Multi-physiological signal fusion sleep staging method and system

Abstract

The invention relates to the technical field of physiological signal processing and sleep monitoring, in particular to a sleep staging method and a sleep staging system for collecting multiple physiological signals, and the sleep staging method and the sleep staging system for collecting the multiple physiological signals synchronously collect auditory meatus photoelectric volume pulse waves, temperature and head micro-motion signals through an in-ear sensor array. According to the method, the signal quality index is calculated, time domain, frequency domain and nonlinear features are extracted, a dynamic weighted fusion mechanism is adopted, feature weights are adjusted according to the signal quality index, a hierarchical depth time sequence learning model is input for sleep staging, and the sleep staging accuracy is improved to 89% or above and is improved by 15-20% compared with a single brain wave method. A sleep state evaluation report and a personalized feedback intervention strategy generated by the system are beneficial for improving sleep quality, a dynamic weighted fusion mechanism enhances system robustness, adapts to different signal qualities and ensures stable performance, and the invention provides an efficient and accurate new method for the field of sleep monitoring.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of physiological signal processing and sleep monitoring, in particular to a sleep staging method and system for multi-physiological signal acquisition. BACKGROUND

[0002] Sleep health is increasingly valued, and sleep monitoring and intervention technology has also developed rapidly. Existing sleep monitoring devices mainly have two types: one is to simply play music to let the user fall asleep, which cannot improve the user's sleep condition; the other focuses on improving the user's experience of wearing earphones, but these devices still cannot fundamentally assist the user in sleeping.

[0003] The current gold standard for sleep monitoring is polysomnography (PSG), which needs to be performed in a professional sleep laboratory, and is complex to operate and costly, which is not suitable for daily use. Traditional home sleep monitoring devices mainly perform sleep staging based on electroencephalogram (EEG) signals, such as CN109999314B discloses an intelligent sleep-aiding system based on EEG monitoring and a sleep earphone thereof, which collects EEG signals, performs specific classification processing, and realizes sleep staging and sleep-aiding.

[0004] However, such sleep monitoring devices based on EEG have the following technical deficiencies:

[0005] 1. Single physiological signal dependence: completely dependent on EEG signals, and the system accuracy will decrease significantly when the electrode contact is poor;

[0006] 2. Wearing inconvenience and comfort problem: multiple electrodes need to be placed on the head, which affects the user's sleep comfort;

[0007] 3. High signal processing complexity: the wavelet transform feature extraction has high computational complexity, which affects real-time performance;

[0008] 4. Limited sleep staging granularity: mainly focuses on several main sleep stages, and lacks analysis of sleep microstructure;

[0009] 5. Limited adaptability to individual differences: limited adaptability to physiological characteristics of different users.

[0010] Therefore, there is an urgent need for a sleep monitoring and intervention system that is comfortable to wear, has high precision, strong robustness, and good adaptability. SUMMARY

[0011] The purpose of the present application is to provide a multi-physiological signal fusion sleep staging method and system, which synchronously acquires multi-source physiological signals through an in-ear multi-modal sensor array, combines adaptive signal processing, dynamic feature fusion, and deep time series learning, and realizes high-precision, high-comfort sleep monitoring and intervention.

[0012] The present application provides a multi-physiological signal fusion sleep staging method, comprising:

[0013] Synchronously collecting multi-source physiological signals including ear canal photoplethysmography (PPG) signals, ear canal temperature signals and head micro-motion signals through an in-ear multi-modal sensor array;

[0014] Performing adaptive signal preprocessing on the multi-source physiological signals and calculating corresponding signal quality indices, comprising:

[0015] Performing adaptive denoising processing to generate preprocessed multi-source physiological signals;

[0016] Calculating corresponding signal quality indices based on the preprocessed multi-source physiological signals;

[0017] Extracting multi-level physiological features based on the preprocessed multi-source physiological signals and performing dynamic weighted fusion according to the signal quality indices, comprising:

[0018] Extracting time domain features, frequency domain features and nonlinear features from the preprocessed multi-source physiological signals;

[0019] Calculating dynamic weights of each signal feature according to the signal quality indices and performing weighted fusion;

[0020] Inputting the dynamically weighted fused features into a hierarchical deep time sequence learning model to perform sleep staging;

[0021] Based on the sleep staging result, generating a sleep state evaluation report and outputting a personalized feedback intervention strategy.

[0022] Preferably, the adaptive signal preprocessing specifically comprises:

[0023] Applying band-pass filtering and wavelet denoising to the ear canal PPG signals;

[0024] Applying Savitzky-Golay filtering to the ear canal temperature signals;

[0025] Applying empirical mode decomposition to remove motion artifacts from the head micro-motion signals.

[0026] 3. The method of claim 1, wherein the signal quality indices comprise a photoplethysmography signal quality index (SQI) calculated by the formula:

[0027]

[0028] wherein SNR is a signal-to-noise ratio, PI is a perfusion index, MA ratio is a motion artifact ratio.

[0029] Temperature signal reliability index (TRI);

[0030] Head micro-motion signal stability index (MSI).

[0031] As preferred, the multi-level physiological features include:

[0032] Cardiovascular feature spectrum extracted from the ear canal photo-plethysmography signal;

[0033] Thermal regulation features extracted from the ear canal temperature signal;

[0034] Body motion feature spectrum extracted from the head micro-motion signal;

[0035] Cross-correlation features among the above features.

[0036] As preferred, the dynamic weighted fusion includes:

[0037] Signal adaptive weights based on the signal quality index;

[0038] Feature weights based on sleep stage relevance;

[0039] Adaptive weights based on individual differences;

[0040] Enhancing feature complementarity through feature cross-projection, high-order feature interaction, and multi-modal feature alignment. As preferred, the hierarchical deep time series learning model includes:

[0041] One-dimensional convolutional neural network for extracting short-time scale features;

[0042] Bidirectional long short-term memory network for capturing medium-time scale dependencies;

[0043] Transformer encoder for modeling long-range dependencies;

[0044] Time series attention mechanism and conditional random field state sequence modeling module.

[0045] As preferred, the sleep staging includes:

[0046] Generating time series context representation through recursive attention update mechanism;

[0047] Outputting sleep stage probability distribution based on time series context representation;

[0048] Constraining sleep stage transitions using conditional random fields;

[0049] Evaluating the reliability of predictions through a Bayesian deep learning framework.

[0050] As preferred, the sleep state assessment report includes:

[0051] sleep continuity index, sleep depth index and sleep stability index;

[0052] sleep microstructure analysis results, including sleep spindles, K-complexes and micro-awakening event statistics;

[0053] an overall sleep quality score, which is calculated as follows:

[0054] SQS = 0.4 x SCI + 0.4 x SDI + 0.2 x SSI,

[0055] wherein SCI is the sleep continuity index, SDI is the sleep depth index, and SSI is the sleep stability index.

[0056] As preferred, the personalized feedback intervention strategy comprises:

[0057] subconscious auditory stimulation based on automatic adjustment of audio characteristics according to current circadian rhythm;

[0058] a closed-loop optimization framework personalized through transfer learning;

[0059] a three-level feedback loop mechanism including short-term, medium-term and long-term feedback loops.

[0060] a multi-physiological signal fusion sleep staging system, comprising:

[0061] an in-ear multi-modal sensor array module for synchronously collecting multi-source physiological signals including ear canal photoplethysmography signals, ear canal temperature signals and head micro-motion signals;

[0062] a signal preprocessing module for performing adaptive signal preprocessing on the multi-source physiological signals and calculating corresponding signal quality indices;

[0063] a feature extraction and fusion module for extracting multi-level physiological features based on the preprocessed multi-source physiological signals and performing dynamic weighted fusion according to the signal quality indices;

[0064] a deep time series learning module for receiving the dynamically weighted fused features and performing sleep staging;

[0065] a sleep evaluation and intervention module for generating a sleep state evaluation report based on the sleep staging results and outputting a personalized feedback intervention strategy.

[0066] The present application has the following beneficial effects:

[0067] 1. Improved monitoring accuracy: through multi-source physiological signal fusion, the sleep staging accuracy is improved to more than 89%, which is 15-20% higher than the single electroencephalogram method;

[0068] 2. Enhanced system robustness: Dynamic weighted fusion mechanism enables the system to adapt to different signal quality conditions, maintaining stable performance even when individual signal quality declines;

[0069] 3. Improved user experience: In-ear design significantly improves wear comfort, with a 60% increase in user ratings and a 45% increase in 28-day continuous use rate;

[0070] 4. Achieved personalized adaptation: Through transfer learning and incremental learning, the system can quickly adapt to individual differences, requiring only 10-15 minutes of calibration data;

[0071] 5. Reduced power consumption: Compared with traditional electroencephalogram monitoring devices, power consumption is reduced by 50%, with a single charge lasting more than 8 hours;

[0072] 6. Significantly improved sleep quality: Average sleep onset time is reduced by 8-12 minutes, deep sleep proportion is increased by 15-20%, and wake-up frequency is reduced by 30-40%. BRIEF DESCRIPTION OF DRAWINGS

[0073] Figure 1 is the overall flowchart of the method of the present application;

[0074] Figure 2 is the adaptive signal preprocessing flowchart;

[0075] Figure 3 is the multi-level physiological feature extraction and dynamic weighted fusion schematic diagram;

[0076] Figure 4 is the hierarchical deep time series learning model architecture diagram;

[0077] Figure 5 is the sleep state evaluation and personalized intervention system framework diagram;

[0078] Figure 6 is the overall architecture diagram of the system of the present application. DETAILED DESCRIPTION

[0079] Please refer to Figures 1-6 , the specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings.

[0080] Example 1: Overall flow of multi-physiological signal fusion sleep staging method

[0081] As shown in Figure 1 , the multi-physiological signal fusion sleep staging method provided by the present application mainly includes the following five steps:

[0082] Step S1: Synchronously collect multi-source physiological signals through an in-ear multi-modal sensor array.

[0083] Specifically, through the multi-modal sensor array integrated on the earplug, multi-source physiological signals including ear canal photoplethysmography (PPG) signal, ear canal temperature signal and head micro-motion signal are synchronously collected. The in-ear sensor array is designed with biomechanical optimization to ensure comfortable fit and stable signal quality.

[0084] Step S2: Perform adaptive signal preprocessing on the multi-source physiological signals and calculate the corresponding signal quality index.

[0085] Specifically, different types of physiological signals are processed using corresponding adaptive signal preprocessing methods, including filtering, denoising and artifact removal, etc. At the same time, the quality index of each signal is calculated based on the signal characteristics, which is used for weight distribution in the subsequent fusion process.

[0086] Step S3: Based on the pre-processed multi-source physiological signals, multi-level physiological features are extracted, and dynamic weighted fusion is performed according to the signal quality index.

[0087] Specifically, multi-level features such as time domain features, frequency domain features and nonlinear features are extracted from the processed physiological signals of various types. According to the signal quality index calculated in step S2, combined with sleep stage correlation and individual differences, dynamic weighted fusion is performed on each feature to enhance feature complementarity.

[0088] Step S4: Input the dynamically weighted fused features into the hierarchical deep time series learning model to perform sleep staging.

[0089] Specifically, a deep learning framework containing multi-scale time series feature extraction, time series attention mechanism and conditional random field state sequence modeling is used to process the fused features, and the sleep stage classification results are output.

[0090] Step S5: Based on the sleep staging results, generate a sleep state evaluation report and output personalized feedback intervention strategies.

[0091] Specifically, according to the staging results, sleep continuity index, sleep depth index and other evaluation indicators are calculated to generate a sleep state evaluation report. At the same time, based on the evaluation results, a customized audio stimulus and sleep improvement suggestion are generated by the personalized feedback intervention system.

[0092] Embodiment 2: In-ear multi-modal sensor array and signal acquisition

[0093] This embodiment describes in detail the in-ear multi-modal sensor array and signal acquisition process in step S1.

[0094] The in-ear multi-modal sensor array uses flexible printed circuit (FPC) technology to arrange a multi-sensor array on the surface of the biomechanically optimized earplug, including the following main sensor components:

[0095] 1. PPG sensor: integrated 940nm infrared LED and silicon photodiode, sampling rate set to 60Hz, can accurately detect the tiny blood flow changes in the ear canal, thus obtaining heart rate information;

[0096] 2. Precision thermistor: temperature accuracy of ±0.05℃, sampling rate of 1Hz, used to monitor the change of ear canal temperature, reflecting the circadian rhythm change of human body;

[0097] 3. Six-axis inertial sensor (BMI160): contains three-axis accelerometer and three-axis gyroscope, sampling rate of 100Hz, used to detect head micro-motion and attitude change.

[0098] The signal synchronous acquisition architecture is based on the design of a master control unit of TI CC2640R2F low-power Bluetooth SoC, which realizes μs-level timestamp synchronization and ensures the alignment of multiple signal timing. Preferably, a dynamic sampling scheduling algorithm is adopted to adaptively adjust the sampling frequency according to the sleep state, so as to realize power optimization, specifically:

[0099] 1. Wakeful state: maintain high sampling rate for all sensors;

[0100] 2. Light sleep stage: reduce PPG sampling rate to 30Hz, and inertial sensor sampling rate to 50Hz;

[0101] 3. Deep sleep stage: further reduce PPG sampling rate to 15Hz, and inertial sensor sampling rate to 20Hz.

[0102] Through this dynamic adjustment strategy, compared with the fixed sampling rate scheme, the power consumption can be reduced by 35-40%, while the integrity of the key physiological information is maintained.

[0103] Embodiment 3: Adaptive signal preprocessing and quality evaluation framework

[0104] As shown in Figure 2 , this embodiment describes in detail the adaptive signal preprocessing and quality evaluation framework in step S2.

[0105] Adaptive signal denoising adopts specific processing methods for different physiological signals:

[0106] 1. Apply band-pass filtering (0.5-8Hz) and wavelet denoising to PPG signal, use Symlet5 wavelet, 5-level decomposition to remove baseline drift and high-frequency noise;

[0107] 2. Apply Savitzky-Golay filter to body temperature signal, window length set to 15, polynomial order set to 3, smooth temperature fluctuations while retaining trend changes;

[0108] 3. Empirical Mode Decomposition (EMD) is applied to the inertial signal to decompose the signal into multiple Intrinsic Mode Functions (IMFs), and the IMF components reflecting micro-motions are retained while the gross motion artifacts are removed.

[0109] The signal quality real-time evaluation mechanism performs quality scoring on various signals, including:

[0110] 1. The calculation formula of the PPG signal quality index (SQI) is:

[0111]

[0112] wherein SNR is the signal-to-noise ratio, PI is the perfusion index, MA ratio is the motion artifact ratio. SNR max and PI max are the standard reference upper limit values of the signal-to-noise ratio and the perfusion index, respectively.

[0113] 2. The calculation formula of the temperature signal reliability index (TRI) is:

[0114]

[0115] wherein ΔT / Δt is the temperature change rate, Thres T is the temperature change rate threshold, σ T is the short-time window temperature standard deviation, is the standard deviation upper limit.

[0116] 3. The calculation formula of the inertial signal stability index (MSI) is:

[0117]

[0118] wherein Acc var is the acceleration variance, Gyr var is the angular velocity variance, and are the respective variance upper limits.

[0119] The data integrity compensation technology is based on the time sequence correlation of physiological signals, uses multivariate Kalman filtering to predict missing values, and when there is a long time continuous missing, uses a sliding window condition to generate a generative adversarial network (SW-CGAN) to repair the signal segment, achieving a signal integrity rate of more than 95%. Preferably, the sliding window length is set to 10 seconds, and the overlap rate is 50% to balance the repair accuracy and computational efficiency.

[0120] Embodiment 4: Multi-level physiological feature extraction

[0121] This embodiment describes in detail the multi-level physiological feature extraction process in step S3.

[0122] Multi-level physiological features are extracted from the pre-processed multi-source physiological signals, including:

[0123] 1. Cardiovascular feature hierarchy extraction:

[0124] Time-domain features: Calculate HRV indices such as SDNN (standard deviation of NN intervals), RMSSD (root mean square of successive differences), pNN50 (percentage of adjacent NN intervals with a difference greater than 50 ms) on a 300-second window (50% overlap rate);

[0125] Frequency-domain features: Apply wavelet packet transform to extract VLF (0.003-0.04 Hz), LF (0.04-0.15 Hz), and HF (0.15-0.4 Hz) band power ratios, reflecting the balance between sympathetic and parasympathetic nervous activity;

[0126] Morphological features: Analyze the second derivative of the PPG waveform (accelerated photoplethysmogram), extract wave peak ratios (a / b, c / d, e / a) and wave peak time differences, reflecting vascular elasticity and peripheral resistance.

[0127] 2. Thermal regulation feature extraction:

[0128] Absolute body temperature and relative change rate;

[0129] Temperature fluctuation features based on wavelet entropy, used to capture circadian rhythm changes;

[0130] Cross-correlation features of body temperature change rate and heart rate variability, reflecting the coordinated changes of the autonomic nervous system.

[0131] 3. Body motion feature hierarchy:

[0132] Statistical features: Calculate mean, standard deviation, kurtosis, skewness, etc., to represent overall activity level;

[0133] Spectral features: Extract power spectral density 0.1-1 Hz, 1-3 Hz, 3-8 Hz band energy ratios to distinguish different types of physical activity;

[0134] Time-frequency features: Apply wavelet transform to extract coefficients and entropy values to capture non-stationary motion patterns.

[0135] Preferably, all feature calculations use 30-second non-overlapping windows, consistent with the traditional sleep staging time unit, facilitating comparison with the gold standard; at the same time, features of 1-minute, 5-minute and 10-minute windows are retained to capture physiological changes at different time scales.

[0136] Example 5: Dynamic weighted fusion mechanism

[0137] As Figure 3As shown, the embodiment details the dynamic weighting fusion mechanism in step S3.

[0138] The dynamic weighting fusion mechanism includes the following key parts:

[0139] 1. Signal quality adaptive weight adjustment: Based on the signal quality indices calculated in embodiment 3, the weights of each signal feature are dynamically adjusted, and the calculation formula is:

[0140]

[0141] Where W PPG , W Temp , and W Motion are the weights of the PPG signal, temperature signal, and motion signal, respectively.

[0142] 2. Sleep stage relevance weighting: According to the distinguishing ability of different features for each sleep stage, the feature weight is adjusted, and the calculation formula is:

[0143]

[0144] Where base_weight i is the base weight of feature i, and stage_relevance i is the distinguishing ability coefficient of feature i for different sleep stages, which is determined by offline analysis of the distinguishing degree of each feature in each sleep stage.

[0145] 3. Individual difference adaptive weight: Through incremental learning, the weights of each feature are continuously optimized, and the basic formula is:

[0146]

[0147] Where α is the learning rate (usually set to 0.05-0.1), and Δperformance i is the performance improvement after adding feature i.

[0148] 4. Feature complementarity enhancement technology:

[0149] Feature cross projection: Project single-modal features into other modal feature spaces to create cross-modal representations;

[0150] High-order feature interaction: Calculate second-order and third-order interaction terms between different modal features to capture complex nonlinear relationships;

[0151] Multi-modal feature alignment: Perform time alignment and feature standardization to ensure comparability of different signal features.

[0152] Preferably, different feature combination strategies are adopted for different sleep stages: shallow sleep stage focuses on heart rate variability indicators; deep sleep stage focuses on body temperature changes and low-frequency body movement features; REM stage focuses on eye movement related head micro-movement features. This dynamic adjustment strategy improves the sleep staging accuracy by 8-12% compared with the fixed weight scheme.

[0153] Embodiment 6: Hierarchical deep time series learning framework

[0154] As shown in Figure 4 , this embodiment describes in detail the hierarchical deep time series learning framework in step S4.

[0155] The hierarchical deep time series learning framework includes the following core modules:

[0156] 1. Multi-scale time series feature extraction network:

[0157] Short time scale (5-30s): 1D-CNN is used to extract local time features, with convolution kernel size set to [3, 5, 7] and channel number set to [32, 64, 128] to capture rapid change patterns;

[0158] Medium time scale (30s-5min): BiLSTM is used to capture medium-length dependencies, with hidden layer unit number set to 128 to capture sequence patterns;

[0159] Long time scale (5min-2h): Transformer encoder is used to model long-range dependencies, with attention head number set to 4 and hidden layer dimension set to 256 to capture long-term rhythm changes.

[0160] 2. Innovative time series attention mechanism:

[0161] Multi-head self-attention: captures the importance of different time scales, with head number set to 4;

[0162] Cross-modal attention: enhances information interaction between different physiological signals, calculates attention weights between different modal features;

[0163] Recursive attention update: continuously optimizes attention allocation through update mechanism, with calculation formula as follows:

[0164] A t = softmax(V·tanh(W1·h t +W2·c t +b)),

[0165] Where h t is the current hidden state, c t is the historical context vector, W1, W2, V are learnable weight matrices, and b is the bias term.

[0166] 3. Sleep state sequence modeling: Conditional random field (CRF) is used to model state transition constraints. Combined with prior rules of sleep physiology, the state transition feature function is:

[0167] Ψ(y t-1 ,y t ) = exp(w trans ·f trans (y t-1 ,y t )),

[0168] Among them, f trans The state transition feature function encodes the sleep cycle pattern, w trans For the corresponding weights.

[0169] 4. Uncertainty Estimation and Reliability Evaluation: Using a Bayesian deep learning framework, Monte Carlo Dropout is used to estimate the predicted uncertainty, and the reliability score is calculated as follows:

[0170] R score =1-(σ pred / σ max ),

[0171] Where, σ pred To predict variance, σ max This represents the upper limit of variance.

[0172] 5. Multi-task learning architecture:

[0173] Main task: Sleep stages (W / N1 / N2 / N3 / REM)

[0174] Assistive tasks: Micro-awake detection, sleep quality scoring, and body movement event recognition.

[0175] Task weights are dynamically adjusted:

[0176] L total =∑ i α i ·L i ,

[0177] Where, α i The weights for task i are dynamically adjusted based on the gradient similarity between tasks.

[0178] Preferably, the model training employs a two-stage strategy: first, pre-training is performed on a large-scale sleep database, followed by fine-tuning using a small amount of target user data. Pre-training uses mixed precision training with a learning rate of 0.001, a batch size of 64, and 500 training epochs; during the fine-tuning stage, the learning rate is reduced to 0.0001, the batch size is 16, and 50 training epochs are performed.

[0179] Example 7: Sleep microstructure analysis

[0180] This example describes in detail the sleep microstructure analysis process in step S4.

[0181] Based on sleep staging, the system can also identify the following sleep microstructure events:

[0182] 1. Sleep spindles detection: Continuous Wavelet Transform (CWT) is used to analyze the 12-16 Hz band energy of PPG signal, combined with statistical threshold decision algorithm to detect spindle events. The detection conditions include:

[0183] Band energy exceeds 3 times the background energy;

[0184] Duration is in the range of 0.5-2 seconds;

[0185] Adjacent event interval is greater than 2 seconds.

[0186] 2. K-complex recognition: By analyzing the morphological features of head micro-motion signal, including:

[0187] Sharp unidirectional acceleration change;

[0188] Amplitude exceeds 4 times the standard deviation of the quiet period;

[0189] Duration is in the range of 0.5-1 second.

[0190] 3. Micro-awakening event detection: Using a multi-signal fusion anomaly detection framework, based on the following features:

[0191] Spectrum mutation of PPG signal;

[0192] Short-term increase in heart rate (more than 15% increase over baseline);

[0193] Short-term increase in body motion amplitude;

[0194] Duration is in the range of 3-15 seconds.

[0195] 4. Periodic breathing event recognition: Through the analysis of respiratory modulation component of PPG signal, detect apnea and hypopnea events:

[0196] Extract PPG envelope signal as respiratory proxy signal;

[0197] Detect events with amplitude reduction of more than 50% and duration of more than 10 seconds;

[0198] Analyze event periodicity to identify periodic breathing patterns.

[0199] Preferably, the microstructure analysis adopts a 10-second sliding window with an overlap rate of 75% to balance the detection accuracy and time resolution. The overall accuracy of microstructure event detection reaches 83%, which is about 20% higher than the traditional single sensor method.

[0200] Embodiment 8: Sleep state evaluation report generation

[0201] This embodiment describes in detail the sleep state evaluation report generation process in step S5.

[0202] Based on the sleep staging results and microstructure analysis, the system generates a multi-dimensional sleep quality evaluation, including:

[0203] 1. Sleep continuity index (SCI), the calculation formula is:

[0204]

[0205] Where N awakenings is the number of night awakenings, N threshold is the threshold number of awakenings (usually set to 4), T awake is the wake-up time, T total is the total recording time.

[0206] 2. Sleep depth index (SDI), the calculation formula is:

[0207]

[0208] Where T N3 is the N3 (deep sleep) time, T REM is the REMM (rapid eye movement) sleep time, T total is the total recording time.

[0209] 3. Sleep stability index (SSI), the calculation formula is:

[0210]

[0211] Where N transitions is the number of sleep stage transitions, N exp is the expected number of transitions, is the PPG variability of the ith window, PPG var_max is the upper limit of variability, and N is the total number of windows.

[0212] 4. Overall sleep quality score (SQS), the calculation formula is:

[0213] 1. SQS = 0.4 × SCI + 0.4 × SDI + 0.2 × SSI,

[0214] In addition to the above indicators, the sleep report also includes:

[0215] Sleep cycle integrity analysis, assessing how many complete sleep cycles the user experienced;

[0216] Sleep microstructure statistics, including spindle density, K-complex frequency, and micro-awakening index, etc.;

[0217] Comparison with healthy reference ranges, based on data of the same age and gender;

[0218] Sleep trend analysis, showing the changes in sleep quality in the past 7 or 30 days.

[0219] Preferably, the report adopts visual display, including sleep stage distribution pie chart, sleep cycle flowchart, and key indicator radar chart, etc., to improve user understanding and acceptance.

[0220] Embodiment 9: Personalized feedback intervention system

[0221] As shown in Figure 5 , this embodiment describes in detail the personalized feedback intervention system in step S5.

[0222] The personalized feedback intervention system includes the following main components:

[0223] 1. Adaptive biofeedback system:

[0224] Subconscious auditory stimulation: automatically adjust audio characteristics according to heart rate variability, frequency range 0.5-4Hz, matching the target sleep stage EEG characteristic frequency;

[0225] Rhythm synchronization technology: dynamically adjust audio rhythm according to current physiological rhythm, promote deep sleep, incremental adjustment step 0.1Hz, maximum deviation not more than 0.5Hz;

[0226] Personalized audio synthesis algorithm: based on past response effect, optimize audio parameters, including frequency, volume, timbre and spatial distribution, etc.

[0227] 2. Transfer learning personalized framework:

[0228] Two-stage transfer learning: from general model → specific population model → personalized model, reducing the amount of data required for personalization;

[0229] Feature extraction layer is frozen, only fine-tuning decision layer parameters, accelerating the adaptation process;

[0230] Efficient use of personalized data: only 10-15 minutes of calibration data can improve 15% accuracy.

[0231] 3. Closed-loop optimization framework:

[0232] Short-term feedback loop: Real-time adjustment of stimulation parameters with a 30-second cycle based on immediate physiological responses.

[0233] Mid-term feedback loop: Intervention strategy updated after each night's sleep based on single-night sleep quality assessment.

[0234] Long-term feedback loop: Personalized model updates on a weekly / monthly scale based on cumulative data analysis.

[0235] Swarm intelligence optimization: Federated learning-based swarm knowledge sharing to protect privacy while enhancing model performance.

[0236] Preferably, the system provides three intervention modes:

[0237] Sleep onset assistance mode: Focuses on reducing sleep onset time, using progressive relaxation techniques.

[0238] Deep sleep enhancement mode: Optimizes N3 stage proportion, using low-frequency rhythmic stimulation.

[0239] Comprehensive sleep optimization mode: Balances all sleep stages, using adaptive cycle strategies.

[0240] Users can choose preferred modes through the mobile application, and the system will also recommend the most suitable intervention strategies based on historical data. In practical applications, the deep sleep enhancement mode can increase the deep sleep proportion by an average of 18%, and the sleep onset assistance mode can reduce the sleep onset time by an average of 11 minutes.

[0241] Embodiment 10: Multi-physiological signal fusion sleep staging system

[0242] As shown in Figure 6 , this embodiment describes in detail a multi-physiological signal fusion sleep staging system corresponding to the above method, which includes the following five main modules:

[0243] 1. In-ear multi-modal sensor array module: for synchronous acquisition of multi-source physiological signals including ear canal photoplethysmography (PPG) signal, ear canal temperature signal and head micro-motion signal. This module includes:

[0244] PPG sensing unit: composed of infrared LED emitter and photodiode receiver, detects changes in ear canal blood flow;

[0245] Temperature sensing unit: uses precision thermistor to monitor changes in ear canal temperature;

[0246] Inertial sensing unit: integrates six-axis sensor to detect head micro-motion;

[0247] Signal conditioning circuit: amplification, filtering and analog-to-digital conversion;

[0248] Synchronization control unit: ensures alignment of multi-signal acquisition time stamps.

[0249] 2. Signal Preprocessing Module: for performing adaptive signal preprocessing on multi-source physiological signals and calculating corresponding signal quality indices. This module includes:

[0250] Filtering and Denoising Unit: adaptive processing for different signal types;

[0251] Artifact Detection and Removal Unit: identifying and compensating for motion artifacts and other disturbances;

[0252] Signal Quality Assessment Unit: real-time calculation of quality indices for each signal;

[0253] Data Integrity Compensation Unit: handling signal missing and abnormalities.

[0254] 3. Feature Extraction and Fusion Module: for extracting multi-level physiological features based on preprocessed multi-source physiological signals and performing dynamic weighted fusion according to signal quality indices. This module includes:

[0255] Multi-modal Feature Extraction Unit: extracting time-domain, frequency-domain, and nonlinear features from each signal;

[0256] Feature Selection and Dimensionality Reduction Unit: screening the most effective feature combinations;

[0257] Dynamic Weight Calculation Unit: calculating weights based on signal quality and sleep stage relevance;

[0258] Feature Fusion Unit: performing weighted feature fusion and feature complementarity enhancement.

[0259] 4. Deep Temporal Learning Module: for receiving dynamically weighted fused features and performing sleep staging. This module includes:

[0260] Multi-scale Feature Extraction Network: processing features of different time scales;

[0261] Temporal Attention Unit: capturing important temporal patterns and relationships;

[0262] State Sequence Modeling Unit: applying conditional random fields to constrain sleep state transitions;

[0263] Multi-task Learning Unit: simultaneously performing sleep staging and microstructure analysis;

[0264] Uncertainty Estimation Unit: evaluating the reliability of prediction results.

[0265] 5. Sleep Assessment and Intervention Module: for generating sleep state assessment reports and outputting personalized feedback intervention strategies based on sleep staging results. This module includes:

[0266] Sleep Quality Assessment Unit: calculating multi-dimensional sleep quality indicators;

[0267] Report generation unit: create visual sleep report;

[0268] Intervention strategy generation unit: determine optimal intervention scheme based on evaluation results;

[0269] Biofeedback control unit: adjust audio stimulation parameters;

[0270] Personalized learning unit: continuously optimize intervention strategy based on user feedback.

[0271] The modules are connected through a data bus to form a complete data processing and feedback loop. The system uses a hierarchical computing architecture, assigning low complexity tasks to the edge side for execution, and high complexity tasks to the mobile terminal or cloud for execution, balancing real-time performance and computing resource requirements.

[0272] The hardware implementation of the system can be divided into three layers:

[0273] Earplug end: integrated sensor, primary signal processing and low power communication module;

[0274] Mobile terminal: perform main signal processing, feature extraction and classification tasks;

[0275] Cloud: provides model training, data storage and group knowledge sharing functions.

[0276] Preferably, the earplug end uses low power design, a single charge can work more than 8 hours; the mobile application supports iOS and Android systems, providing a user-friendly interface; the cloud service uses secure encrypted transmission to ensure data privacy protection.

[0277] Example 11: Application case analysis

[0278] This example shows the application effect of the present application in actual sleep monitoring scenarios.

[0279] 30 subjects of different ages (20-60 years old) were selected, including 15 healthy controls and 15 mild sleep disorder patients, and the present system and traditional polysomnography (PSG) were used for sleep monitoring, and the consistency and user experience of the two methods were compared.

[0280] Comparison of sleep staging accuracy:

[0281] The overall accuracy of the present system is 89.3% (relative to the PSG gold standard)

[0282] Accuracy of each sleep stage classification:

[0283] Wakefulness (W): 92.5%;

[0284] Light sleep (N1): 78.4%;

[0285] Light sleep (N2): 87.6%;

[0286] Deep sleep (N3): 91.2%;

[0287] REM sleep: 89.8%;

[0288] Compared with traditional single brain wave-based sleep monitoring devices, the system of the present application has the following significant advantages:

[0289] Accuracy improvement: overall accuracy improved by 15-20%;

[0290] Robustness enhancement: performance degradation under body motion interference reduced by 60%;

[0291] User comfort: comfort score improved by 60% (based on a 10-point user survey);

[0292] Continuous use rate: 28-day continuous use rate increased by 45%;

[0293] Through the application of the personalized intervention function of the system of the present application, the sleep quality of the subjects was significantly improved after a period of 4 weeks of use:

[0294] Sleep onset time: average reduction of 9.4 minutes;

[0295] Deep sleep ratio: average increase of 17.3%;

[0296] Number of night awakenings: average reduction of 38.2%;

[0297] Subjective sleep quality score: average improvement of 35.7% (based on the PSQI scale);

[0298] This case verifies the effectiveness and superiority of the present application in practical application, especially in the daily continuous use scenario in non-laboratory environment, showing high practical value.

[0299] The present application has significant industrial practicality:

[0300] 1. High hardware implementation feasibility: the sensors used are all mature commercial products (PPG sensor selects OSRAM SFH 7060, temperature sensor selects TI TMP117, and inertial sensor selects Bosch BMI160), which can be integrated through flexible printed circuits and medical-grade silicone shells to realize industrial-scale production;

[0301] 2. Reasonable power consumption control: through dynamic sampling scheduling and hierarchical computing architecture, the average power consumption is controlled below 10mW, and a single charge can be used for more than 8 hours, meeting the whole night monitoring demand;

[0302] 3. Algorithm deployment is efficient: lightweight model (<5MB) can run smoothly on mid-end mobile devices, signal processing delay <100ms, sleep state determination delay <1s, meeting real-time feedback requirements;

[0303] 4. Cost control is reasonable: sensor and electronic component cost is about $25, PCB and structural part cost is about $10, total BOM cost is controlled within $50, and it has strong market competitiveness;

[0304] 5. Application scenarios are extensive: suitable for home sleep monitoring, chronic sleep disorder management, sleep health research and other scenarios, and the market potential is huge.

[0305] In summary, the present application realizes a high-precision and high-comfort sleep monitoring and intervention system through an innovative ear canal multi-physiological signal fusion method, and has high industrial application value and market promotion potential.

[0306] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Various modifications and changes can be made by those skilled in the art based on the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-physiological signal fusion sleep staging method, characterized in that, The method comprises the following steps: Synchronously collecting multi-source physiological signals including ear canal photoplethysmography (PPG) signals, ear canal temperature signals, and head micro-motion signals through an intra-aural multi-modal sensor array; Performing adaptive signal preprocessing on the multi-source physiological signals and calculating corresponding signal quality indices, including: Performing adaptive denoising processing to generate preprocessed multi-source physiological signals; Based on the preprocessed multi-source physiological signals, calculating corresponding signal quality indices; Based on the preprocessed multi-source physiological signals, extracting multi-level physiological features and performing dynamic weighted fusion according to the signal quality indices, including: Extracting time-domain features, frequency-domain features, and nonlinear features from the preprocessed multi-source physiological signals; Calculating dynamic weights of each signal feature according to the signal quality indices and performing weighted fusion; Inputting the dynamically weighted and fused features into a hierarchical deep time series learning model to perform sleep staging; Based on the sleep staging results, generating a sleep state evaluation report and outputting a personalized feedback intervention strategy.

2. The method of claim 1, wherein, The adaptive signal preprocessing specifically includes: Applying band-pass filtering and wavelet denoising to the ear canal PPG signals; Applying Savitzky-Golay filtering to the ear canal temperature signals; Applying empirical mode decomposition to remove motion artifacts from the head micro-motion signals.

3. The method of claim 1, wherein, The signal quality indices include: A photoplethysmography signal quality index (SQI) whose calculation formula is: where SNR is the signal-to-noise ratio, PI is the perfusion index, MA ratio is the motion artifact ratio; A temperature signal reliability index (TRI); A head micro-motion signal stability index (MSI).

4. The method of claim 1, wherein, The multi-level physiological features include: A cardiovascular feature spectrum extracted from the ear canal PPG signals; A thermal regulation feature extracted from the ear canal temperature signals; A body motion feature spectrum extracted from the head micro-motion signals; Cross-correlation features among the above features.

5. The method of claim 1, wherein, The dynamic weighted fusion includes: Calculating signal adaptive weights based on the signal quality indices; Calculating feature weights based on sleep stage correlations; Calculating adaptive weights based on individual differences; Enhancing feature complementarity through feature cross-projection, high-order feature interaction, and multi-modal feature alignment.

6. The method of claim 1, wherein, The hierarchical deep time series learning model includes: A one-dimensional convolutional neural network for extracting short-time scale features; A bidirectional long short-term memory network for capturing medium-time scale dependencies; A Transformer encoder for modeling long-range dependencies; A time series attention mechanism and a conditional random field state sequence modeling module.

7. The method of claim 1, wherein, The sleep staging includes: Generating a time series context representation through a recursive attention update mechanism; Outputting a sleep stage probability distribution based on the time series context representation; Constraining sleep stage transitions using a conditional random field; Evaluating the reliability of predictions through a Bayesian deep learning framework.

8. The method of claim 1, wherein, The sleep state evaluation report includes: A sleep continuity index, a sleep depth index, and a sleep stability index; Sleep microstructure analysis results, including sleep spindle, K-complex, and micro-awakening event statistics; An overall sleep quality score whose calculation formula is: SQS = 0.4 × SCI + 0.4 × SDI + 0.2 × SSI, Wherein, SCI is sleep continuity index, SDI is sleep depth index, and SSI is sleep stability index.

9. The method of claim 1, wherein, The personalized feedback intervention strategy comprises: Subconscious auditory stimulation based on automatic adjustment of audio characteristics of current circadian rhythm; A closed-loop optimization framework personalized by transfer learning; A three-level feedback loop mechanism including short-term, medium-term and long-term.

10. A multi-physiological signal fusion sleep staging system, characterized by, Comprise: An in-ear multi-modal sensor array module for synchronously collecting multi-source physiological signals including ear canal photoplethysmography pulse wave signals, ear canal temperature signals and head micro-motion signals; A signal preprocessing module for performing adaptive signal preprocessing on the multi-source physiological signals and calculating corresponding signal quality indexes; A feature extraction and fusion module for extracting multi-level physiological features based on the preprocessed multi-source physiological signals and performing dynamic weighted fusion according to the signal quality indexes; A deep time series learning module for receiving the dynamically weighted fused features and performing sleep staging; A sleep evaluation and intervention module for generating a sleep state evaluation report based on the sleep staging result and outputting a personalized feedback intervention strategy.

Images