Sleep disorder prediction method and system based on deep learning and signal decomposition

Abstract

The invention discloses a sleep disorder prediction method and a sleep disorder prediction system based on deep learning and signal decomposition, the system adaptively determines parameters of variational mode decomposition (VMD) through an RIME optimization algorithm, and decomposes s-EEG signals into intrinsic mode functions (IMFs); the IMFs are dynamically weighted by using a multi-head self-attention mechanism (MHSA), and key features are extracted, and then the IMFs are subjected to dynamic weighting by using the multi-head self-attention mechanism (MHSA), and the IMFs are subjected to dynamic weighting by using the multi-head self-attention mechanism (MHSA); and finally, time sequence modeling is carried out through a long and short time memory network (LSTM), and high-precision prediction is realized. According to the method, the accuracy and stability of sleep disorder prediction are remarkably improved, technical support is provided for early diagnosis and intervention, and clinical application value is achieved.

Description

Technical Field

[0001] The present invention belongs to the cross - field of biomedical engineering and artificial intelligence, and focuses on methods and systems for predicting sleep disorders using s - EEG signals. Background Art

[0002] Sleep disorders are a class of diseases that seriously affect an individual's health and quality of life, covering symptoms such as difficulty falling asleep, sleep maintenance disorders, sleep rhythm disorders, and excessive daytime sleepiness. These symptoms not only lead to short - term decline in cognitive function, mood swings, and memory loss, but may also trigger long - term health problems such as cardiovascular diseases, metabolic syndrome, and mental health disorders. Therefore, early detection and intervention of sleep disorders have important clinical significance. As the core biomarker for sleep disorder prediction, s - EEG signals can achieve sleep staging, quality assessment, and abnormal pattern detection by analyzing their characteristics. However, the non - linear and non - stationary nature of s - EEG signals, as well as their susceptibility to noise interference, pose significant challenges to high - precision sleep disorder prediction.

[0003] Traditional s - EEG signal analysis methods rely on manually extracting time - domain, frequency - domain, and time - frequency - domain features and combining classical machine learning algorithms for sleep state assessment. However, these methods have significant limitations in capturing non - linear features, processing non - stationary signals, and high - dimensional data, with low feature extraction efficiency and difficulty in adaptive adjustment. In recent years, the introduction of deep learning technology has provided a new direction for sleep disorder prediction. In particular, LSTM has been widely used due to its advantages in time - series modeling. The LSTM model combined with the attention mechanism further improves the ability to capture the dynamic features of s - EEG signals. However, existing deep learning methods still have deficiencies in signal decomposition and parameter optimization. Directly analyzing s - EEG signals may lead to frequency band aliasing, affecting the prediction accuracy.

[0004] As an advanced signal decomposition technology, VMD can decompose s - EEG signals into IMFs of different frequencies, thus facilitating the extraction of features in each frequency band to improve the prediction effect. Although VMD performs excellently in spectral analysis, its number of modes and penalty parameters usually rely on empirical selection, resulting in limited decomposition accuracy and stability. To solve this problem, researchers have tried to use meta - heuristic optimization algorithms (such as particle swarm optimization, genetic algorithm) to optimize the VMD parameters to improve the decomposition quality. However, these methods are still insufficient in global parameter search and adaptability to complex signals, often falling into local optima and difficult to fully explore the deep information of s - EEG signals.

[0005] The disadvantages of the existing technology are summarized as follows:

[0006] (1) Traditional s-EEG signal analysis methods rely on manual feature extraction and classical machine learning, making it difficult to effectively capture the non-linear and non-stationary characteristics of signals, resulting in low feature extraction efficiency and limited prediction accuracy.

[0007] (2) Existing VMD decomposition methods rely on experience in parameter selection. The non-self-adaptability of the number of modes and penalty parameters reduces the accuracy and stability of signal decomposition, and is prone to frequency band aliasing.

[0008] (3) Existing meta-heuristic optimization algorithms lack global search ability in VMD parameter optimization, are prone to falling into local optima, and are strongly dependent on initial parameters, which may lead to poor decomposition effects and thus affect the performance of subsequent sleep disorder prediction.

[0009] (4) Existing sleep disorder prediction systems lack the ability of collaborative optimization between modules, resulting in low overall efficiency of signal processing and feature analysis, and limiting the reliability of prediction results. Summary of the Invention

[0010] To overcome the limitations of s-EEG signal analysis and sleep disorder prediction in the prior art, the purpose of the present invention is to propose a sleep disorder prediction method and system based on RIME optimization and multi-headed self-attention mechanism-long short-term memory network (Multi-Headed Self-Attention Long Short-Term Memory, MHSA-LSTM). This method combines adaptive optimization of VMD parameters, dynamic weighted feature extraction and time series modeling to effectively solve the non-linear and non-stationary problems of s-EEG signals, achieve high-precision sleep disorder prediction, and is applicable to clinical diagnosis and intervention. The present invention mainly solves the following problems:

[0011] (1) Aiming at the problem that the traditional VMD decomposition parameters rely on empirical selection, resulting in low signal decomposition accuracy and frequency band aliasing, the RIME optimization algorithm is used to adaptively determine the VMD parameters to improve the decomposition quality.

[0012] (2) Aiming at the problem that existing feature extraction methods lack a dynamic weighting mechanism and cannot highlight key features, MHSA is introduced to dynamically weight IMFs to enhance the feature extraction ability.

[0013] (3) Aiming at the problem that existing deep learning models are difficult to capture the time series dependence relationship of s-EEG signals, resulting in limited classification accuracy, LSTM is used for time series modeling to improve the prediction accuracy.

[0014] (4) Aiming at the problem that existing systems lack the ability of collaborative optimization between modules, resulting in low overall efficiency, a modular progressive system is designed to improve the collaborative efficiency of signal processing and feature analysis.

[0015] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0016] A sleep disorder prediction method based on deep learning and signal decomposition, comprising the following steps:

[0017] a) Signal acquisition: Obtain s-EEG signals;

[0018] b) Parameter optimization: Use the RIME optimization algorithm to adaptively determine the number of modes K and the penalty parameter α of VMD;

[0019] c) Signal decomposition: Perform VMD decomposition on the s-EEG signals using the determined K and α in step b) to obtain a series of IMFs;

[0020] d) Feature extraction: Extract and screen features from the decomposed IMFs to obtain a feature matrix;

[0021] e) Feature weighting: Use MHSA to perform weighting processing on the feature matrix to obtain weighted features;

[0022] f) Temporal prediction: Input the weighted features into LSTM, and predict sleep disorders through temporal modeling;

[0023] g) Result output: Generate prediction information of sleep disorders according to the output result of LSTM, including sleep stage classification and anomaly detection.

[0024] Further, the signal acquisition in step a) includes: using a multi-channel electroencephalogram device to collect s-EEG signals, and the channels include F3-A2, F4-A1, C3-A2, C4-A1, O1-A2, O2-A1; preprocessing the collected signals, including wavelet threshold denoising and 4 - 40 Hz band-pass filtering, to remove noise and artifacts.

[0025] Further, the parameter optimization in step b) specifically includes: simulating the dynamic behavior of a frost particle population through the RIME optimization algorithm to search for the optimal number of modes K and the penalty parameter α of VMD;

[0026] The optimization objective is to minimize the weighted sum of the reconstruction error (RE) and the spectral separation degree (FS), defined as:

[0027] Loss = λ 1 ·RE + λ 2 ·FS

[0028] Where, represents the reconstruction error, FS measures the frequency discrimination between modes, and λ 1 , λ 2 are adjustable weight coefficients;

[0029] Update the positions of frost particles iteratively until the objective function converges or the maximum number of iterations is reached, and output the optimal parameters K and α.

[0030] Further, the signal decomposition in step c) is specifically as follows: Based on the optimized K and α in step b), the s-EEG signal is decomposed into a series of IMFs by VMD, and the mathematical expression is:

[0031]

[0032] where x(t) is the s-EEG signal, and u k (t) is the k-th intrinsic mode function;

[0033] By optimizing the central frequency and spectral width of the modes, the frequency separability of the IMFs is ensured. The central frequency update formula is:

[0034]

[0035] where U k (f) is the Fourier transform of u k (t), and ω k is the central frequency of the k-th mode.

[0036] Further, the feature extraction in step d) specifically includes:

[0037] ⑤ Time-domain features: Calculate the root mean square (RMS) and zero crossing rate (ZCR). The RMS formula is:

[0038]

[0039] ⑥ Frequency-domain features: Calculate the band energy, and the formula is:

[0040]

[0041] where f 1 and f 2 are the target frequency band ranges;

[0042] ⑦ Time-frequency domain features: Calculate the energy by short-time Fourier transform (STFT), and the formula is:

[0043]

[0044] where ω(τ - t) is the window function;

[0045] ⑧ Feature screening: Screen the IMFs based on the frequency criteria for sleep staging. The specific conditions are:

[0046] WAKE stage: The dominant frequency f dom ≥ 8Hz, and the zero crossing rate (ZCR) > 0.3;

[0047] REM phase: 4 Hz ≤ f dom ≤ 13 Hz;

[0048] NREM phase: 0.5 Hz ≤ f dom ≤ 7 Hz;

[0049] Calculate the Pearson correlation coefficient between the IMFs and the original signal, and retain the IMFs with a correlation coefficient higher than the threshold.

[0050] Furthermore, the feature weighting in step e) is specifically as follows: Use MHSA to calculate the feature weights in parallel through multiple attention heads

[0051] The attention score of each head is calculated as:

[0052]

[0053] where Q i , K j are the i-th query vector and the j-th key vector respectively, N is the number of input features, and d k is the dimension of the key vector;

[0054] Generate weighted features through weighted summation, and the formula is:

[0055]

[0056] where V j is the value vector; Concatenate the multi-head outputs and generate the final feature representation through a linear transformation.

[0057] Furthermore, the temporal prediction in step f) is specifically as follows: Input the weighted features into the LSTM network, and the hidden state

[0058] update formula is:

[0059] h t = o t · tanh(c t )

[0060] where o t is the output gate, and c t is the cell state;

[0061] The LSTM network uses the ReLU activation function, which is defined as:

[0062] σ(x) = max(0, x)

[0063] Map the LSTM output to the sleep stage classification through a fully connected layer, and output the probabilities of the NREM, REM, and WAKE stages.

[0064] The present invention also proposes a sleep disorder prediction system based on deep learning and signal decomposition, including:

[0065] a) Signal acquisition module: providing a clean s-EEG signal, serving as the starting point of the entire system and laying a foundation for parameter optimization and signal decomposition;

[0066] b) Parameter optimization module: generating the optimal parameters of VMD, directly supporting the accurate operation of the signal decomposition module;

[0067] c) Signal decomposition module: outputting IMFs and providing the decomposed signal components for the feature extraction module;

[0068] d) Feature extraction module: generating a feature matrix and providing key feature data for the feature weighting module;

[0069] e) Feature weighting module: optimizing the feature representation and providing more targeted inputs for the time series prediction module;

[0070] f) Time series prediction module: completing the sleep disorder prediction and providing the core basis for the result output module;

[0071] g) Result output module: generating the final diagnosis information and completing the system function.

[0072] Furthermore, step a) the signal acquisition module is configured to collect and preprocess the s-EEG signal through a multi-channel electroencephalogram device, generate a high-quality clean signal, directly provide input data for the parameter optimization module and the signal decomposition module, and support the accuracy of subsequent signal analysis.

[0073] Furthermore, step b) the parameter optimization module is configured to automatically search for the optimal parameters of VMD through a processor, including the number of modes and the penalty parameter, to ensure that the signal decomposition module can efficiently and accurately decompose the s-EEG signal into IMFs.

[0074] Furthermore, step c) the signal decomposition module is configured to decompose the input s-EEG signal using the parameters provided by the parameter optimization module, generate IMFs with distinct frequencies, and serve as the signal basis for the feature extraction module.

[0075] Furthermore, step d) the feature extraction module is configured to extract multi-dimensional features from the IMFs output by the signal decomposition module and generate a structured feature matrix, providing clear and key data support for the feature weighting module.

[0076] Furthermore, step e) the feature weighting module is configured to perform dynamic weighting processing on the feature matrix through a processor, enhance the distinguishability of the features, generate an optimized feature representation, and directly support the prediction accuracy of the time series prediction module.

[0077] Further, the timing prediction module in step f) is configured to receive the output of the feature weighting module, generate a prediction result of sleep disorder through timing analysis, and provide a reliable core basis for the result output module.

[0078] The beneficial effects of the present invention are as follows:

[0079] Through the integration of RIME optimization, multi-head self-attention mechanism (MHSA) and long short-term memory network (LSTM), the present invention significantly improves the performance of s-EEG signal analysis and sleep disorder prediction, and has the following beneficial effects:

[0080] (1) The VMD parameters are adaptively determined by the RIME optimization algorithm, which overcomes the limitation of parameter selection depending on experience in the traditional method, improves the accuracy and stability of signal decomposition, effectively reduces frequency band aliasing, and provides high-quality IMFs for subsequent feature extraction.

[0081] (2) The introduction of the MHSA dynamic weighting mechanism enhances the pertinence of feature extraction, can highlight the key features related to sleep disorder, and improves the expression ability of the model for complex non-stationary signals.

[0082] (3) Using LSTM for timing modeling improves the ability of the deep learning model to capture the timing dependence relationship of s-EEG signals, thereby improving the prediction accuracy of sleep stage classification and anomaly detection.

[0083] (4) Designing a modular progressive system optimizes the collaborative efficiency of signal processing and feature analysis, overcomes the deficiency of lack of collaboration between modules in the existing system, and improves the reliability and robustness of the overall prediction. Description of the Drawings

[0084] Figure 1 : Framework diagram of the RIME optimization algorithm;

[0085] Figure 2 : VMD decomposition framework diagram;

[0086] Figure 3 : Feature extraction-selection-reconstruction framework diagram;

[0087] Figure 4 : Multi-head self-attention (MHSA) framework diagram;

[0088] Figure 5 : LSTM structure framework diagram;

[0089] Figure 6 : Overall framework diagram of the sleep disorder prediction system. Detailed Embodiments

[0090] In the following, the present invention will be further described in conjunction with the accompanying drawings and embodiments. A sleep disorder prediction method and system based on RIME optimization and MHSA-LSTM proposed by the present invention are not limited to the examples described below, but are applicable to all sleep analysis scenarios based on s-EEG signals.

[0091] To describe the technical solution of the present invention more clearly, as a specific embodiment, the accompanying drawings are combined for detailed description. The present invention includes a signal acquisition module, an RIME parameter optimization module, a signal decomposition module, a feature extraction module, an MHSA feature weighting module, an LSTM time series prediction module, and a result output module.

[0092] Embodiment 1: Sleep disorder prediction based on single-channel s-EEG signal

[0093] ① Step 1: Define the model architecture

[0094] The model architecture includes the following parts:

[0095] VMD decomposition: Decompose the s-EEG signal into intrinsic mode functions (IMFs), and the number of modes and penalty parameters are determined by RIME optimization.

[0096] MHSA module: 4 attention heads, each head with a dimension of 32 and a hidden layer dimension of 128, for feature weighting.

[0097] LSTM module: 1 layer of LSTM, with a hidden layer dimension of 128 and 3 output classes (NREM, REM, WAKE).

[0098] Overall process: Signal input → Preprocessing → VMD decomposition → Feature extraction → MHSA weighting → LSTM prediction → Classification output.

[0099] ② Step 2: Define the optimization and threshold

[0100] Set the RIME optimization objective to minimize the signal reconstruction error and spectral separation degree, with 20 iteration times and 30 frost particle numbers. Classification performance threshold: F1 score > 0.85, AUC value > 0.90. Feature screening threshold: Pearson correlation coefficient > 0.5.

[0101] ③ Step 3: Initialize the data

[0102] Collect single-channel F3-A2 s-EEG signals (200Hz, 8 hours, approximately 5,760,000 samples). Initialize the parameters as follows:

[0103] Preprocessing: Use db4 wavelet denoising and 4 - 40Hz band-pass filtering.

[0104] Sample segmentation: One segment every 30 seconds, a total of 6000 samples.

[0105] Number of IMFs: Initially set to 8, and then optimized and adjusted by RIME.

[0106] Feature dimension: Initially, 18-dimensional features (root mean square, zero-crossing rate, band energy, etc.) are extracted, and 15 dimensions are retained after screening.

[0107] ④ Step 4: Model training

[0108] The training process is as follows:

[0109] Preprocessing: The signal undergoes wavelet denoising and band-pass filtering to generate a clean signal. For details, see the signal acquisition module in Figure 6 .

[0110] RIME optimization: Optimize the VMD parameters. The specific optimization process is as shown in Figure 1 . The number of modes obtained is 9, the penalty parameter is 1815, and the target loss is reduced by 11.85%.

[0111] VMD decomposition: Decompose the signal into 9 IMFs. The low-frequency mode captures slow-wave activity, and the medium-high frequency modes reflect REM and WAKE characteristics. The decomposition framework is as shown in Figure 2 .

[0112] Feature extraction and screening: Extract time-domain (root mean square, zero-crossing rate), frequency-domain (band energy, main frequency), and time-frequency domain (STFT energy) features from the IMFs. Based on the AASM standard (NREM: 0.5 - 7Hz, REM: 4 - 13Hz, WAKE: >8Hz), screen and retain features with high correlation to generate a feature matrix (6000×15). The specific process is as shown in Figure 3 .

[0113] MHSA weighting: Weight the features with a 4-head attention mechanism to enhance the expression of key information. The MHSA framework is as shown in Figure 4 .

[0114] LSTM prediction: Train an LSTM (hidden layer 128, ReLU activation), use the AdamW optimizer (learning rate 0.001), for 40 epochs, batch size 48. The LSTM structure is as shown in Figure 5 .

[0115] After training, the model accuracy reaches 96.2%, and the validation loss is 0.164.

[0116] ⑤ Step 5: Performance evaluation and optimization

[0117] Evaluate the model performance:

[0118] Metrics: NREM F1 score is 0.94, REM F1 score is 0.89, WAKE F1 score is 0.92; the AUC values are 0.96, 0.94, and 0.98 respectively.

[0119] Adjustment: If the F1 score is less than 0.85, increase the number of RIME iterations to 25 or the number of MHSA heads to 6 and retrain.

[0120] The final performance is stable, and the validation loss is reduced by 53% - 57% compared to the baseline model.

[0121] The specific process of the above training process is as Figure 6 shown.

[0122] Example 2: Sleep disorder prediction based on multi-channel s-EEG signals

[0123] ① Step 1: Adopt the architecture of Example 1 and expand it to 6 channels (F3 - A2, F4 - A1, C3 - A2, C4 - A1, O1 - A2, O2 - A1), and the fusion feature matrix is 6000×90.

[0124] ② Step 2: Independently run RIME optimization for each channel to obtain different parameters (e.g., for F3 - A2: the number of modes is 10, and the penalty parameter is 3197).

[0125] ③ Step 3: Preprocess the signals of each channel, perform VMD decomposition, extract features, and then fuse the feature matrix.

[0126] ④ Step 4: The training process is the same as that of Example 1, for 40 rounds, and the validation loss is 0.162.

[0127] ⑤ Step 5: Performance evaluation shows that the NREM precision is 0.94, the REM F1 score is 0.89, the WAKE recall rate is 0.96, and Cohen's kappa reaches 0.9. Compared with the baseline models (CNN, LSTM, VMD - CNN), the F1 score of the present invention is increased by 5 - 10%, the loss converges quickly, and the smoothing factor is 0.003.

[0128] Example 3: Ablation experiment

[0129] To verify the effectiveness of the module, an ablation experiment is carried out:

[0130] ① Only VMD decomposition: Fix the number of modes at 8 and the penalty parameter at 2000, without RIME optimization.

[0131] ② Only LSTM prediction: Skip MHSA and directly input the features.

[0132] ③ No feature screening: Retain all IMF features.

[0133] The results show:

[0134] ①Only VMD decomposition: NREM F1 score is 0.85, and the decomposition error is 20% higher.

[0135] ②Only LSTM prediction: REM F1 score is 0.80, and the feature expression is weak.

[0136] ③Without feature screening: overfitting, and the validation loss is 0.324.

[0137] ④The complete model has the highest F1 score (0.94) and the lowest validation loss (0.164).

Claims

1. A sleep disorder prediction method based on deep learning and signal decomposition, characterized in that: The following steps are involved: a) Signal acquisition: acquiring s-EEG signals; b) Parameter optimization: The RIME optimization algorithm is used to adaptively determine the number of VMD modes K and the penalty parameter α; c) Signal decomposition: Perform VMD decomposition on the s-EEG signal using K and α determined in step b) to obtain a series of IMFs; d) Feature extraction: Extract and screen the features of the decomposed IMFs to obtain a feature matrix; e) Feature weighting: Use MHSA to weight the feature matrix to obtain weighted features; f) Time series prediction: The weighted features are input into LSTM to predict sleep disorders through time series modeling; g) Result output: Based on the output results of LSTM, predictive information of sleep disorders is generated, including sleep stage classification and anomaly detection.

2. The sleep disorder prediction method according to claim 1, characterized in that: The signal acquisition in step a) includes: using a multi-channel EEG device to acquire s-EEG signals, the channels including F3-A2, F4-A1, C3-A2, C4-A1, O1-A2, O2-A1; preprocessing the acquired signals, including wavelet threshold denoising and 4-40 Hz bandpass filtering, to remove noise and artifacts.

3. The sleep disorder prediction method according to claim 1, characterized in that: The parameter optimization in step b) specifically includes: simulating the dynamic behavior of the frost particle group through the RIME optimization algorithm, searching for the optimal mode number K and penalty parameter α of VMD; The optimization objective is to minimize the weighted sum of reconstruction error (RE) and spectral separation (FS), which is defined as: Loss = λ1·RE+λ2·FS in, represents the reconstruction error, FS measures the frequency discrimination between modes, and λ1,λ2 are adjustable weight coefficients; The frost particle positions are updated iteratively until the objective function converges or the maximum number of iterations is reached, and the optimal parameters K and α are output.

4. The sleep disorder prediction method according to claim 1, characterized in that: The signal decomposition in step c) is specifically as follows: based on K and α optimized in step b), the s-EEG signal is decomposed into a series of IMFs by VMD, which can be mathematically expressed as: Where x(t) is the s-EEG signal, u k (t) is the kth intrinsic mode function; By optimizing the center frequency and spectrum width of the mode, the frequency separation of IMFs is ensured. The center frequency update formula is: Among them, U k (f) is u k The Fourier transform of (t), ω k is the center frequency of the kth mode.

5. The sleep disorder prediction method according to claim 1, characterized in that: The feature extraction in step d) specifically includes: ① Time domain characteristics: Calculate the root mean square (RMS) and zero crossing rate (ZCR). The RMS formula is: ② Frequency domain characteristics: Calculate the frequency band energy, the formula is: Among them, f1 and f2 are the target frequency band ranges; ③ Time-frequency domain features: Energy is calculated by short-time Fourier transform (STFT), the formula is: Among them, ω(τ-t) is the window function; ④ Feature screening: Screen IMFs based on the frequency standard of sleep stages. The specific conditions are: WAKE stage: main frequency f dom ≥8Hz, and zero crossing rate (ZCR)>0.3; REM stage: 4Hz≤f dom ≤13Hz; NREM stage: 0.5Hz≤f dom ≤7Hz; Calculate the Pearson correlation coefficient between IMFs and the original signal, and retain IMFs with correlation coefficients higher than the threshold.

6. The sleep disorder prediction method according to claim 1, characterized in that: The feature weighting in step e) is specifically as follows: MHSA is used to calculate the feature weights in parallel through multiple attention heads, and the attention score of each head is calculated as: Among them, Q i ,K j are the i-th query vector and the j-th key vector, respectively, N is the number of input features, d k is the dimension of the key vector; The weighted features are generated by weighted summation, the formula is: Among them, V j is a value vector; concatenates multiple head outputs and generates the final feature representation through linear transformation.

7. The sleep disorder prediction method according to claim 1, characterized in that: The time series prediction in step f) is specifically as follows: the weighted features are input into the LSTM network, and the hidden state update formula of the LSTM is: h t =o t ·tanh(c t ) Among them, t is the output gate, c t is the unit state; The LSTM network uses the ReLU activation function, which is defined as: σ(x)=max(0,x) The LSTM output is mapped to the sleep stage classification through a fully connected layer, outputting the probabilities of NREM, REM, and WAKE stages.

8. A sleep disorder prediction system based on deep learning and signal decomposition, characterized in that: include: a) Signal acquisition module: provides clean s-EEG signals as the starting point of the entire system and lays the foundation for parameter optimization and signal decomposition; b) Parameter optimization module: generates the optimal parameters of VMD, directly supporting the accurate operation of the signal decomposition module; c) Signal decomposition module: outputs IMFs and provides decomposed signal components for feature extraction module; d) Feature extraction module: generates feature matrix and provides key feature data for feature weighting module; e) Feature weighting module: optimizes feature representation and provides more targeted input for the time series prediction module; f) Time series prediction module: completes the prediction of sleep disorders and provides the core basis for the result output module; g) Result output module: generates final diagnostic information and completes system functions.

9. The sleep disorder prediction system according to claim 8, characterized in that: Step a) The signal acquisition module is configured to collect and preprocess s-EEG signals through a multi-channel EEG device to generate high-quality clean signals, directly provide input data for the parameter optimization module and the signal decomposition module, and support the accuracy of subsequent signal analysis.

10. The sleep disorder prediction system according to claim 8, characterized in that: Step b) the parameter optimization module is configured to automatically search for optimal parameters of VMD, including the number of modalities and penalty parameters, through a processor, to ensure that the signal decomposition module can efficiently and accurately decompose the s-EEG signal into IMFs.

11. The sleep disorder prediction system according to claim 8, characterized in that: Step c) The signal decomposition module is configured to decompose the input s-EEG signal using the parameters provided by the parameter optimization module to generate IMFs with distinct frequencies as the signal basis for the feature extraction module.

12. The sleep disorder prediction system according to claim 8, characterized in that: Step d) The feature extraction module is configured to extract multi-dimensional features from the IMFs output by the signal decomposition module and generate a structured feature matrix to provide clear and critical data support for the feature weighting module.

13. The sleep disorder prediction system according to claim 8, characterized in that: Step e) The feature weighting module is configured to dynamically weight the feature matrix through a processor to enhance the distinguishability of the features, generate an optimized feature representation, and directly support the prediction accuracy of the time series prediction module.

14. The sleep disorder prediction system according to claim 8, characterized in that: Step f) The time series prediction module is configured to receive the output of the feature weighting module, generate a prediction result of sleep disorder through time series analysis, and provide a reliable core basis for the result output module.

Images