaction signal.

3. The sleep apnea-hypopnea detection system as described in claim 1, characterized in that, The change point detection module includes: a segmentation unit, a detection unit, and a fragment extraction unit; The segmentation unit is used to segment the optimized cardiac impaction signal into multiple non-overlapping time-series signal segments of a predetermined time period, and calculate the mean, variance, and sample entropy of each time-series signal segment as variation features. The detection unit is used to detect abrupt changes in the changing features using a Bayesian online change point detection algorithm; The segment extraction unit is used to extract signal segments at preset times before and after the mutation point to form the suspected event signal segment.

4. The sleep apnea-hypopnea detection system as described in claim 1, characterized in that, The classification decision module includes: a dual-branch neural network, which includes a time-frequency analysis branch, the sequence feature branch, a feature fusion unit, and the classifier; The time-frequency analysis branch is used to perform feature enhancement and multi-scale temporal pattern extraction on the time-frequency features using a lightweight CNN network, and outputs the time-frequency morphological features; The sequence feature branch is used to capture bidirectional temporal dependencies from the temporal sequence features through a bidirectional gated recurrent unit, capture key time point information from the temporal sequence features through a temporal attention mechanism, and output the complexity feature; The feature fusion unit is used to fuse the time-frequency morphological features with the complexity features to obtain fused features; The classifier is used to determine the classification result of respiratory events based on the fusion features.

5. The sleep apnea-hypopnea detection system as described in claim 4, characterized in that, The time-frequency analysis branch includes: a first input layer, a standard convolutional layer, a first depthwise separable convolutional block, a second depthwise separable convolutional block, a multi-scale residual block, and a global average pooling layer; The first input layer is used to receive the time-frequency features and transmit them to the standard convolutional layer; The standard convolutional layer is used to perform fast downsampling and primary feature detection on the time-frequency features, and output primary features; The first depthwise separable convolutional block is used to extract channel-level features from the primary features and, in conjunction with the SE attention mechanism, calibrate the feature response to output the first depthwise features; The second depthwise separable convolutional block is used to perform secondary downsampling on the first depth features to expand the receptive field and capture intermediate abstract time-frequency semantics, outputting the second depth features; The multi-scale residual block is used to simultaneously capture time-frequency patterns at different scales through multi-size convolutional paths; The global average pooling layer is used to aggregate spatial feature information based on the output features of the multi-scale residual block and output the time-frequency morphological features.

6. The apnea-hypopnea detection system as described in claim 4, characterized in that, The sequence feature branch includes: a second input layer, a bidirectional GRU layer, an attention layer, and a first fully connected layer; The second input layer is used to receive the time-domain sequence features and transmit them to the bidirectional GRU layer; The bidirectional GRU layer is used to process the time-domain sequence features bidirectionally, capture forward and backward temporal dependencies, and output multi-dimensional temporal features. The attention layer is used to calculate the attention weight of each time step in the multidimensional temporal feature, perform a weighted summation of the attention weights of each time step, determine the key time point information of the multidimensional temporal feature, and output the attention-weighted temporal feature. The first fully connected layer is used to map the attention-weighted temporal features to a semantic space aligned with the time-frequency analysis branch, and output the complexity features.

7. The apnea-hypopnea detection system as described in claim 4, characterized in that, The classifier includes: a second fully connected layer and an output layer; The second fully connected layer is used to perform advanced nonlinear interactive learning on the fused features and suppress overfitting through Dropout to output classification features that fit the classification. The output layer is used to map the classification features to a probability distribution of each classification result using a Softmax activation function, and to determine the classification result of the respiratory event based on the probability of each classification result.

8. The sleep apnea-hypopnea detection system as described in claim 1, characterized in that, The sleep apnea-hypopnea detection system also includes: a diagnostic module; The diagnostic module is used to obtain the user's total sleep time, calculate the AHI index based on the respiratory event classification results obtained within the sleep cycle and the total sleep time, and determine the respiratory abnormality grading result based on the AHI index.

9. A method for detecting apnea-hypopnea, characterized in that, include: Obtain the training sample set; Based on the training sample set and the weighted cross-entropy loss function, the apnea-hypopnea detection system as described in any one of claims 1 to 8 is trained using a domain adversarial training strategy to obtain a target apnea-hypopnea detection model. The original cardiac impaction signal is input into the target apnea-hypopnea detection model to obtain respiratory event classification results.

10. A method for detecting apnea-hypopnea, characterized in that, include: The raw cardiac impactogram signal obtained from detecting user breathing events is acquired, and the raw cardiac impactogram signal is preprocessed to obtain an optimized cardiac impactogram signal; A preset change point detection algorithm is used to locate the suspected event signal segment corresponding to the suspected apnea event from the optimized cardiac impact map signal; Time-frequency features and time-domain sequence features are extracted from each of the suspected event signal fragments; The time-frequency features are enhanced and multi-scale time-series patterns are extracted through the time-frequency analysis branch, and the time-frequency morphological features are output. The bidirectional time-series dependence and key time point information are captured from the time-domain sequence features through the sequence feature branch, and the complexity features are output. The time-frequency morphological features and the complexity features are fused to obtain the fused features. Based on the fusion features, the classification result of the respiratory event is determined by a classifier.

Images