Sleep apnea detection method, apparatus, embedded device, and storage medium

Abstract

The application relates to a sleep apnea detection method and device, an embedded device and a storage medium. The method comprises the following steps: acquiring PPG signal data of a user during sleep, preprocessing the PPG signal data to obtain a plurality of PPG signal segments, extracting a one-dimensional time sequence feature vector reflecting the change of heart rate and blood flow for each PPG signal segment, converting the PPG signal segment into a two-dimensional spectrum image, extracting a two-dimensional image feature vector reflecting the spectrum structure characteristics based on the two-dimensional spectrum image, fusing the one-dimensional time sequence feature vector and the two-dimensional image feature vector to obtain a fusion feature vector, performing standardization and dimension reduction processing on the fusion feature vector to obtain a target feature vector, inputting the target feature vector into a trained classification model for classification to obtain a classification result of whether the PPG signal segment has an apnea event. The method can improve the accuracy of sleep apnea detection.

Description

Technical Field

[0001] This application relates to the field of sleep monitoring technology, and in particular to a method, apparatus, embedded device, computer-readable storage medium, and computer program product for detecting sleep apnea. Background Technology

[0002] Sleep apnea is a common sleep disorder closely associated with health problems such as cardiovascular disease, hypertension, and diabetes. Traditional methods for detecting sleep apnea, such as polysomnography (PSG), while highly accurate, are limited in their use in homes and daily life due to the high cost of equipment, complex operation, and the need for hospital or sleep laboratory environments.

[0003] In recent years, photoplethysmography (PPG) has been widely used in the field of health monitoring. PPG signals can reflect changes in blood volume and indirectly reflect physiological parameters such as heart rate and respiratory rate. Therefore, the detection of sleep apnea based on PPG signals has been increasingly widely used.

[0004] However, currently, traditional sleep apnea detection methods based on PPG signals can only capture relatively macroscopic physiological changes and are difficult to accurately identify short-term, minute apnea events. Summary of the Invention

[0005] Therefore, it is necessary to provide a more accurate method, device, embedded device, computer-readable storage medium, and computer program product for detecting sleep apnea, addressing the aforementioned technical problems.

[0006] Firstly, this application provides a method for detecting sleep apnea, including:

[0007] Acquire PPG signal data of the user during sleep, preprocess the PPG signal data to obtain multiple PPG signal segments;

[0008] For each PPG signal segment, a one-dimensional time-series feature vector is extracted from the PPG signal segment. The one-dimensional time-series feature vector is used to reflect the changes in heart rate and blood flow.

[0009] The PPG signal segment is converted into a two-dimensional spectrum image, and a two-dimensional image feature vector is extracted based on the two-dimensional spectrum image. The two-dimensional image feature vector is used to reflect the spectrum structure characteristics.

[0010] The one-dimensional temporal feature vector is fused with the two-dimensional image feature vector to obtain a fused feature vector;

[0011] The fused feature vector is standardized and dimensionality reduced to obtain the target feature vector;

[0012] The target feature vector is input into a trained classification model for classification to obtain a classification result indicating whether the PPG signal fragment contains an apnea event. The classification model is trained based on historical PPG signal fragments carrying apnea classification labels.

[0013] In one embodiment, the PPG signal segment is converted into a two-dimensional spectral image, and a two-dimensional image feature vector is extracted based on the two-dimensional spectral image, including:

[0014] The PPG signal segment is subjected to time-frequency transformation to obtain a two-dimensional spectrum image;

[0015] The two-dimensional spectral image is subjected to frequency domain cropping and geometric normalization to obtain the target two-dimensional image;

[0016] Based on the target two-dimensional image, extract the two-dimensional image feature vector.

[0017] In one embodiment, extracting the two-dimensional image feature vector from the target two-dimensional image includes at least one of the following methods:

[0018] The ratio of frequency energy to total energy in the target two-dimensional image within a preset low-frequency range is determined to obtain the low-frequency energy proportion.

[0019] For each pixel in the target two-dimensional image, the pixel is compared with its eight neighboring pixels, and the local binary mode of the target two-dimensional image is determined based on the comparison results.

[0020] Based on the target two-dimensional image, the two-dimensional spectrum image is equivalent to a probability distribution, and the two-dimensional spectrum image is compared with a preset respiratory spectrum template to determine the relative entropy between the two-dimensional spectrum image and the preset respiratory spectrum template.

[0021] The two-dimensional image feature vector includes at least one of the following: low-frequency energy proportion, local binary pattern, and relative entropy.

[0022] In one embodiment, the preprocessing of the PPG signal data to obtain multiple PPG signal segments includes:

[0023] The PPG signal data is filtered and then segmented into frames according to a fixed time window and a sliding step size to obtain multiple PPG signal segments with partial overlap between frames.

[0024] In one embodiment, extracting a one-dimensional time-series feature vector from the PPG signal segment includes:

[0025] For each PPG signal segment, the first-order difference of the PPG signal segment is determined, and the pulse peak is located based on the first-order difference.

[0026] Based on the position and amplitude of each pulse peak, determine the pulse interval sequence and pulse amplitude sequence;

[0027] Based on the pulse interval sequence and the pulse amplitude sequence, determine the root mean square of the mean pulse period, the standard deviation of the pulse interval, the mean pulse amplitude, and the difference between adjacent pulse amplitudes;

[0028] The sample entropy of each PPG signal segment under different preset scale factors is determined to obtain the multi-scale entropy;

[0029] The one-dimensional temporal feature vector includes at least one of the following: mean pulse period, standard deviation of pulse interval, mean pulse amplitude, root mean square of the difference between adjacent pulse amplitudes, and multi-scale entropy.

[0030] In one embodiment, the step of inputting the target feature vector into a trained classification model for classification to obtain a classification result indicating whether the PPG signal fragment contains an apnea event includes:

[0031] The target feature vector is input into a trained support vector machine classifier for classification to obtain a classification label for whether the PPG signal fragment contains a sleep apnea event;

[0032] The support vector machine classifier uses a radial basis function kernel.

[0033] Secondly, this application also provides a sleep apnea detection device, comprising:

[0034] The data acquisition module is used to acquire PPG signal data of the user during sleep, and preprocess the PPG signal data to obtain multiple PPG signal segments.

[0035] A one-dimensional feature extraction module is used to extract a one-dimensional temporal feature vector from each PPG signal segment. The one-dimensional temporal feature vector is used to reflect the changes in heart rate and blood flow.

[0036] A two-dimensional feature extraction module is used to convert the PPG signal segment into a two-dimensional spectrum image, and extract a two-dimensional image feature vector based on the two-dimensional spectrum image. The two-dimensional image feature vector is used to reflect the spectrum structure characteristics.

[0037] The feature fusion module is used to fuse the one-dimensional temporal feature vector with the two-dimensional image feature vector to obtain a fused feature vector;

[0038] The dimension reduction processing module is used to standardize and reduce the dimension of the fused feature vector to obtain the target feature vector.

[0039] The apnea detection module is used to input the target feature vector into a trained classification model for classification, and obtain a classification result of whether the PPG signal segment contains an apnea event. The classification model is trained based on historical PPG signal segments carrying apnea classification labels.

[0040] Thirdly, this application also provides an embedded device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in any of the above embodiments of the sleep apnea detection method.

[0041] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in any of the above embodiments of the sleep apnea detection method.

[0042] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps in any of the above-described embodiments of the sleep apnea detection method.

[0043] The aforementioned sleep apnea detection methods, devices, embedded devices, computer-readable storage media, and computer program products only require the acquisition of PPG signals. By extracting a one-dimensional temporal feature vector reflecting changes in heart rate and blood flow, they can effectively capture physiological changes such as increased heart rate variability and aggravated blood flow fluctuations caused by apnea. By converting PPG signal segments into two-dimensional spectral images and extracting two-dimensional image feature vectors reflecting spectral structure characteristics, and then fusing one-dimensional temporal features and two-dimensional spectral image features, they can fully leverage the advantages of temporal dynamic sensitivity and frequency domain pattern recognition to comprehensively capture minute changes in apnea events, thereby significantly improving the accuracy and robustness of sleep apnea detection. Subsequently, by standardizing and reducing the dimensionality of the fused feature vector, a target feature vector is obtained, simplifying the algorithm complexity. The target feature vector is then input into a classification model trained on historical PPG signal segments carrying apnea classification labels for classification, resulting in an accurate classification result indicating whether an apnea event exists in the PPG signal segment. The entire method only requires the acquisition of PPG signals to achieve accurate sleep apnea detection. It is also low in complexity, can be deployed in wearable devices for real-time monitoring, and can be widely used for long-term sleep monitoring in home environments. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating a sleep apnea detection method in one embodiment;

[0046] Figure 2 This is a flowchart illustrating a sleep apnea detection method in another embodiment;

[0047] Figure 3 This is a flowchart illustrating the steps for extracting a one-dimensional temporal feature vector in one embodiment.

[0048] Figure 4 This is a flowchart illustrating the step of extracting two-dimensional image feature vectors in another embodiment;

[0049] Figure 5 This is a flowchart illustrating the sleep apnea detection method in yet another embodiment;

[0050] Figure 6 This is a structural block diagram of a sleep apnea detection device in one embodiment;

[0051] Figure 7 This is an internal structural diagram of a computer device in one embodiment;

[0052] Figure 8 This is an internal structure diagram of an embedded device in one embodiment. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] It should be noted that the terms "comprising" and "having," and any variations thereof, as used in this application, are intended to cover non-exclusive inclusion. The term "multiple" as used in this application refers to two or more. The term "and / or" as used in this application refers to one of the solutions, or any combination of multiple solutions.

[0055] In one embodiment, such as Figure 1As shown, a method for detecting sleep apnea is provided. This embodiment illustrates the application of this method to a wearable health monitoring device. It is understood that this method can also be applied to a server, or to a system including both a wearable health monitoring device and a server, and can be implemented through the interaction between the wearable health monitoring device and the server. In this embodiment, the method includes the following steps:

[0056] Step 100: Obtain PPG signal data of the user during sleep, preprocess the PPG signal data to obtain multiple PPG signal segments.

[0057] PPG signal data refers to pulse wave signals acquired by a dual-wavelength (red light + infrared light) photoplethysmography sensor, reflecting changes in blood's absorption of light. PPG signal segment data refers to data segments obtained by dividing a continuous signal into fixed time windows, also known as PPG signal frames.

[0058] In practical applications, wearable health monitoring devices may include, but are not limited to, smart rings, smart bracelets, smartwatches, and smart bands, all of which incorporate dual-wavelength PPG sensors. The PPG signal acquisition module in a wearable health monitoring device includes red and infrared light emitting and receiving sensors to acquire information about the absorption of red and infrared light by the blood at the user's sampling site. The dual-wavelength PPG sensor can be worn on the user's finger, earlobe, or other suitable location. The data sampling rate can be set from 100Hz to 256Hz to ensure the capture of detailed pulse waveforms. In this embodiment, the preset sampling rate is 100Hz.

[0059] In practice, after the user falls asleep, the wearable health monitoring device (hereinafter referred to as the device) activates its dual-wavelength PPG sensor to continuously collect PPG signal data from the fingertip or earlobe at a sampling rate of 100Hz. After collecting the PPG signal data, the device preprocesses the PPG signal data, such as filtering, noise reduction, and frame segmentation, dividing the original PPG signal data into multiple overlapping PPG signal segments. The frame length is 10 seconds, and the step size is 5 seconds, meaning that each frame of PPG signal data has a 50% overlap.

[0060] Step 200: For each PPG signal segment, extract a one-dimensional time-series feature vector from the PPG signal segment. The one-dimensional time-series feature vector is used to reflect the changes in heart rate and blood flow.

[0061] One-dimensional temporal feature vectors are numerical features extracted from each frame of PPG signal that reflect cardiovascular dynamics, including but not limited to the mean pulse interval reflecting heart rate, the mean pulse amplitude reflecting blood flow intensity, and the root mean square of the difference between adjacent pulse amplitudes reflecting blood flow fluctuations.

[0062] After obtaining multiple PPG signal segment data, for each PPG signal segment, the current PPG signal segment can be normalized. Then, pulse peak detection is performed, and statistical analysis is performed on each detected pulse peak to obtain a one-dimensional time-domain feature vector, such as the mean pulse interval, the mean pulse amplitude, and the root mean square of the difference between adjacent pulse amplitudes.

[0063] Step 300: Convert the PPG signal segment into a two-dimensional spectrum image, and extract the two-dimensional image feature vector based on the two-dimensional spectrum image. The two-dimensional image feature vector is used to reflect the spectrum structure characteristics.

[0064] In this embodiment, the two-dimensional spectral image refers to a two-dimensional time-frequency energy distribution map obtained by performing a time-domain transformation on a PPG signal segment. The two-dimensional image feature vector is an image-level feature extracted from this time-frequency energy distribution map, including energy ratios and histograms, etc.

[0065] Following the previous step, the device can perform time-domain transformation on the PPG signal segment to obtain a two-dimensional spectrum image. Then, it can extract the content of the preset frequency band from the two-dimensional spectrum image and normalize it into a fixed-size image. Based on the fixed-size image, the two-dimensional image feature vector is extracted from the two-dimensional spectrum image.

[0066] Step 400: The one-dimensional temporal feature vector and the two-dimensional image feature vector are fused to obtain the fused feature vector.

[0067] Specifically, a one-dimensional temporal feature vector can be concatenated and fused with a two-dimensional image feature vector to obtain a fused feature vector. For example, if each frame of the PPG signal, after feature extraction, yields a 7-dimensional one-dimensional temporal feature vector and a 61-dimensional two-dimensional image feature vector, the two-dimensional image feature vector can be flattened and transformed into a one-dimensional form, and then concatenated with the original one-dimensional temporal feature vector to obtain a multi-dimensional fused feature vector for each frame.

[0068] Step 500: Standardize and reduce the dimensionality of the fused feature vector to obtain the target feature vector.

[0069] For example, in order to integrate the algorithm into embedded devices and achieve 68-dimensional fused feature vectors, principal component analysis (PCA) pre-trained on a large amount of data is used to reduce the dimensionality of the features.

[0070] First, the multidimensional features of each frame are standardized:

[0071]

[0072] in For each frame, the fused feature vector is... and These are the mean and standard deviation saved from the training phase. This is the fused feature vector after standardization.

[0073] Next, using the pre-trained principal component analysis matrix, the standardized data is projected onto the dimensionality-reduced space:

[0074]

[0075] in The principal component analysis matrix is ​​trained in the previous stage. This is the target feature vector after dimensionality reduction. The target dimension of the target feature vector after dimensionality reduction is approximately 15-20.

[0076] In this embodiment, the target feature vector can be reduced from 68 dimensions to 15 dimensions through the above-described standardization and dimensionality reduction processes.

[0077] Step 600: Input the target feature vector into the trained classification model for classification to obtain the classification result of whether the PPG signal segment contains an apnea event. The classification model is trained based on historical PPG signal segments carrying apnea classification labels.

[0078] In this embodiment, the classification model refers to an SVM or other classifier that has been pre-trained on a large number of historical PPG signal segments carrying apnea classification labels. The parameters are stored in the device and used to determine in real time whether apnea has occurred.

[0079] Following the previous step, after obtaining the dimensionality-reduced target feature vector, it can be input into a trained classification model for classification. The model outputs in real time the classification result indicating whether apnea events exist within that time period. The classification result can be a binary label of "0" (normal) or "1" (apnea), or it can be a condition where the output probability is greater than a preset probability threshold, indicating the presence of apnea; otherwise, sleep is considered normal. The classification result can be updated every 5-10 seconds. It is understood that the update frequency of the classification result can be set according to actual conditions and is not uniquely limited. Furthermore, the classification result can be cached locally and synchronized to a mobile application (APP) the next day to generate a relevant sleep apnea detection report.

[0080] The aforementioned sleep apnea detection method only requires the acquisition of PPG signals. By extracting a one-dimensional temporal feature vector reflecting changes in heart rate and blood flow, it can effectively capture physiological changes such as increased heart rate variability and aggravated blood flow fluctuations caused by apnea. By converting PPG signal segments into two-dimensional spectral images and extracting two-dimensional image feature vectors reflecting spectral structure characteristics, and then fusing the one-dimensional temporal features and two-dimensional spectral image features, it fully leverages the advantages of temporal dynamic sensitivity and frequency domain pattern recognition to comprehensively capture subtle changes in apnea events, thus significantly improving the accuracy and robustness of sleep apnea detection. Subsequently, by standardizing and reducing the dimensionality of the fused feature vector, a target feature vector is obtained, simplifying the algorithm complexity. This target feature vector is then input into a classification model trained on historical PPG signal segments carrying apnea classification labels for classification, resulting in an accurate classification result indicating whether an apnea event exists in the PPG signal segment. The entire method only requires the acquisition of PPG signals to achieve accurate sleep apnea detection, has low complexity, can be deployed in wearable devices for real-time monitoring, and can also be widely applied to long-term sleep monitoring in home environments.

[0081] The preprocessing method is not limited. For example... Figure 2 As shown, in an exemplary embodiment, step 100 includes:

[0082] Step 120: Obtain PPG signal data of the user during sleep, filter the PPG signal data, and perform frame processing according to a fixed time window and sliding step size to obtain multiple PPG signal segments with partial overlap between frames.

[0083] In this embodiment, the sampling frequency of the PPG signal data is 100Hz. Smoothing filtering is performed on the PPG signal data within a preset 1-second window. This operation can be expressed as:

[0084]

[0085] in, This is the PPG signal data after smoothing and filtering. The current PPG signal data is represented by L, which is the preset time window length.

[0086] The baseline interference of the PPG signal data can be filtered out through the above smoothing filtering operation. After the PPG signal is smoothed and filtered, it is then subjected to FIR (Finite Impulse Response) low-pass filtering. Specifically, the cutoff frequency of the PPG signal is 5Hz to filter out high-frequency noise such as power frequency interference in the PPG signal. After the filtering process is completed, the signal is framed according to a preset time window of 10 seconds, with a window sliding step of 5 seconds. Thus, each PPG signal segment (i.e., each frame of PPG signal) has a 50% overlap with the previous PPG signal segment. It can be understood that the window length and window sliding step can be set according to the actual situation and are not limited here.

[0087] In this embodiment, a preprocessing strategy combining filtering and overlapping framing effectively improves signal quality and event capture capabilities. The 50% overlap design enhances the temporal localization accuracy of brief breathing apnea events and avoids missed detections caused by frame boundary cutting. Filtering ensures the stability of subsequent feature extraction, providing high-quality input for high-precision detection.

[0088] like Figure 3 As shown, in an exemplary embodiment, extracting a one-dimensional temporal feature vector from a PPG signal segment includes:

[0089] Step 220: For each PPG signal segment, determine the first-order difference of the PPG signal segment, and locate the pulse peak based on the first-order difference.

[0090] The first-order difference is the difference between adjacent sampling points, used to detect the rising edge of the waveform.

[0091] In this embodiment, for each PPG signal segment (i.e. each PPG signal frame), the first-order difference of each PPG signal frame can be calculated, and the peak position of each pulse wave in the PPG signal can be detected by the positive zero-crossing point of the first-order difference.

[0092] Specifically, the formula for calculating its first-order difference is:

[0093]

[0094] Where n is the index of the current data point, and the range of n is from 2 to the length of each frame of data points. This refers to the pulse wave data value at that point. This is the first-order difference value at that point.

[0095] Step 240: Based on the position and amplitude of each pulse peak, determine the pulse interval sequence and pulse amplitude sequence.

[0096] Step 260: Based on the pulse interval sequence and pulse amplitude sequence, determine the root mean square of the mean pulse period, the standard deviation of the pulse interval, the mean pulse amplitude, and the difference between adjacent pulse amplitudes.

[0097] Mean pulse cycle ( This can help identify changes in heart rate, especially accelerations or decelerations during sleep or sleep apnea. The calculation formula is:

[0098]

[0099] Where N is the length of the pulse interval sequence, and n is the index of the current data point, ranging from 1 to N. The pulse interval at that point. This represents the average pulse cycle of the frame.

[0100] The standard deviation of normal-to-normal intervals (SDNN) is an important indicator for assessing the function of the autonomic nervous system. The heart rate of healthy individuals typically exhibits a certain degree of fluctuation, primarily regulated by the sympathetic and parasympathetic nervous systems. During sleep apnea, sympathetic nerve excitability increases, leading to greater fluctuations in the heart rate intervals and consequently increasing its standard deviation. A larger standard deviation usually indicates significant heart rate fluctuations, which may be a manifestation of sleep apnea or other cardiovascular abnormalities. The formula for calculating SDNN is:

[0101]

[0102] Where N is the length of the pulse interval sequence, and n is the index of the current data point, ranging from 1 to N. The pulse interval at that point. This represents the average pulse cycle of the frame. This represents the standard deviation of the pulse interval for that frame.

[0103] Mean pulse amplitude ( The mean pulse amplitude is directly related to changes in blood flow and is typically influenced by cardiac output and peripheral vascular resistance. During apnea, airway obstruction increases sympathetic nerve activity, leading to elevated blood pressure and peripheral vasoconstriction, which in turn affects pulse amplitude. During apnea, the mean pulse amplitude may exhibit periodic decreases or increases, reflecting hemodynamic changes. The calculation formula is as follows:

[0104]

[0105] Where N is the length of the pulse interval sequence, and n is the index of the current data point, ranging from 1 to N. This represents the pulse amplitude at that data point. This represents the average pulse amplitude for that frame.

[0106] The root mean square of successive differences (RMSD) reflects rapid changes in pulse amplitude. During apnea, pulse amplitude can change significantly over time due to fluctuations in blood flow and blood pressure. During apnea, sympathetic activation causes vasoconstriction, which can lead to periodic fluctuations in pulse amplitude, thus affecting amplitude differences. A high RMSD usually indicates drastic changes in pulse amplitude, which may be an indication of apnea, hypoxia, or hemodynamic instability. Specifically, its calculation formula is:

[0107]

[0108] Where N is the length of the pulse interval sequence, and n is the index of the current data point. This represents the pulse amplitude at that point. It is the root mean square of the difference in amplitude between adjacent pulses in this PPG signal frame.

[0109] Step 280: Determine the sample entropy of each PPG signal segment under different preset scale factors to obtain the multi-scale entropy.

[0110] Multi-scale entropy (MSE) is a nonlinear feature used to assess signal complexity. It provides a more comprehensive description by analyzing the complexity of a signal across multiple scales. In this embodiment, MSE is used to capture changes in the PPG signal at different time scales, helping to better identify abnormal signal patterns in sleep apnea. Under normal circumstances, the complexity of the PPG signal is relatively stable. During apnea, due to decreased blood oxygenation and sympathetic nerve activation, the complexity of the PPG signal changes significantly, and MSE can capture this change. Specifically, the steps for determining the multi-scale entropy are as follows:

[0111] 1. Resample the original signal to generate signals at different scales. ,in It is a scaling factor, for each Scale, by calculating each This is achieved by taking the mean of several data points. In this embodiment, You can choose 1, 2, or 3. The formula is as follows, where... For scale The resampled signal below:

[0112]

[0113] 2. For the downsampled signal, extract subsequences of length m and m+1 (embedding dimension) from the signal. In this embodiment, m is 2. The implementation formula is as follows, where... and The resulting subsequence:

[0114]

[0115]

[0116] 3. For each pair and The maximum distance is used to calculate the distance between the two, and the calculation formula is as follows:

[0117]

[0118] 4. Given a tolerance level Statistical analysis is performed when the standard deviation of the original data is 20%. The quantity.

[0119] 5. Sample Entropy The complexity of a signal is measured by calculating the similarity ratio between subsequences of length m and m+1. The formula for sample entropy is as follows: where A is a pair of subsequences of length m, satisfying the number of distances. B is a pair of subsequences of length m+1, satisfying the number of distances.

[0120]

[0121] 6. Calculate at different scales The sample entropy is then used to obtain the MSE value:

[0122]

[0123] For multi-scale entropy, the sample entropy of each frame of signal is calculated at scale factors τ=1, 2, 3 to obtain the multi-scale entropy features of that frame. When calculating the sample entropy, the embedding dimension is preset to 2, and the tolerance threshold is preset to 20% of the variance.

[0124] In this embodiment, the one-dimensional temporal feature vector includes the average pulse period, the standard deviation of the pulse interval, the average pulse amplitude, the root mean square of the difference between adjacent pulse amplitudes, and multi-scale entropy. It is understood that in other embodiments, the one-dimensional temporal feature vector may include at least one of the following: the average pulse period, the standard deviation of the pulse interval, the average pulse amplitude, the root mean square of the difference between adjacent pulse amplitudes, and multi-scale entropy; the specific selection can be made according to actual circumstances and requirements.

[0125] In this embodiment, by extracting a set of highly sensitive time-domain physiological features from the PPG signal, especially by introducing multi-scale entropy, it is possible to capture the physiological signal "decomplication" phenomenon caused by autonomic nervous system imbalance during sleep apnea, which significantly improves the detection sensitivity for early or mild sleep apnea events.

[0126] In one exemplary embodiment, such as Figure 4 As shown, the PPG signal segment is converted into a two-dimensional spectral image, and the two-dimensional image feature vector is extracted from the two-dimensional spectral image, including:

[0127] Step 320: Perform time-frequency transformation on the PPG signal segment to obtain a two-dimensional spectrum image.

[0128] A two-dimensional spectral image refers to a two-dimensional time-frequency energy distribution map. In this embodiment, the PPG signal of each frame can be converted into a two-dimensional time-frequency energy distribution map using Stockwell transform. Stockwell transform (ST) is a time-frequency analysis tool that combines the advantages of Fourier transform and wavelet transform, providing precise signal localization in the time and frequency domains. ST generates a two-dimensional time-frequency map, where the horizontal axis represents time, the vertical axis represents frequency, and each point represents the signal strength at that time and frequency. The conversion formula is as follows:

[0129]

[0130] in, The input PPG signal; The result of the transformation over time t and frequency f is represented as a point on a two-dimensional spectrum image; It is a Gaussian window function, the size of which determines the locality of the signal in time and frequency; It is a complex exponential function used to calculate the frequency components of a signal.

[0131] It is understood that in other embodiments, in addition to the Stockwell transform, other time-frequency transformation algorithms, such as short-time Fourier transform and continuous wavelet transform, may be used, depending on the specific circumstances.

[0132] Step 340: Perform frequency domain cropping and geometric normalization on the two-dimensional spectral image to obtain the target two-dimensional image.

[0133] Since apnea and related physiological phenomena typically exhibit low-frequency changes, a frequency range related to the apnea frequency, such as 0.05–0.5 Hz, can be extracted from the obtained two-dimensional spectral image. This facilitates focused analysis of low-frequency components related to respiration and heart rate. Therefore, in practice, only the 0.05–0.5 Hz frequency range can be extracted from the two-dimensional spectral image.

[0134]

[0135] After extracting a frequency range of 0.05–0.5 Hz from the two-dimensional spectrogram image, the size of the extracted two-dimensional spectrogram image is normalized to a target two-dimensional image of 64×64. It can be understood that the normalized image size can also be other sizes, depending on the specific circumstances.

[0136] Step 360: Extract the two-dimensional image feature vector based on the target two-dimensional image.

[0137] After obtaining a 64×64 target 2D image, the 2D image feature vector can be extracted from the target 2D image.

[0138] In this embodiment, a two-dimensional spectral image is generated by introducing time-frequency transformation, overcoming the limitations of traditional methods that only analyze the time or frequency domains. This allows for the intuitive capture of typical patterns in PPG signals during apnea, such as low-frequency energy enhancement and spectral disturbances. This image-based representation lays the foundation for subsequent advanced feature extraction methods such as texture analysis and template comparison, significantly improving the ability to characterize non-stationary physiological signals.

[0139] In an exemplary embodiment, a two-dimensional image feature vector is extracted from a two-dimensional spectral image based on a target two-dimensional image, including at least one of the following methods:

[0140] The first step is to determine the ratio of frequency energy to total energy in the target two-dimensional image within a preset low-frequency range, thus obtaining the low-frequency energy percentage.

[0141] Low-frequency energy ratio refers to the ratio of low-frequency energy (e.g., 0.1–0.3 Hz) to total energy, used to quantify the relative activity of respiratory-related frequency components. This ratio often increases significantly during sleep apnea.

[0142] In practical applications, low-frequency energy may increase or decrease during sleep apnea. For example, an increase in low-frequency energy after apnea indicates an overreaction of the sympathetic nervous system. These characteristic changes can be seen as manifestations of respiratory-heart rate coupling changes caused by apnea, with the energy in the 0.1Hz to 0.3Hz frequency band primarily related to respiration and cardiac activity. Therefore, in this embodiment, the ratio of the low-frequency energy (0.1–0.3Hz) to the total frequency energy is calculated to reflect the proportion of low-frequency components, directly measuring the degree of respiratory regulation imbalance and facilitating the detection of transient but significant respiratory change patterns. The calculation formula is as follows, where ER is the calculated proportion of low-frequency energy:

[0143]

[0144] The second step involves comparing each pixel in the target 2D image with its eight neighboring pixels and determining the local binary pattern of the target 2D image based on the comparison results.

[0145] Local Binary Pattern (LBP) is the extracted local texture information of a time-frequency image (spectral image), that is, the gray-level pattern changes between adjacent pixels in the spectrum. It is usually used to measure the texture roughness, detail structure, etc. of an image.

[0146] LBP extracts local texture information from the spectrogram, specifically the grayscale pattern changes between adjacent pixels in the spectrum. It is typically used to measure image texture roughness, detail, and other structural features. Sleep apnea causes sympathetic nerve excitation, heart rate fluctuations, and changes in vascular tone, resulting in textural perturbations or changes in the PPG spectrum. LBP can capture these changes in the local structure of the spectrogram before and after sleep apnea events, such as sudden energy enhancement, local blurring, and boundary changes.

[0147] Specifically, LBP feature extraction on a two-dimensional spectrogram can be performed as follows: for each pixel... The pixel is compared with its eight neighboring pixels to obtain the binary comparison result, which is then converted to decimal:

[0148]

[0149] in, For pixels If the comparison result with each pixel is greater than the center pixel, then... =1, otherwise =0.

[0150] Next, the LBP value is calculated for each pixel in the two-dimensional spectrogram, generating an LBP histogram. Specifically, in this embodiment, the histogram has a dimension of 59, representing different texture modes of rotation-invariant LBP. This histogram describes the distribution of local textures in the image. It is understood that in other embodiments, the dimension of the LBP histogram can be other values, which can be set according to the actual situation.

[0151] The third step involves transforming the target two-dimensional image into a probability distribution and comparing it with a preset respiratory spectrum template to determine the relative entropy between the two-dimensional image and the preset respiratory spectrum template.

[0152] Relative entropy, also known as KL (Kullback-Leibler Divergence), measures the degree of difference between the current spectral energy distribution and a baseline distribution; it is a measure of "energy anomaly." The preset respiratory spectrogram template is a standard reference model built upon the average time-frequency energy distribution of nighttime PPG data from a large number of healthy individuals and stored in the device. Its characteristics include energy concentrated in the 0.1–0.3 Hz range with uniform distribution and clear texture.

[0153] KL divergence measures the degree of difference between the current spectral energy distribution and a baseline distribution, serving as a measure of "energy anomaly." During apnea, the energy distribution in the PPG spectrum differs significantly, such as increased low frequencies and peak shift. KL divergence effectively captures these distributional differences. A large KL divergence indicates that the energy distribution deviates from the normal state, potentially reflecting apnea or abnormal heart rate regulation. Its calculation formula is:

[0154]

[0155] in, This is a spectrogram template for normal sleep breathing conditions, determined based on a large amount of experimental data from previous studies. This refers to the degree of difference between the pre-specified distribution and the respiratory spectrogram template.

[0156] Specifically, a two-dimensional image feature vector may include at least one of low-frequency energy proportion, local binary pattern, and relative entropy. For example, in this embodiment, the two-dimensional image feature vector including low-frequency energy proportion, local binary pattern, and KL divergence is used as an example for illustration.

[0157] In this embodiment, by converting the PPG signal into a two-dimensional spectral image and extracting three types of image features—low-frequency energy ratio, LBP texture features, and KL divergence—it is helpful to achieve high-precision and robust recognition of sleep apnea events.

[0158] like Figure 5 As shown, in an exemplary embodiment, step 600 includes:

[0159] Step 620: Input the target feature vector into the trained support vector machine classifier for classification to obtain the classification label of whether the PPG signal segment contains a sleep apnea event. The vector machine classifier uses a radial basis function kernel.

[0160] In this embodiment, a Support Vector Machine (SVM) classifier is selected as the classifier for sleep apnea event recognition. The goal of SVM is to separate samples of different classes by selecting a hyperplane that maximizes the margin between classes. This hyperplane can be represented by a linear function:

[0161]

[0162] in The normal vector is the direction of the hyperplane. It is the input feature vector, which in this application is the dimensionality-reduced feature vector. , This is the bias term.

[0163] In this application, an SVM classifier using a radial basis function (RBF) kernel was pre-trained and stored using a large amount of historical PPG signal fragment data carrying apnea classification labels. In this embodiment, the SVM classifier is applied as follows:

[0164]

[0165] in Let be the Lagrange multiplier of the support vectors. This is the result of the RBF kernel function calculation. This is the output of the SVM classifier. Where if If the value is greater than 0, the prediction result indicates that a sleep apnea event exists in the PPG signal segment, meaning that a sleep apnea event occurred during the corresponding time period of the PPG signal segment; if... If the value is less than 0, the prediction result is that the PPG signal segment is in a normal state and no sleep apnea event has occurred.

[0166] It is understood that in other embodiments, support vector machines may also employ kernel functions such as linear kernel functions and polynomial kernel functions, depending on the specific circumstances.

[0167] In this embodiment, a support vector machine classifier based on RBF kernel is used, which can not only efficiently and accurately complete the final discrimination from "target feature vector" to "apnea event label", but also take into account algorithm performance, computational efficiency and engineering feasibility.

[0168] To provide a clearer explanation of the sleep apnea detection method provided in this application, a specific embodiment is described below, which includes the following:

[0169] After the user falls asleep, the wearable health monitoring device activates the dual-wavelength PPG sensor to continuously collect raw PPG signals from the fingertip or earlobe at a sampling rate of 100Hz. The raw signal is smoothed by a 1-second sliding window to remove baseline drift, and then filtered by a 5Hz FIR low-pass filter to remove high-frequency noise, resulting in a filtered signal. The filtered signal is then framed according to a 10-second window and a 5-second step size to generate overlapping (50% overlap between frames) PPG signal segments.

[0170] Next, for each frame of PPG signal: calculate the first-order difference → detect positive zero-crossing points to locate the pulse peak → calculate the time difference between adjacent peaks to obtain the pulse interval sequence, and calculate the peak amplitude to obtain the pulse amplitude sequence → based on the pulse interval sequence and pulse amplitude sequence, calculate the root mean square of the mean pulse period, pulse interval standard deviation, mean pulse amplitude, and the difference between adjacent pulse amplitudes → then perform multi-scale resampling on the original signal and calculate the sample entropy (m=2, r=0.2×std) → obtain a 7-dimensional one-dimensional time series feature vector containing 5 types of indicators.

[0171] A Stockwell transform is performed on the PPG signal segment to generate a two-dimensional spectrum image. The frequency range corresponding to 0.1–0.5 Hz is extracted from the two-dimensional spectrum image and the extracted content is normalized to a target two-dimensional image of 64×64. Based on the target two-dimensional image, 61-dimensional two-dimensional image features are extracted, including low-frequency energy ratio, LBP histogram, and KL divergence.

[0172] The two types of features, one-dimensional time-series feature vector and two-dimensional image feature vector, are concatenated into a 68-dimensional fusion vector. The fusion vector is then standardized and reduced to 15 dimensions using a pre-stored PCA matrix to obtain the target feature vector.

[0173] Input the target feature vector into a trained SVM classifier with RBF kernel, and output a binary label indicating whether the frame is a sleep apnea event.

[0174] Finally, all classification results are cached locally and synchronized to the mobile app via Bluetooth the next day for generating monitoring reports and health recommendations.

[0175] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0176] Based on the same inventive concept, this application also provides a sleep apnea detection device for implementing the sleep apnea detection method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations of the one or more sleep apnea detection device embodiments provided below can be found in the limitations of the sleep apnea detection method described above, and will not be repeated here.

[0177] In one exemplary embodiment, such as Figure 6 As shown, a sleep apnea detection device 600 is provided, including: a data acquisition module 610, a one-dimensional feature extraction module 620, a two-dimensional feature extraction module 630, a feature fusion module 640, a dimensionality reduction processing module 650, and an apnea detection module 660, wherein:

[0178] The data acquisition module 610 is used to acquire PPG signal data of the user during sleep, preprocess the PPG signal data to obtain multiple PPG signal segments;

[0179] The one-dimensional feature extraction module 620 is used to extract a one-dimensional time-series feature vector from each PPG signal segment. The one-dimensional time-series feature vector is used to reflect the changes in heart rate and blood flow.

[0180] The two-dimensional feature extraction module 630 is used to convert PPG signal segments into two-dimensional spectrum images, and extract two-dimensional image feature vectors based on the two-dimensional spectrum images. The two-dimensional image feature vectors are used to reflect the spectrum structure characteristics.

[0181] The feature fusion module 640 is used to fuse one-dimensional temporal feature vectors with two-dimensional image feature vectors to obtain fused feature vectors.

[0182] The dimensionality reduction module 650 is used to standardize and reduce the dimensionality of the fused feature vector to obtain the target feature vector.

[0183] The apnea detection module 660 is used to input the target feature vector into a trained classification model for classification, and obtain the classification result of whether the PPG signal segment contains an apnea event. The classification model is trained based on historical PPG signal segments carrying apnea classification labels.

[0184] In one embodiment, the two-dimensional feature extraction module 630 is further configured to perform time-frequency transformation on the PPG signal segment to obtain a two-dimensional spectrum image, perform frequency domain cropping and geometric normalization processing on the two-dimensional spectrum image to obtain a target two-dimensional image, and extract the two-dimensional image feature vector based on the target two-dimensional image.

[0185] In one embodiment, the two-dimensional feature extraction module 630 is further configured to perform any of the following:

[0186] The first step is to determine the ratio of frequency energy to total energy in the target two-dimensional image within a preset low-frequency range, thus obtaining the low-frequency energy percentage.

[0187] The second step involves comparing each pixel in the target two-dimensional image with its eight neighboring pixels and determining the local binary pattern of the target two-dimensional image based on the comparison results.

[0188] The third step involves, based on the target two-dimensional image, converting the two-dimensional spectral image into a probability distribution, and comparing the two-dimensional spectral image with a preset respiratory spectral image template to determine the relative entropy between the two-dimensional spectral image and the preset respiratory spectral image template.

[0189] The two-dimensional image feature vector includes at least one of low-frequency energy proportion, local binary pattern, and relative entropy.

[0190] In one embodiment, the data acquisition module 610 is further configured to filter the PPG signal data and perform frame segmentation processing according to a fixed time window and sliding step size to obtain multiple PPG signal segments with partial overlap between frames.

[0191] In one embodiment, the one-dimensional feature extraction module 620 is further configured to determine the first-order difference of each PPG signal segment, locate the pulse peak based on the first-order difference, determine the pulse interval sequence and pulse amplitude sequence based on the position and amplitude of each pulse peak, and determine the mean pulse period, pulse interval standard deviation, mean pulse amplitude, and root mean square of the difference between adjacent pulse amplitudes based on the pulse interval sequence and pulse amplitude sequence. The sample entropy of each PPG signal segment under preset scale factors is determined to obtain multi-scale entropy; wherein, the one-dimensional temporal feature vector includes at least one of the following: mean pulse period, pulse interval standard deviation, mean pulse amplitude, root mean square of the difference between adjacent pulse amplitudes, and multi-scale entropy.

[0192] In one embodiment, the sleep apnea detection module 660 is further configured to input the target feature vector into a trained support vector machine classifier for classification to obtain a classification label indicating whether a sleep apnea event exists in the PPG signal segment, wherein the support vector machine classifier employs a radial basis function kernel.

[0193] The modules in the aforementioned sleep apnea detection device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0194] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data such as PPG signals and classification results. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a sleep apnea detection method.

[0195] In one exemplary embodiment, an embedded device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in any of the above embodiments of the sleep apnea detection method.

[0196] Specifically, the internal structure diagram of an embedded device can be as follows: Figure 8 As shown, the embedded device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a sleep apnea detection method. The display unit of the embedded device is used to form a visually visible image. It can be a display screen. The input device of the embedded device can be a touch layer covering the display screen, or a button, trackball, or touchpad set on the housing of the embedded device, or an external keyboard, touchpad, or mouse, etc.

[0197] For example, embedded devices include, but are not limited to, smart wearable devices such as smart bracelets, smartwatches, and smart rings, as well as other health monitoring devices with physiological monitoring functions.

[0198] In some exemplary embodiments, taking an embedded device as an example of a sleep monitoring device, the sensor portion of the sleep monitoring device is made of medical-grade flexible material to ensure no irritation to the user's skin and avoid pressure on the skin, while also avoiding affecting the user's activities. The overall design of the device is lightweight and convenient, ensuring stable signal acquisition while minimizing restrictions on the user's activities, thus achieving non-invasive and harmless continuous signal acquisition.

[0199] Those skilled in the art will understand that Figure 8The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0200] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in any of the above embodiments of the sleep apnea detection method.

[0201] In one exemplary embodiment, an embedded device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in any of the above embodiments of the sleep apnea detection method.

[0202] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in any of the above embodiments of the sleep apnea detection method.

[0203] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the sleep apnea detection method.

[0204] 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.

[0205] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0206] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0207] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A sleep apnea detection method, characterized by, The method comprises: acquiring PPG signal data of a user during sleep, pre-processing the PPG signal data to obtain a plurality of PPG signal segments; for each PPG signal segment, extracting a one-dimensional time sequence feature vector from the PPG signal segment, the one-dimensional time sequence feature vector being used to reflect the changes of heart rate and blood flow, including average pulse period, pulse interval standard deviation, average pulse amplitude, root mean square of adjacent pulse amplitude difference, and multi-scale entropy; performing time-frequency transformation on the PPG signal segment to obtain a two-dimensional frequency spectrum image, performing frequency domain clipping and geometric normalization processing on the two-dimensional frequency spectrum image to obtain a target two-dimensional image, and extracting a two-dimensional image feature vector from the target two-dimensional image, the two-dimensional image feature vector being used to reflect the spectral structure characteristics, including low-frequency energy proportion, local binary pattern, and relative entropy between the two-dimensional frequency spectrum image and a preset respiratory spectrum image template; fusing the one-dimensional time sequence feature vector and the two-dimensional image feature vector to obtain a fused feature vector; performing standardization and dimensionality reduction processing on the fused feature vector to obtain a target feature vector; inputting the target feature vector into a trained classification model for classification to obtain a classification result of whether the PPG signal segment has an apnea event, the classification model being trained based on historical PPG signal segments carrying apnea classification labels; wherein, according to the target two-dimensional image, the two-dimensional image feature vector is extracted, including: determining the ratio of frequency energy to total energy of the target two-dimensional image in a preset low-frequency range to obtain a low-frequency energy proportion, the low-frequency energy proportion being used to quantify the relative activity degree of the respiratory frequency component; for each pixel point in the target two-dimensional image, comparing the pixel point with eight adjacent pixel points, and determining the local binary pattern of the target two-dimensional image according to the comparison result; based on the target two-dimensional image, the two-dimensional frequency spectrum image is equivalent to a probability distribution, and the two-dimensional frequency spectrum image is compared with a preset respiratory spectrum image template to determine the relative entropy between the two-dimensional frequency spectrum image and the preset respiratory spectrum image template.

2. The method of claim 1, wherein, The pre-processing of the PPG signal data to obtain a plurality of PPG signal segments comprises: filtering the PPG signal data and performing frame processing according to a fixed time window and a sliding step to obtain a plurality of PPG signal segments with partially overlapping frames.

3. The method of claim 1, wherein, The one-dimensional time sequence feature vector extracted from the PPG signal segment comprises: for each PPG signal segment, determining a first-order difference of the PPG signal segment and locating a pulse peak based on the first-order difference; based on the position and amplitude of each pulse peak, determining a pulse interval sequence and a pulse amplitude sequence; based on the pulse interval sequence and the pulse amplitude sequence, determining the average pulse period, the pulse interval standard deviation, the average pulse amplitude, and the root mean square of the adjacent pulse amplitude difference; determining the sample entropy of each PPG signal segment under a preset different scale factor to obtain the multi-scale entropy.

4. The method according to any one of claims 1 to 3, characterized in that, The target feature vector is input into a trained classification model for classification to obtain a classification result of whether the PPG signal segment has an apnea event. The target feature vector is input into a trained support vector machine classifier for classification to obtain a classification label of whether the PPG signal segment has an apnea event. The support vector machine classifier adopts a radial basis function kernel.

5. A sleep apnoea detection apparatus characterised by The device comprises: A data acquisition module configured to acquire PPG signal data of a user during sleep, and to pre-process the PPG signal data to obtain a plurality of PPG signal segments. A one-dimensional feature extraction module configured to extract, for each PPG signal segment, a one-dimensional time series feature vector from the PPG signal segment, the one-dimensional time series feature vector being used to reflect changes in heart rate and blood flow, and comprising average pulse period, pulse interval standard deviation, average pulse amplitude, root mean square of adjacent pulse amplitude difference, and multi-scale entropy. A two-dimensional feature extraction module configured to perform time-frequency transformation on the PPG signal segment to obtain a two-dimensional frequency spectrum image, to perform frequency domain clipping and geometric normalization processing on the two-dimensional frequency spectrum image to obtain a target two-dimensional image, and to extract a two-dimensional image feature vector from the target two-dimensional image, the two-dimensional image feature vector being used to reflect frequency spectrum structure characteristics, and comprising low-frequency energy proportion, local binary pattern, and relative entropy between the two-dimensional frequency spectrum image and a preset respiratory frequency spectrum image template. A feature fusion module configured to fuse the one-dimensional time series feature vector and the two-dimensional image feature vector to obtain a fused feature vector. A dimension reduction processing module configured to perform standardization and dimension reduction processing on the fused feature vector to obtain a target feature vector. An apnea detection module configured to input the target feature vector into a trained classification model for classification to obtain a classification result of whether the PPG signal segment has an apnea event, the classification model being trained based on historical PPG signal segments carrying apnea classification labels. The two-dimensional feature extraction module is further configured to determine a ratio of frequency energy of the target two-dimensional image in a preset low-frequency range to total energy to obtain a low-frequency energy proportion, the low-frequency energy proportion being used to quantify a relative activity level of a respiratory frequency component, to compare each pixel point in the target two-dimensional image with eight adjacent pixel points to determine a local binary pattern of the target two-dimensional image according to a comparison result, to equivalently convert the two-dimensional frequency spectrum image into a probability distribution based on the target two-dimensional image, and to compare the two-dimensional frequency spectrum image with a preset respiratory frequency spectrum image template to determine relative entropy between the two-dimensional frequency spectrum image and the preset respiratory frequency spectrum image template.

6. The apparatus of claim 5, wherein, The data acquisition module is further configured to perform filtering processing on the PPG signal data, and to perform frame processing on the PPG signal data according to a fixed time window and a sliding step to obtain a plurality of PPG signal segments with partially overlapped frames.

7. The apparatus of claim 5, wherein, The one-dimensional feature extraction module is further configured to determine, for each PPG signal segment, a first-order difference of the PPG signal segment, locate pulse wave peaks based on the first-order difference, determine a pulse interval sequence and a pulse amplitude sequence based on positions and amplitudes of each pulse wave peak, determine an average pulse period, a pulse interval standard deviation, an average pulse amplitude, a root mean square of adjacent pulse amplitude differences based on the pulse interval sequence and the pulse amplitude sequence, and determine a sample entropy of each PPG signal segment at different scale factors to obtain a multi-scale entropy.

8. An embedded device comprising a memory and a processor, said memory storing a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 4.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 4.

10. A computer program product comprising a computer program, characterized in that, The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 4. The computer program, when executed by the processor, implements the steps of the method of any one of claims 1 to 4.

Images