Non-contact sleep staging method and system based on lightweight network

Abstract

The invention discloses a non-contact sleep staging method and system based on a lightweight network, and the method comprises the steps: carrying out the feature extraction from a time domain feature, a frequency domain feature, a phase feature, a nonlinear feature, a linkage feature and a wavelet transformation statistical feature, quantifying the rhythm coupling relation of vital signs in a sleep stage, providing a new discrimination dimension for sleep staging, and improving the sleep staging efficiency. The extraction of the local features of different dimensions effectively retains the local time sequence correlation and improves the feature abstraction ability, two kinds of data are fused, the data meet the requirements of small-scale comprehensiveness, the data are lighter, the problem of data redundancy is avoided, division does not need to be performed depending on artificial experience, and the efficiency is improved. The data obtained by fusion is simpler in later model training and real-time calculation, the calculation complexity is reduced, the later staging result is more accurate, and the complexity of the algorithm is reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of sleep staging, and relates to a non-contact sleep staging method and system based on a lightweight network. Background Art

[0002] Sleep, as the core link in the restoration of human physiological functions, directly affects individual health and quality of life. However, with the accelerating social pace and the intensifying aging population, sleep disorders have become a global public health problem. Although traditional sleep monitoring technologies can provide accurate data, their limitations such as high cost and complex wearing are becoming increasingly prominent, making it urgent to promote the innovation of non-contact technologies.

[0003] For a long time, polysomnography (PSG) has been regarded as the "gold standard" for sleep staging. Through multi-modal signals such as electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG), sleep can be divided into the wake stage, rapid eye movement (REM) stage, and non-rapid eye movement (NREM) stage. However, the application of this technology has always been limited by its complex operation and high cost. Patients need to wear a large number of electrodes in a laboratory environment, which not only causes skin discomfort but also makes it difficult to fall asleep due to the unfamiliar environment, greatly reducing the authenticity of the data. For infants, electrode pasting may interfere with natural sleep and even increase the risk of infection; for the elderly living alone, traditional devices cannot achieve real-time early warning of sudden risks at night. More seriously, primary medical institutions generally lack PSG devices and professional technical personnel, making it difficult for a large number of potential patients to obtain early diagnosis. These pain points indicate that traditional contact technologies are no longer able to meet the inclusive needs of national health management.

[0004] The vast majority of them take the EEG signal or multi-channel fusion signal in PSG as the research object, and through artificial feature extraction combined with traditional machine learning classifiers or deep learning methods, automatic sleep stage staging from binary classification to six-classification has been achieved. Existing methods have to a certain extent proved the feasibility of using multi-channel physiological signals such as heart rate, EEG, EOG, and EMG for sleep staging. However, the existing sleep staging research has a relatively high dependence on researchers' knowledge in this field and is also relatively sensitive to unknown noise and interference in the signals. In addition, the fusion of multi-channel signals not only increases the algorithm complexity but also limits the generality of the model in other sensing technologies and new test environments. Traditional PSG devices themselves have deficiencies in wearing comfort, real-time monitoring, and long-term continuous monitoring, making it difficult to meet the needs of long-term, non-invasive, and convenient sleep health monitoring in practical applications. In recent years, with the development of radio frequency sensors and millimeter-wave radar technologies, using non-contact monitoring methods to obtain physiological signals such as respiration, heartbeat, and body movement has become a research hotspot, which provides new ideas and directions for the development of sleep staging technologies, but it still faces many challenges in signal processing, feature extraction, and cross-domain data adaptability.

[0005] After years of exploration, significant progress has been made in sleep staging technology based on multi-channel physiological signals. However, the existing technologies still have the following main drawbacks:

[0006] Relying on human experience for classification: Currently, most clinical studies perform manual sleep staging based on traditional signals such as EEG, ECG, and EOG. This method highly depends on the prior knowledge of researchers, requires a large amount of manpower and time costs, has a high degree of subjectivity in stage classification, and is difficult to handle signal variations in complex, non-linear, and high-noise environments.

[0007] Signal fusion and system complexity: Although the acquisition and fusion of multi-channel and multi-lead signals can provide rich multi-dimensional information, it also leads to problems such as high data redundancy and a significant increase in the complexity of model training and real-time calculation, reducing the generality and application scope of the system.

[0008] Insufficient anti-noise performance and domain adaptability: Traditional methods have poor robustness to unknown noise and interference signals, and there are significant domain shift problems among different devices, test environments, and populations. The generalization ability of the model is insufficient, and it is difficult to obtain stable and reliable results in long-term clinical and home monitoring.

[0009] Inconvenient device wearing and monitoring: Traditional PSG monitoring devices are large in size, complex to wear, and not very comfortable, which are not suitable for long-term continuous monitoring. This is significantly contradictory to the requirements of modern health management and telemedicine for portable and non-invasive devices. Summary of the Invention

[0010] The purpose of the present invention is to solve the problems in the existing technology that rely on human experience for classification, have complex signal fusion, resulting in high data redundancy, a significant increase in the complexity of model training and real-time calculation, and affecting the accuracy of staging results. A non-contact sleep staging method and system based on a lightweight network are provided.

[0011] To achieve the above purpose, the present invention adopts the following technical solutions:

[0012] A non-contact sleep staging method based on a lightweight network, comprising the following steps:

[0013] Obtain multi-dimensional data of human sleep;

[0014] Based on the multi-dimensional data, extract time-domain features, frequency-domain features, phase features, non-linear features, linkage features, and wavelet transform statistical features, and fuse the extracted features to obtain data-domain features;

[0015] Extract local features of different dimensions in the multi-dimensional data to form a fused feature vector; splice the fused feature vector and the data-domain features to obtain an enhanced feature vector;

[0016] Classify the enhanced feature vectors to complete sleep staging.

[0017] A further improvement of the present invention lies in:

[0018] The obtaining of multi-dimensional data of human sleep includes:

[0019] Perform non-contact measurement on the human body through a millimeter-wave radar, and combine the PSG data measured synchronously to label the data, obtaining non-contact data with labels.

[0020] In the non-contact data, each sample data includes five dimensions. The 0th dimension and the 1st dimension are respectively the time series data of respiration and heart rate, the 2nd dimension and the 3rd dimension are respectively the amplitude spectra of respiration and heart rate, and the 4th dimension is the distance spectrum data.

[0021] When extracting the non-linear features, it includes:

[0022] Extract the detrended fluctuation scaling exponent, power spectrum entropy, singular spectrum entropy, wavelet transform energy entropy, differential entropy, sample entropy, fuzzy entropy, and permutation entropy of the respiration and heartbeat signals.

[0023] When extracting the associated features, it includes:

[0024] Extract the correlation features between the respiration and heartbeat signals. The correlation features include phase difference statistics, phase locking value, frequency domain coherence, and cross-correlation time delay characteristics.

[0025] The steps of extracting local features of different dimensions in the multi-dimensional data specifically include: constructing a deep learning heterogeneous network, and extracting local features of different dimensions in the multi-dimensional data based on the deep learning heterogeneous network.

[0026] The deep learning heterogeneous network includes a 1D convolutional layer, batch normalization and ReLU activation function, adaptive average pooling, and a feature fusion module.

[0027] The data domain features include 153-dimensional features extracted from the data domain, and the local features of different dimensions include 320-dimensional features extracted by the neural network. The 153-dimensional features and the 320-dimensional features are concatenated and fused to obtain an enhanced feature vector with a total dimension of 473.

[0028] The steps of classifying the enhanced feature vectors include:

[0029] Train the LightGBM model, and classify the enhanced feature vectors through the trained LightGBM model.

[0030] A non-contact sleep staging system based on a lightweight network includes:

[0031] An original data acquisition module for acquiring multi-dimensional data of human sleep;

[0032] A feature extraction module for extracting features from time-domain features, frequency-domain features, phase features, non-linear features, linkage features, and wavelet transform statistical features based on the multi-dimensional data, fusing the extracted features to obtain data-domain features, extracting local features of different dimensions in the multi-dimensional data, and forming a fused feature vector;

[0033] A feature fusion module for fusing and splicing local features and data-domain features to obtain an enhanced feature vector;

[0034] A staging module for classifying the enhanced feature vector to complete sleep staging.

[0035] A terminal device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of any method of the present invention are implemented.

[0036] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, the steps of any method of the present invention are implemented.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention discloses a non-contact sleep staging method based on a lightweight network. Features are extracted from time-domain features, frequency-domain features, phase features, non-linear features, linkage features, and wavelet transform statistical features, quantifying the rhythm coupling relationship of vital signs in the sleep stage, providing a new discrimination dimension for sleep staging. The extraction of local features of different dimensions effectively retains local temporal correlation and improves feature abstraction ability. The two types of data are fused, meeting the requirements of small-scale comprehensiveness, making the data more lightweight, avoiding the problem of data redundancy, and not requiring manual experience for division. The fused data is simpler in later model training and real-time calculation, reducing the computational complexity, contributing to more accurate later staging results, and reducing the complexity of the algorithm.

[0039] Furthermore, in the present invention, human body data is collected by a millimeter-wave radar without any physical contact, completely eliminating the wearing burden, improving the comfort of patients. Its penetration detection ability can penetrate clothes and quilts, adapt to complex environments, and achieve non-sensing monitoring. The millimeter-wave radar analyzes physiological signals through radio waves without recording images or sounds, completely avoiding the risk of privacy leakage.

[0040] Furthermore, in the present invention, a lightweight network is used to construct a sleep staging model to reduce the computational complexity. Brief Description of the Drawings

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0042] Figure 1 It is the overall staging flowchart of the present invention;

[0043] Figure 2 It is the schematic diagram of the data structure of the present invention;

[0044] Figure 3 It is the structure diagram of the data domain feature extraction data set of the present invention;

[0045] Figure 4 It is the MVNet network structure diagram of the present invention;

[0046] Figure 5 It is the comparison schematic diagram of the machine learning integration algorithms of the embodiments of the present invention;

[0047] Figure 6 It is the flowchart of the integration learning algorithm of the embodiments of the present invention;

[0048] Figure 7 It is the loss function diagram of the embodiments of the present invention, (where a is the loss curve of the original data of LightGBM; b is the loss curve of the data augmentation of LightGBM);

[0049] Figure 8 It is the radar chart of the data metrics of the embodiments of the present invention (where a is the radar chart of the metrics of the original data of LightGBM; b is the radar chart of the metrics of the data augmentation of LightGBM; c is the radar chart of the metrics of the original data of XGBoost; d is the radar chart of the metrics of the original data of CatBoost);

[0050] Figure 9 It is the confusion matrix of the embodiments of the present invention (where a is the confusion matrix of the original data of LightGBM; b is the confusion matrix of the data augmentation of LightGBM; c is the confusion matrix of the original data of XGBoost; d is the confusion matrix of the original data of CatBoost);

[0051] Figure 10 It is the heat map of feature importance of the embodiments of the present invention (where a is the feature importance of the original data of LightGBM; b is the feature importance of the data augmentation of LightGBM; c is the feature importance of the original data of XGBoost; d is the feature importance of the original data of CatBoost);

[0052] Figure 11 This is the decision tree visualization diagram of the embodiment of the present invention (where a is the visualization of a single decision tree of the original data of LightGBM; b is the visualization of a single decision tree of the data augmentation of LightGBM). Detailed implementation manners

[0053] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Generally, the components of the embodiments of the present invention described and illustrated in the drawings here can be arranged and designed in various different configurations.

[0054] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents the selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0055] It should be noted that: like reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined and explained in subsequent drawings.

[0056] In the description of the embodiments of the present invention, it should be noted that if terms such as "upper", "lower", "horizontal", "inner", etc. are used to indicate the orientation or positional relationship, it is based on the orientation or positional relationship shown in the drawings, or the orientation or positional relationship in which the product of the invention is usually placed during use. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation to the present invention. In addition, terms such as "first", "second", etc. are only used for distinguishing descriptions and cannot be understood as indicating or implying relative importance.

[0057] In addition, if the term "horizontal" appears, it does not mean that the component is required to be absolutely horizontal, but it can be slightly inclined. For example, "horizontal" only means that its direction is more horizontal relative to "vertical", and does not mean that the structure must be completely horizontal, but it can be slightly inclined.

[0058] In the description of the embodiments of the present invention, it should also be noted that unless otherwise clearly specified and limited, if the terms "set", "install", "connect", and "link" are used, they should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific situations.

[0059] The following further describes the present invention in detail with reference to the accompanying drawings:

[0060] See Figure 1 , the present invention discloses a non-contact sleep staging method based on a lightweight network, including four parts: data feature extraction and fusion, integrated learning data preprocessing, integrated algorithm selection and model training, and model evaluation.

[0061] Specifically, it includes the following steps:

[0062] Step 1: Obtain multi-dimensional data of human sleep

[0063] In this solution, a millimeter-wave radar is first used to perform non-contact measurement on the human body, and the PSG data measured synchronously is combined to label the data, obtaining a series of non-contact data with labels.

[0064] A total of 4952 sample data are collected in the experiment. Each sample data has five dimensions. Among them, the 0th dimension and the 1st dimension are the time series data of respiration and heart rate respectively, the 2nd dimension and the 3rd dimension are the amplitude spectra of respiration and heart rate respectively, and the 4th dimension is the distance spectrum data, where the horizontal axis represents the distance position and the vertical axis represents the radar echo amplitude. The five-dimensional data is described as follows Figure 2 shown.

[0065] Step 2: For data in multiple dimensions, respectively adopt two methods of data domain feature extraction and deep learning feature extraction to achieve feature capture in time, frequency, and distance.

[0066] Step 2.1: Data domain feature extraction

[0067] The way of data domain feature extraction includes time domain features, frequency domain features, phase features, non-linear features (mainly multiple entropy value characteristics), linkage features of respiration and heart rate, and wavelet transform statistical features of respiration and heart rate, specifically as Figure 3 shown.

[0068] Step 2.1.1: Time domain features

[0069] Assume that a time series of respiratory or heart rate signals is \(x(n)\), where \(1\leq n\leq N\), and \(N = 512\) in the experiment, and calculate its mean \(\mu\). x :

[0070]

[0071] Step 2.1.2: Frequency-domain features

[0072] Perform K-point FFT transformation (K = 2048 in the experiment) on the time series of respiration and heart rate respectively, with a sampling rate of 20 Hz, and take the magnitude spectrum of the positive frequencies, that is, half of the K points as the magnitude spectrum data \(Y(f\) i )), where \(f\) i represents the \(i\)-th frequency point, \(1\leq i\leq M\).

[0073] Step 2.1.2.1: Mean of power spectrum

[0074] Between the 0 - 3 Hz frequency band data of respiration and heart rate, divide several frequency bands at intervals of 0.2 Hz, and extract the mean \(P\) i ) of the power spectral density PSD(\(f\) i ) = |Y(\(f\) 2 )| mean , and the calculation formula is as follows:

[0075]

[0076] where \(D\) j represents the total number of frequency points in the \(j\)-th frequency band, \(1\leq j\leq 15\).

[0077] Step 2.1.2.2: Dominant frequency

[0078] The dominant frequency represents the frequency corresponding to the maximum value of the power spectral density, reflecting the frequency component where the signal energy is most concentrated:

[0079]

[0080] Step 2.1.2.3: Frequency centroid

[0081] The frequency centroid mainly represents the weighted average frequency of the power spectral density, characterizing the concentrated position of the signal energy in the frequency domain:

[0082]

[0083] Step 2.1.2.4: Root mean square frequency

[0084] The root mean square frequency mainly represents the square root of the second moment of the frequency domain energy, describing the degree of dispersion of the signal frequency distribution:

[0085]

[0086] Step 2.1.2.6: Frequency kurtosis

[0087] It mainly represents the sharpness of the power spectral density distribution and reflects the concentration and pulse characteristics of the signal energy in the frequency domain:

[0088]

[0089] Among them, respectively represent the mean and standard deviation of the spectrum Y(f i ).

[0090] Step 2.1.2.7: Frequency bandwidth

[0091] This characteristic is based on the effective frequency range of the energy accumulation ratio and reflects the frequency band width of the main energy distribution of the signal. First, the power spectral density is normalized to a probability distribution:

[0092]

[0093] Then, the cumulative power probability is calculated:

[0094]

[0095] Set the bandwidth threshold α. In the experiment, α = 0.5 is adopted, representing 50% of the energy. Find the minimum frequency f k that satisfies CPSD(f BW ):

[0096] f BW = min{f k | CPSD(f k ) ≥ α} (9)

[0097] Step 2.1.3: Phase characteristics

[0098] Perform the Hilbert transform on the respiratory and heart rate time-domain data x(n) respectively to obtain the analytic signal H(n):

[0099] H(n) = x(n) + jH{x(n)} (10)

[0100] Among them, H represents the Hilbert transform operator, and the instantaneous phase of the analytic signal is extracted:

[0101]

[0102] Step 2.1.3.1 Phase mean

[0103] The phase mean mainly reflects the overall offset of the phase, and the expression is as follows:

[0104]

[0105] Step 2.1.3.2 Phase standard deviation

[0106] The phase standard deviation quantifies the degree of phase fluctuation. A high standard deviation indicates unstable phase, which can reflect the conditions of different sleep stages. The expression is as follows:

[0107] Step 2.1.3.3 Average phase change rate

[0108]

[0109] The average phase change rate characterizes the average rate of phase change over time and is related to the signal frequency stability:

[0110]

[0111] Step 2.1.4: Nonlinear features

[0112] Extract the detrended fluctuation scaling exponent, power spectral entropy, singular spectral entropy, wavelet transform energy entropy, differential entropy, sample entropy, fuzzy entropy, and permutation entropy of the respiration and heartbeat signals as the nonlinear features of the respiration and heartbeat signals.

[0113] Step 2.1.4.1: Detrended fluctuation scaling exponent DFA

[0114] Detrended fluctuation analysis (DFA) is a method for judging the statistical autocorrelation properties of a signal, which can effectively detect and analyze the non-stationary characteristics of the signal, and can effectively distinguish random noise from faulty signals with memory. The specific steps of the DFA algorithm are as follows:

[0115] Calculate the cumulative difference y(n) of the signal sequence:

[0116]

[0117] Divide y(n) into non-overlapping equal-length subintervals with a subinterval window length of s, where s ∈ [s min , s max . In the experiment, take s min = 16, s max = N / 4. Select a certain window length s and use the least squares method to fit the local trend y fit (n) of each window interval, and calculate the detrended residual fluctuation:

[0118]

[0119] Change the window length s within the interval range and repeat the above operations to obtain the residual fluctuation F = [F(smin ), F(s1), …, F(s max ). According to the calculated parameters, the linear relationship between log(F) and log(s) is fitted by the least squares method, and a curve is plotted. The slope of the curve is the detrended fluctuation scaling exponent DFA of the signal, which represents the correlation between the respiratory signal sequence and the heartbeat signal sequence in the experiment.

[0120] Step 2.1.4.2: Power spectral entropy

[0121] Power spectral entropy is used to represent the uncertainty of signal energy under the power spectrum division. When the frequency composition in the signal is simple, the power spectrum is concentrated in some frequency components, and the corresponding frequency spectral lines will be fewer. The corresponding component probabilities will be less, resulting in a smaller value of the obtained power spectral entropy. On the contrary, if the signal is more complex, the corresponding power spectrum of the signal is more dispersed, the corresponding power spectral lines will increase, and the obtained power spectral entropy value will become larger. Therefore, power spectral entropy is a quantitative description of the complexity of the energy distribution of the signal in the frequency domain. The closer the signal is to white noise, the more dispersed the power spectrum and the larger the power spectral entropy. Combining this energy distribution with information entropy can realize the quantitative description of the complexity of the signal in the frequency domain space. Based on the information entropy formula, the power spectral entropy H1 is defined as:

[0122]

[0123] where p(k) represents the probability distribution of the power spectrum, and p(k) = PSD norm (f k ), 1 ≤ k ≤ M.

[0124] Step 2.1.4.3: Singular spectrum entropy

[0125] Singular Spectrum Entropy (SSE) is a signal complexity quantification method based on Singular Value Decomposition (SVD). It characterizes the randomness or orderliness of the signal by analyzing the singular value distribution of the signal trajectory matrix, and it is applicable to the application scenarios with fewer sampling points and noise in the collected signals.

[0126] The signal x(n) is converted into a trajectory matrix (Hankel matrix) to represent the time-domain structure of the signal. The window length L is selected, usually 2 ≤ L ≤ N / 2, and L = 65 is taken in the experiment. A matrix X of L × K is constructed, where K = N - L + 1:

[0127]

[0128] Each row of this matrix is a delayed version of the signal, which is used to capture the time-domain correlation of the signal.

[0129] Perform singular value decomposition on the trajectory matrix X:

[0130] X = UΣV T (19)

[0131] Where:

[0132] 1) U is an L×L orthogonal matrix (left singular vector);

[0133] 2) Σ is an L×K diagonal matrix, and the diagonal elements are singular values σ1≥σ2≥…≥σ r ≥0 (r = min(L, K));

[0134] 3) V is a K×K orthogonal matrix (right singular vector).

[0135] The square of the singular value represents the proportion of the energy of the corresponding component in the signal. Normalize the singular values to a probability distribution:

[0136]

[0137] Based on the information entropy formula, define the singular spectrum entropy H2 as:

[0138]

[0139] Through the singular spectrum entropy, the complexity of the signal in the time domain can be quantified, and it is widely used in fields such as biomedical signal analysis and non-stationary signal processing.

[0140] Step 2.1.4.4: Wavelet transform energy entropy

[0141] The energy entropy distributes the energy of the respiration and heart rate signals to different frequency bands through wavelet decomposition, and quantifies the complexity of its energy distribution based on the Shannon entropy. First, perform L-layer wavelet decomposition on the signal x(n) to obtain a set of coefficients. In the experiment, the db4 wavelet basis is selected for L = 5-layer decomposition:

[0142] coeffs = [cA L ,cD L ,cD L-1 ,…,cD1] (22)

[0143] Where, cA L is the approximation coefficient of the Lth layer in the low-frequency part, and cD k is the detail coefficient of the kth layer in the high-frequency part (k = 1, 2,…, L).

[0144] The energy of each sub-band is the sum of the squares of its coefficients:

[0145]

[0146] The total energy is the sum of the energies of all sub-bands:

[0147]

[0148] Normalize the sub-band energy into a probability distribution:

[0149]

[0150] Therefore, the energy entropy H3 is defined as:

[0151]

[0152] Step 2.1.4.5: Differential entropy

[0153] Differential entropy is an extension of the entropy of a continuous random variable and is defined as:

[0154]

[0155] In the experiment, the Gaussian kernel and Silverman's rule are used to estimate the probability density function. The higher the differential entropy value, the closer the signal is to a uniform distribution (such as white noise); the lower the value, the more concentrated the signal is (such as a sine wave). Therefore, this feature is used to evaluate the randomness characteristics of respiratory and heart rate signals.

[0156] Step 2.1.4.6: Sample entropy

[0157] Sample entropy measures the complexity of a time series. For a sequence x(n) of length N, count the number of matches B m [i] - x m [j]|| < r, and for a sequence x(m) of length N + 1, count the number of matches A n (r). In the experiment m (r), where is the standard deviation of the sample sequence x(n). Define the sample entropy SampEn as:

[0158]

[0159] The lower the sample entropy value, the more regular the sequence; the higher the value, the more complex the sequence, which is applicable to short-term physiological signals such as heart rate variability analysis.

[0160] Step 2.1.4.7: Fuzzy entropy

[0161] Fuzzy entropy introduces a fuzzy membership function to measure similarity. First, de-mean the sequence x(n), and then use an exponential function to calculate the similarity S between sequences at different times:

[0162]

[0163] Among them, r = 0.2, and the calculation formula for the fuzzy entropy FuzzyEn is defined as:

[0164] FuzzyEn = lnφ n (r) - lnφ n+1 (r) (30)

[0165] Among them, It smooths the boundary through the membership function and is more robust in dealing with the amplitude fluctuations of respiratory and heartbeat signals.

[0166] Step 2.1.4.8: Permutation entropy

[0167] Permutation entropy measures the signal complexity through the permutation pattern of symbolized time. The sequence is segmented into windows to generate a certain permutation pattern π (such as ascending or descending order), and then the probability p(π) of each permutation pattern is statistically calculated. Then, the permutation entropy PermEn is defined as:

[0168]

[0169] Among them, m is the permutation dimension.

[0170] Step 2.1.5: Respiratory and heart rate coupling characteristics

[0171] To quantify the rhythmic coupling characteristics of respiratory and heart rate signals, multi-dimensional coupling characteristics are proposed to facilitate obtaining the correlation between vital sign signals in different sleep stages. For the respiratory signal x bre (n) and the heart rate signal x hea (n) in a certain sample, the coupling characteristics include phase difference statistics, phase locking value, frequency domain coherence, and cross-correlation time delay characteristics.

[0172] Step 2.1.5.1: Mean phase difference

[0173] Through the Hilbert transform mentioned above, the instantaneous phases φ bre (n) and φ hea (n) of the respiratory and heart rate signals are obtained respectively, and the phase difference between the two signals is calculated:

[0174] Δφ(n) = φ bre (n) - φ hea (n) (32)

[0175] The mean phase difference is obtained:

[0176]

[0177] This feature can reflect the overall offset of the phases of the two signals.

[0178] Step 2.1.5.2: Standard deviation of phase difference

[0179] The standard deviation can further quantify the phase synchronization stability:

[0180]

[0181] Step 2.1.5.3: Phase Locking Value

[0182] When analyzing electroencephalogram (EEG) data, the Phase Locking Value (PLV) is used to evaluate the synchronous activity between different brain regions. In this experiment, the phase locking value is used to analyze the synchronization of respiratory and heart rate signals:

[0183]

[0184] Step 2.1.5.4: Frequency Domain Coherence

[0185] This feature tests the strength of the linear correlation between two signals in the frequency domain to detect specific frequency resonances. First, calculate the power spectral density P xx and P yy , and the cross-spectral density P xy :

[0186]

[0187] where PSD bre (f i ) represents the power spectral density of the respiratory signal, PSD hea (f i ) represents the power spectral density of the heart rate signal, Y bre * (f i ) represents the complex conjugate of the FFT transform of the respiratory time-domain signal, Y hea (f i ) represents the FFT transform of the heart rate signal, 1 ≤ i ≤ M.

[0188] Calculate the average coherence coefficient in the frequency domain:

[0189]

[0190] Step 2.1.5.5: Cross-Correlation Time Delay Characteristics

[0191] Cross-correlation delay estimation calculates the time delay between two signals by comparing the similarity of the respiratory and heart rate signals. From a physiological perspective, heart rate changes lag behind respiration. When apnea occurs, situations such as extended time delay and weakened correlation may occur. The definition of the cross-correlation function is as follows:

[0192]

[0193] The maximum value of the cross-correlation function corresponds to the optimal time delay between two signals, which is the time-delay estimation feature of the respiration and heart rate signals.

[0194]

[0195] Step 2.1.6. Statistical features of wavelet transform

[0196] Regarding the time-varying spectral characteristics of the respiration and heart rate signals, wavelet transform of formula (22) is adopted, and 5-layer multi-scale decomposition is performed using the db4 wavelet basis to extract 7 statistics such as the mean, standard deviation, and energy of each sub-band coefficient.

[0197] Table 1 Statistical quantities of wavelet transform sub-bands

[0198]

[0199]

[0200] According to Table 1, there are 7 statistics for each sub-band, a total of 6 sub-bands, and 2 types of signals, so a total of 84 statistical features of wavelet transform are extracted.

[0201] Combining the above 153 features extracted in the data domain, which include multiple aspects such as the time domain, frequency domain, and non-linear domain of respiration and heart rate, comprehensively excavates the hidden physiological information in respiration and heart rate data, providing a physical basis for the subsequent stage staging.

[0202] Table 2 Explanation of feature extraction data

[0203]

[0204]

[0205] Step 2.2 Automatic extraction of neural network features

[0206] Based on the features extracted from the original data in the data domain, this study further introduces a neural network for automatic feature capture to enhance the data representation ability and mine the deep information about the range spectrum in radar data to achieve information extension. Although the traditional flattening operation can adapt multi-dimensional data to the model input, it may lose local correlations. Therefore, this embodiment designs a heterogeneous network based on the MVNet architecture, and its multi-branch 1D convolutional structure can capture local features of different dimensions respectively, and generate a high-dimensional abstract representation through adaptive pooling and feature fusion.

[0207] The dimension of the original dataset is (4952, 5, 512). Adopting the logic of grouped input of the heterogeneous graph network, the five dimensions of the respiration waveform, heartbeat waveform, respiration frequency, heartbeat frequency, and range spectrum are reserved with a three-dimensional structure and input into independent branch networks respectively. The specific steps of its branch structure are as follows, for details seeFigure 4 :

[0208] (1) The 1D convolutional layer (convolution kernel = 3, stride = 2, padding = 2) extracts local temporal features;

[0209] (2) Batch normalization and ReLU activation function enhance the non - linear expression ability;

[0210] (3) Adaptive average pooling compresses the temporal dimension to a fixed length (16), and the output dimension is (4, 16)

[0211] (4) Feature fusion: The outputs of 5 branches are concatenated along the channel dimension to form a fused feature vector (dimension: 4×16×5 = 320), which is used as the abstract feature extracted by the neural network;

[0212] (5) Feature fusion and standardization: The 320 - dimensional features extracted by the neural network are concatenated with the 153 - dimensional features extracted from the data to form an enhanced feature vector with a total dimension of 473, and finally put into a machine - learning classifier to complete the staging.

[0213] This improvement enables the subsequent ensemble learning model to utilize richer feature information, improving the classification performance and generalization ability while maintaining the computational efficiency.

[0214] Step 3: Ensemble learning data pre - processing:

[0215] Feature data and label data are respectively input into the machine - learning classifier. The dimension of the feature data is (4952, 473), indicating 4952 samples, and each sample has 473 features. Among them, the first 153 features are the respiratory and heartbeat features extracted in the data domain, and the last 320 are the features automatically extracted by deep learning; the dimension of the label data is (4952, 1), where "1" represents the sleep state (positive class) and "0" represents the awake state (negative class).

[0216] To eliminate the influence of feature dimension differences on model training, standardization processing (Standard Scaler) is adopted. The features are normalized through the formula x′=(x - μ) / σ, where μ is the feature mean and σ is the feature standard deviation, ensuring that each feature is in the same scale space and improving the stability and convergence efficiency of model training.

[0217] Regarding the problem of data class imbalance (1224 samples of class "0" and 3728 samples of class "1" in the original training set), a multiple data augmentation strategy is implemented. First, based on the original data, 2504 new samples with the label "0" are generated using the SMOTE (Synthetic Minority OverSampling Technique) oversampling algorithm. By interpolating to expand the minority class samples, the skewness of the class distribution is alleviated, making the number of samples with the label "0" equal to that of samples with the label "1". Second, Gaussian noise perturbation with a mean of 0 and a standard deviation of 0.05 is applied to the original features to enhance the robustness of the model to data perturbation, generating noise data with a feature dimension of (7456, 473). During testing, 20% of the original data is maintained in the test set to ensure the stability of the evaluation.

[0218] Step 4: Selection of Ensemble Learning Algorithm and Model Training

[0219] Ensemble learning methods, as an effective strategy to improve prediction accuracy by combining multiple base learners, have shown significant advantages in various classification tasks in recent years. Especially when dealing with data with strong noise and complex features, ensemble learning can effectively enhance the robustness and generalization ability of the model. Therefore, the present invention aims to explore a lightweight radar measurement data sleep staging technique based on ensemble learning methods and provide an efficient solution by comparing three gradient boosting (Gradient Boosting Decision Tree, GBDT) machine learning algorithms, namely LightGBM, XGBoost, and CatBoost, as shown in Figure 5 .

[0220] LightGBM: LightGBM is based on the gradient boosting tree algorithm and achieves high training efficiency and low memory consumption through a tree structure with a leaf-first strategy. It can effectively handle large-scale data and avoid overfitting, enhancing the generalization ability of the model.

[0221] XGBoost: XGBoost optimizes the loss function and uses second-order gradient information, having a higher training speed and strong regularization ability. It can better capture complex non-linear relationships in the data and improve the prediction ability of the model.

[0222] CatBoost: CatBoost is designed specifically for categorical features, uses an ordered categorical encoding method to effectively avoid overfitting problems, and has strong missing value handling capabilities, adapting to different data types and data

[0223] In this embodiment, the data obtained from the above-mentioned data domain feature extraction and neural network feature extraction is used as the original data for the classifier. First, SMOTE oversampling is adopted to balance the sample labels of the dataset, and Gaussian white noise is added to enhance the data robustness. On this basis, three high-performance ensemble algorithms, namely LightGBM, XGBoost, and CatBoost, are used for sleep staging training. Finally, performance metrics such as the accuracy, F1 score, and confusion matrix of the test set are calculated for algorithm comparison. To improve the interpretability of the algorithm, SHAP is used to analyze the feature contribution degree, and the decision tree structure is visualized to trace the model decision path. The specific process is as follows Figure 6 as shown

[0224] Step 5: Model Evaluation and Performance Metrics:

[0225] (1) Evaluation Metrics

[0226] To comprehensively quantify the classification performance of different ensemble learning models in the sleep staging task of radar measurement data, this embodiment uses multiple evaluation metrics, including plotting the confusion matrix and calculating metrics such as accuracy, F1 value, recall rate, precision, and ROC-AUC.

[0227] Confusion Matrix: Intuitively shows the classification results of the model for different classes, where TP (True Positive) represents the correctly classified positive samples, TN (True Negative) represents the correctly classified negative samples, FP (False Positive) represents the negative samples misclassified as positive samples, and FN (False Negative) represents the positive samples misclassified as negative samples.

[0228] Accuracy: Measures the proportion of correct classifications of the model as a whole, and the calculation formula is:

[0229]

[0230] F1 Score: Considers the harmonic mean of Precision and Recall to balance the performance of the model on positive and negative samples, and the calculation formula is:

[0231]

[0232] Recall: Measures the ability of the model to identify all positive samples, and the calculation formula is:

[0233]

[0234] Precision: Represents the proportion of actual positive samples among the data classified as positive samples, and the calculation formula is:

[0235]

[0236] ROC-AUC (Area Under Curve): Measures the ability of the model to distinguish between positive and negative samples. The closer the value is to 1, the stronger the classification ability of the model.

[0237] (2) Analysis of experimental results

[0238] In this embodiment, the performances of three ensemble learning models, LightGBM, XGBoost, and CatBoost, on the original dataset are evaluated respectively, and the influence of data augmentation (SMOTE + Gaussian noise) on LightGBM is further analyzed.

[0239] a. Comparison of classification performance

[0240] In this embodiment, three ensemble learning models, LightGBM, XGBoost, and CatBoost, are compared, and a data augmentation experiment is conducted on LightGBM. Figure 7 Shows the loss curves of LightGBM for the original data and the augmented data on the training set and the validation set.

[0241] Through Figure 7 As can be seen from Figure a, for the original data, the training loss and the validation loss gradually decrease with the increase of the training rounds and finally tend to converge. This indicates that the model can effectively learn the data features and there is no obvious overfitting phenomenon.

[0242] Through Figure 7 As can be obtained from Figure b, compared with the original data, the overall downward trend of the loss curve after data augmentation is similar, but the oscillation is more obvious. This may be due to the introduction of synthetic samples, increasing the data diversity, thus leading to fluctuations in the adaptability of the model in the early training stage. In the later stage of training, the loss curve still remains stable and convergent, indicating that the model can still effectively learn the augmented data, and this augmentation strategy has no obvious negative impact on the training stability of the model.

[0243] From the results, LightGBM performs best in terms of indicators such as Accuracy (0.9465), F1 Score (0.9652), and Recall (0.9853), indicating its strong classification ability in this task. Data augmentation (SMOTE + Gaussian noise) improves the Precision (0.9481) and AUC (0.9795) of LightGBM, but the Accuracy (0.9445) and Recall (0.9799) decrease slightly, indicating that while data augmentation improves the class balance, it may introduce certain noise.

[0244] The overall performance of XGBoost is close to that of LightGBM, showing good balance. Its Accuracy (0.9415) and AUC (0.9755) are slightly lower than those of LightGBM, and its Precision (0.9444) is relatively low, indicating that there may be certain misclassification problems.

[0245] CatBoost performs better in terms of Recall (0.9839), but its AUC (0.9022) is significantly lower than that of other models, which may be due to some information loss caused by its categorical feature encoding strategy.

[0246] Table 1 Model Index Table

[0247]

[0248] b. Misclassification Analysis

[0249] Table 2 Confusion Matrix Table

[0250]

[0251] From Table 4 and Figure 9 As can be seen, all models have relatively high Recall for the positive class (sleep state), indicating that they have good sensitivity in identifying the sleep state. However, for the negative class (awake state), the number of false positives (FP) is relatively high, indicating that there are still certain misjudgments when the models distinguish the awake state.

[0252] Feature Importance Analysis

[0253] To further explore the decision-making mechanisms of each model, the feature importance analysis method is used to evaluate the degree of dependence of the models on different features. Figure 10 The feature importance heatmaps of each model are shown.

[0254] As Figure 10 shown, LightGBM mainly relies on the respiratory and heartbeat features extracted from data features and some features extracted by deep learning. XGBoost shows relatively balanced feature selection, and some deep learning features have relatively high contribution degrees. CatBoost relatively pays more attention to the respiratory and heartbeat features extracted from data features.

[0255] From the decision tree of LightGBM after data augmentation ( Figure 11 b), it can be seen that the depth of the decision tree after data augmentation is slightly increased compared with that of the decision tree of the original data ( Figure 11 a), indicating that the model introduces more complex decision-making paths after increasing the data samples. The feature thresholds of some key nodes have changed, indicating that data augmentation affects the distribution characteristics of the samples and makes the model make different judgments during the decision-making process.

[0256] (2) Result Analysis

[0257] Model performance comparison: Due to its efficient leaf-first strategy, LightGBM achieves the best performance among all models. XGBoost demonstrates strong generalization ability through second-order gradient optimization, while CatBoost has unique advantages in handling categorical features but shows inferior AUC-ROC performance compared to the other two algorithms in this task.

[0258] Impact of data augmentation: Although SMOTE oversampling improves class balance, its impact on the Recall of LightGBM is small, and the Precision increases slightly, indicating that oversampling methods need to be combined with more refined sample generation strategies.

[0259] Robustness analysis: As can be seen from the confusion matrix, all models have relatively high Recall for the positive class (sleep state), indicating that they have good sensitivity in detecting the sleep state. However, for the negative class (awake state), the number of FPs is relatively high, suggesting that there is still room for improvement in the model's misjudgment of the awake state.

[0260] Generally speaking, LightGBM performs the best in this task and is suitable for the sleep staging task of millimeter-wave radar measurement data; XGBoost performs stably and can be used as an alternative; CatBoost is suitable for data containing more categorical variables, but its AUC in this task still needs to be optimized. Therefore, LightGBM is finally selected as the optimal classifier in this patent.

[0261] The non-contact sleep signal extraction and staging system based on millimeter-wave radar of the present invention has the following significant advantages compared with traditional contact monitoring technologies:

[0262] (1) Non-contact monitoring, improving comfort and applicability: Traditional contact devices (such as PSG) need to attach electrodes to the body surface, which can easily cause discomfort, especially interfering with infants, the elderly, and those with limited mobility, and even affecting sleep quality. This system uses millimeter-wave radar to emit high-frequency signals and captures physiological parameters such as respiration and heartbeat in real time by detecting minute chest wall displacements without any physical contact, completely eliminating the wearing burden. Its penetration detection ability can penetrate clothing and bedding, adapt to complex environments, and the device is highly integrated and can be concealedly installed in home or medical scenarios to achieve non-intrusive monitoring, significantly improving the user experience.

[0263] (2) Comprehensive privacy and security protection: Compared with non-contact solutions such as video and infrared that may involve image acquisition or environmental restrictions, millimeter-wave radar only analyzes physiological signals through radio waves, does not record images or sounds, completely avoiding the risk of privacy leakage, ensuring the security of user information, and is especially suitable for privacy-sensitive scenarios such as homes and nursing homes.

[0264] (3) Millimeter-wave radar penetration monitoring technology: A non-contact signal acquisition scheme based on a continuous wave frequency modulation millimeter-wave radar in the 77 - 80.5 GHz band is proposed to achieve precise capture of chest wall displacement. This technology breaks through the comfort limit of traditional contact electrodes, can penetrate clothing and bedding, and improves the user experience while maintaining monitoring accuracy, especially suitable for special groups such as infants and the elderly.

[0265] (4) Multimodal physiological signal coupling analysis: A respiratory and heart rate linkage feature system including mean phase difference, phase locking value, frequency domain coherence, and cross-correlation time delay characteristics is proposed to quantify the rhythmic coupling relationship of vital signs during the sleep stage, providing a new discrimination dimension for sleep staging.

[0266] (5) Hybrid feature fusion and heterogeneous network architecture: A composite feature space including 153-dimensional data domain features (covering time domain, frequency domain, non-linear, and respiratory and heart rate linkage features) and 320-dimensional features extracted by deep learning is constructed. The heterogeneous network designed based on MVNet effectively retains local temporal correlation and improves feature abstraction ability.

[0267] (6) Lightweight integrated learning model optimization: By comparing three integrated learning models, the system uses LightGBM to construct a lightweight gradient boosting mechanism to build a sleep staging model, achieving relatively accurate and fast stage division for the signal data extracted by the millimeter-wave radar. Combining a hybrid data augmentation strategy of SMOTE oversampling and Gaussian noise injection to alleviate the class imbalance problem.

[0268] (7) Feature interpretability analysis framework: Using the method of SHAP value contribution analysis and visualizing decision trees, the influence weight of each feature on the sleep staging result is quantified. Through this method, clinicians can intuitively identify key physiological indicators (such as standard deviation of respiratory rate, detrended fluctuation scaling exponent, phase locking value), and trace the model decision path, improving the credibility and practicality of the algorithm in medical scenarios.

[0269] This embodiment also discloses a non-contact sleep staging system based on a lightweight network, including:

[0270] An original data acquisition module for acquiring multi-dimensional data of human sleep;

[0271] A feature extraction module for extracting features from time domain features, frequency domain features, phase features, non-linear features, linkage features, and wavelet transform statistical features based on multi-dimensional data, fusing the extracted features to obtain data domain features, and extracting local features of different dimensions in the multi-dimensional data to form a fused feature vector;

[0272] A feature fusion module, configured to fuse and splice local features and data domain features to obtain an enhanced feature vector;

[0273] A staging module, configured to classify the enhanced feature vector to complete sleep staging.

[0274] A schematic diagram of a terminal device provided by an embodiment of the present invention. The terminal device in this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps in the above-mentioned method embodiments are implemented. Alternatively, when the processor executes the computer program, the functions of each module / unit in the above-mentioned device embodiments are implemented.

[0275] The computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory and executed by the processor to complete the present invention.

[0276] The terminal device may be a computing device such as a desktop computer, a notebook, a palm computer, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory.

[0277] The processor may be a central processing unit (CPU), or may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0278] The memory may be used to store the computer program and / or module, and the processor realizes various functions of the terminal device by running or executing the computer program and / or module stored in the memory, and by calling the data stored in the memory.

[0279] If the modules / units integrated in the terminal device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by a computer program instructing the relevant hardware. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, the steps of the above-described various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals.

[0280] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A non-contact sleep staging method based on a lightweight network, characterized in that It includes the following steps: Obtain multi-dimensional data of human sleep; Based on the multi-dimensional data, extract time-domain features, frequency-domain features, phase features, non-linear features, linkage features, and wavelet transform statistical features, and fuse the extracted features to obtain data-domain features; Extract local features of different dimensions in the multi-dimensional data to form a fused feature vector; splice the fused feature vector and the data-domain features to obtain an enhanced feature vector; Classify the enhanced feature vector to complete sleep staging.

2. The non-contact sleep staging method based on a lightweight network according to claim 1, wherein The obtaining of the multi-dimensional data of human sleep includes: Perform non-contact measurement on the human body through a millimeter-wave radar, and combine the PSG data measured synchronously to label the data, obtaining non-contact data with labels; In the non-contact data, each sample data includes five dimensions. The 0th dimension and the 1st dimension are respectively the time-series data of respiration and heart rate, the 2nd dimension and the 3rd dimension are respectively the amplitude spectra of respiration and heart rate, and the 4th dimension is the range spectrum data.

3. A non-contact sleep staging method based on a lightweight network according to claim 1, characterized in that, When extracting the non-linear features, it includes: Extract the detrended fluctuation scaling exponent, power spectral entropy, singular spectral entropy, wavelet transform energy entropy, differential entropy, sample entropy, fuzzy entropy, and permutation entropy of the respiration and heartbeat signals; When extracting the linkage features, it includes: Extract the correlation features between the respiration and heartbeat signals. The correlation features include phase difference statistics, phase-locking value, frequency-domain coherence, and cross-correlation time-delay characteristics.

4. A non-contact sleep staging method based on a lightweight network according to claim 1, characterized in that The step of extracting local features of different dimensions in the multi-dimensional data specifically includes: constructing a deep learning heterogeneous network, and extracting local features of different dimensions in the multi-dimensional data based on the deep learning heterogeneous network.

5. A non-contact sleep staging method based on a lightweight network according to claim 4, characterized in that, The deep learning heterogeneous network includes a 1D convolutional layer, batch normalization and ReLU activation function, adaptive average pooling, and a feature fusion module.

6. A non-contact sleep staging method based on a lightweight network according to claim 1, characterized in that, The data-domain features include 153-dimensional features extracted from the data domain, and the local features of different dimensions include 320-dimensional features extracted by the neural network. Splice and fuse the 153-dimensional features and the 320-dimensional features to obtain an enhanced feature vector with a total dimension of 473.

7. A non-contact sleep staging method based on a lightweight network according to claim 1, characterized in that, The step of classifying the enhanced feature vector includes: Train the LightGBM model, and classify the enhanced feature vector through the trained LightGBM model.

8. A non-contact sleep staging system based on a lightweight network, characterized in that, It includes: An original data acquisition module for obtaining multi-dimensional data of human sleep; A feature extraction module for extracting features from time-domain features, frequency-domain features, phase features, non-linear features, linkage features, and wavelet transform statistical features based on the multi-dimensional data, and fusing the extracted features to obtain data-domain features, and extracting local features of different dimensions in the multi-dimensional data to form a fused feature vector; A feature fusion module for fusing and splicing local features and data-domain features to obtain an enhanced feature vector; A staging module for classifying the enhanced feature vector to complete sleep staging.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1-7.

Images