A sleep disorder risk prediction and intervention method and device

Abstract

The present application provides a sleep disorder risk prediction and intervention method and device, collects multi-modal time series data of a user; processes the multi-modal time series data and inputs it into a risk prediction model, outputs a probability time series prediction result; generates an intervention instruction according to the probability time series prediction result, and executes the intervention instruction. By fusing multi-modal data such as night sleep, daytime activity, environmental factors and user subjective feelings, a systematic full-cycle health view is constructed, so that the root cause of individual sleep problems is more comprehensively understood. Secondly, intelligent and personalized active closed-loop intervention is realized, the intervention measures are adjusted according to the risk level, and the strategy is continuously optimized through the feedback mechanism. Finally, the time series deep learning model based on the advanced architecture is adopted, which effectively captures the complex time dependence of multi-modal data, significantly improves the prediction accuracy and reliability. The present application can predict the probability of sleep disorder risk events, providing a valuable window for active intervention and risk prevention.

Description

Technical Field

[0001] This invention relates to the field of intelligent health monitoring technology, and in particular to a method and device for predicting and intervening in the risk of sleep disorders. Background Technology

[0002] With changing lifestyles, sleep disorders such as obstructive sleep apnea, insomnia, and periodic limb movement disorder have become significant public health issues. While traditional polysomnography (PSG) offers high accuracy in professional medical institutions, its expensive equipment, complex operation, and potential sleep disturbances make it unsuitable for ordinary households. Therefore, in recent years, advancements in sensor technology and the Internet of Things (IoT) have driven rapid progress in home sleep monitoring technology.

[0003] However, current sleep monitoring devices, whether wearable or contactless, primarily focus on physiological signals detected during sleep. This singular data source leads to "data silos," making it difficult to effectively establish the relationship between lifestyle and sleep health. Furthermore, existing data analysis models typically process data after sleep has ended, generating assessment reports on sleep quality, and their intervention methods are relatively limited, failing to provide intelligent adjustments based on risk levels.

[0004] Finally, most monitoring products rely on general algorithm models, resulting in limited adaptability and accuracy when dealing with different individuals, making it difficult to provide truly personalized health management solutions. Therefore, there is an urgent need for innovative technologies to integrate multi-dimensional data and achieve proactive and personalized sleep health management. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a method and apparatus for predicting and intervening in the risk of sleep disorders, in order to solve the problem of low accuracy in sleep monitoring and intervention.

[0006] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions:

[0007] The first aspect of this invention discloses a method for predicting and intervening in the risk of sleep disorders, the method comprising:

[0008] Collect users' multimodal time-series data;

[0009] The multimodal time series data is processed and input into the risk prediction model, and the probabilistic time series prediction results are output.

[0010] An intervention instruction is generated based on the probability time series prediction results, and the intervention instruction is executed.

[0011] Preferably, the multimodal time-series data includes sleep physiological data, environmental data, wakefulness physiological data and behavioral data, as well as health logs;

[0012] The collected multimodal time-series data of users includes:

[0013] The system collects sleep physiological data of users during sleep using non-contact sensors and environmental data of users during sleep using environmental sensors.

[0014] Use wearable devices to collect users' physiological and behavioral data when they are awake;

[0015] Obtain health logs entered by the user through the terminal.

[0016] Preferably, the step of processing the multimodal time series data and inputting it into the risk prediction model to output probabilistic time series prediction results includes:

[0017] The multimodal time series data is processed into a multidimensional feature vector sequence and position encoding is added to obtain the target multidimensional feature vector sequence;

[0018] The target multidimensional feature vector sequence is input into the risk prediction model. In the risk prediction model, the encoder encodes the target multidimensional feature vector sequence through a multi-head self-attention mechanism to generate a contextual representation.

[0019] The decoder in the risk prediction model decodes the context representation in an autoregressive manner to obtain the probability time series prediction result, and outputs the probability time series prediction result.

[0020] Preferably, generating the intervention instruction based on the probability time series prediction result includes:

[0021] The comprehensive risk score corresponding to the probability time series prediction result is calculated based on the multi-criteria decision fusion logic.

[0022] The comprehensive risk score is mapped to a preset risk level system to obtain the risk level;

[0023] The intervention strategy corresponding to the risk level is obtained from the intervention strategy library to generate an intervention instruction.

[0024] Preferably, the method further includes:

[0025] Collect physiological data of the user after intervention, and construct formatted experience data based on the intervention instructions and the physiological data;

[0026] Based on the offline reinforcement learning learning framework, the strategies in the intervention strategy library are updated according to the empirical data, and the updated intervention strategy library is output.

[0027] Based on the empirical data, the parameters of the risk prediction model are updated using a low-rank adaptive method to obtain the latest risk prediction model.

[0028] A second aspect of this invention discloses a sleep disorder risk prediction and intervention device, the device comprising:

[0029] The acquisition unit is used to acquire users' multimodal time-series data;

[0030] The prediction unit is used to process the multimodal time series data and input it into the risk prediction model, and output the probability time series prediction result.

[0031] An intervention unit is used to generate intervention instructions based on the probability time series prediction results and to execute the intervention instructions.

[0032] Preferably, the multimodal time-series data includes sleep physiological data, environmental data, wakefulness physiological data and behavioral data, as well as health logs;

[0033] The acquisition unit includes:

[0034] The first acquisition module is used to acquire sleep physiological data of the user during sleep through non-contact sensors and to acquire environmental data of the user during sleep through environmental sensors.

[0035] The second data acquisition module is used to collect the user's physiological and behavioral data when the user is awake using wearable devices;

[0036] The acquisition module is used to acquire the health logs entered by the user through the terminal.

[0037] Preferably, the prediction unit includes:

[0038] The processing module is used to process the multimodal time series data into a multidimensional feature vector sequence and add position encoding to obtain the target multidimensional feature vector sequence;

[0039] The encoding module is used to input the target multidimensional feature vector sequence into the risk prediction model. In the risk prediction model, the encoder encodes the target multidimensional feature vector sequence through a multi-head self-attention mechanism to generate a contextual representation.

[0040] The decoding module is used by the decoder in the risk prediction model to decode the context representation in an autoregressive manner, obtain the probability time series prediction result, and output the probability time series prediction result.

[0041] Preferably, the intervention unit includes:

[0042] The calculation module is used to calculate the comprehensive risk score corresponding to the probability time series prediction result based on the multi-criteria decision fusion logic;

[0043] The mapping module is used to map the comprehensive risk score to a preset risk level system to obtain a risk level;

[0044] The generation module is used to obtain the intervention strategy corresponding to the risk level from the intervention strategy library in order to generate intervention instructions.

[0045] Preferably, the device further includes:

[0046] A construction unit is used to collect the user's physiological data after intervention and to construct formatted experience data based on the intervention instructions and the physiological data;

[0047] The first update unit is used to update the strategies in the intervention strategy library based on the offline reinforcement learning learning framework, according to the empirical data, and output the updated intervention strategy library.

[0048] The second update unit is used to update the parameters of the risk prediction model based on the empirical data using a low-rank adaptive method, so as to obtain the latest risk prediction model.

[0049] Based on the above embodiments of the present invention, a method and apparatus for predicting and intervening in sleep disorder risks are provided. This method collects multimodal time-series data from users; processes the multimodal time-series data and inputs it into a risk prediction model, outputting probabilistic time-series prediction results; generates intervention instructions based on the probabilistic time-series prediction results, and executes the intervention instructions. By integrating multimodal data such as nighttime sleep, daytime activities, environmental factors, and user subjective feelings, a systematic, full-cycle health view is constructed, thereby providing a more comprehensive understanding of the root causes of individual sleep problems. Secondly, it achieves intelligent and personalized proactive closed-loop intervention, adjusting intervention measures according to risk levels and continuously optimizing strategies through a feedback mechanism. Finally, it employs a time-series deep learning model based on an advanced architecture to effectively capture the complex temporal dependencies of multimodal data, significantly improving prediction accuracy and reliability. This invention can predict the probability of sleep disorder risk events, providing a valuable window for proactive intervention and risk prevention. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0051] Figure 1 A flowchart illustrating a method for predicting and intervening in the risk of sleep disorders provided in an embodiment of the present invention;

[0052] Figure 2 This is a structural block diagram of a sleep disorder risk prediction and intervention device provided in an embodiment of the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] In this application, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0055] As can be seen from the background technology, current sleep monitoring devices suffer from limited data sources, outdated analysis models, and reliance on general algorithms, making it difficult to effectively link lifestyle and sleep health, and resulting in a lack of accurate intervention and management solutions.

[0056] Therefore, embodiments of the present invention provide a method and apparatus for predicting and intervening in sleep disorder risks. This method collects multimodal time-series data from users; processes the multimodal time-series data and inputs it into a risk prediction model, outputting probabilistic time-series prediction results; generates intervention instructions based on the probabilistic time-series prediction results, and executes the intervention instructions. By integrating multimodal data such as nighttime sleep, daytime activities, environmental factors, and user subjective feelings, a systematic, full-cycle health view is constructed, thereby providing a more comprehensive understanding of the root causes of individual sleep problems. Secondly, it achieves intelligent and personalized proactive closed-loop intervention, adjusting intervention measures according to risk levels and continuously optimizing strategies through a feedback mechanism. Finally, it employs a time-series deep learning model based on an advanced architecture to effectively capture the complex temporal dependencies of multimodal data, significantly improving prediction accuracy and reliability. This invention can predict the probability of sleep disorder risk events, providing a valuable window for proactive intervention and risk prevention.

[0057] See Figure 1 The diagram illustrates a flowchart of a sleep disorder risk prediction and intervention method provided by an embodiment of the present invention. The method includes:

[0058] Understandably, sleep disorder risk events include, but are not limited to, events with an abnormal Apnea-Hypopnea Index (AHI), abnormal heart rate events during sleep, and awakening events.

[0059] Step S101: Collect the user's multimodal time series data.

[0060] It should be noted that multimodal time-series data includes sleep physiological data, environmental data, wakefulness physiological data and behavioral data, as well as health logs.

[0061] In the specific implementation step S101, sleep physiological data of the user during sleep is collected through non-contact sensors, and environmental data of the user during sleep is collected through environmental sensors. Wearable devices are used to collect the user's wakefulness physiological and behavioral data; and health logs input by the user through a terminal are also obtained.

[0062] Understandably, sleep physiological data includes, but is not limited to, respiratory rate, heart rate, and body movement information. Wake-up physiological and behavioral data includes, but is not limited to, heart rate, heart rate variability (HRV), blood oxygen saturation, steps, activity intensity, and stress index, as well as temperature, humidity, CO2 concentration, light intensity, and ambient noise levels.

[0063] Specifically, non-contact sensors, such as frequency-modulated continuous-wave (FMCW) millimeter-wave radar operating at 60 GHz, are typically positioned above or to the side of the user's bedside. They can capture micro-Doppler signals generated by subtle undulations in the user's chest and abdominal cavities with high precision in a non-contact manner. By sequentially performing phase demodulation, filtering, and spectral analysis (such as Fast Fourier Transform, FFT) on these raw signals, the user's respiratory rate and heart rate can be calculated in real time, and the kinetic energy values ​​characterizing body movements or subtle shifts can be extracted.

[0064] Non-contact sensors also include integrated environmental sensors for comprehensive real-time monitoring of the sleep environment. For example, it integrates temperature and humidity sensors as well as a carbon dioxide sensor to accurately measure indoor temperature, humidity, and carbon dioxide concentration. Simultaneously, it monitors light intensity via a photoresistor and uses a microphone to collect ambient sound to assess noise levels (i.e., decibel values).

[0065] It should be noted that the wearable device and the controller of this embodiment of the invention are connected via the standard Bluetooth Low Energy (BLE) protocol. The wearable device continuously collects the user's physiological and activity data through a variety of integrated sensors. On the one hand, based on a photoplethysmography (PPG) sensor, the device can obtain heart rate and blood oxygen saturation data throughout the day, and further analyze them to derive time-domain and frequency-domain indices of heart rate variability characterizing the state of autonomic nervous function. On the other hand, based on an accelerometer, the device can record daytime steps and activity intensity, and monitor body rotation and posture information during sleep to assess the continuity of sleep stages and body movement.

[0066] Understandably, users can supplement their subjective health logs manually through applications on their devices. Specifically, users can easily record key daily behaviors and states by adding tags or entering short text, such as "drinking coffee," "drinking alcohol," "engaging in high-intensity exercise," "high work stress," or "taking medication." This subjective log data effectively complements objective physiological and environmental data from sensors, jointly building a more comprehensive foundation for health analysis.

[0067] Step S102: Process the multimodal time series data and input it into the risk prediction model, and output the probability time series prediction results.

[0068] In step S102, the multimodal time-series data is processed to obtain a target multidimensional feature vector sequence. This target multidimensional feature vector sequence is then input into the risk prediction model, outputting the probability time-series prediction result. The specific implementation process is as follows (processes A1 to A3):

[0069] Process A1: Process the multimodal time series data into a multidimensional feature vector sequence and add position encoding to obtain the target multidimensional feature vector sequence.

[0070] In implementing process A1, since the sampling rates of different data points in the multimodal time series data are different, the multimodal time series data is first cleaned, including removing outliers and resampling to a uniform time resolution (e.g., 1 minute), and then timestamp alignment is performed. Normalization is then applied to eliminate the influence of different dimensions between features. The processed multimodal time series data is then concatenated into a multidimensional feature vector sequence, and positional encoding is added to obtain the target multidimensional feature vector sequence.

[0071] Process A2: Input the target multidimensional feature vector sequence into the risk prediction model. In the risk prediction model, the encoder encodes the target multidimensional feature vector sequence through a multi-head self-attention mechanism to generate a contextual representation.

[0072] Understandably, a risk prediction model is obtained by training an encoder-decoder deep learning model based on a Transformer architecture using multi-dimensional feature vector sequences of sample targets in advance. In this risk prediction model, a multi-head self-attention mechanism is used to process these feature vector sequences, thereby capturing the intrinsic correlations between different modalities and within the same modality over time.

[0073] Specifically, the encoder of this risk prediction model consists of multiple stacked encoder layers with identical structures. Each encoder layer contains two core components: a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism is the key to the model. It can calculate the correlation weights between any two time points and between any two feature dimensions in the input multi-dimensional feature sequence. Through this mechanism, the model can automatically uncover deep dependencies across time and modalities.

[0074] For example, the model can learn autonomously how much the "high pressure index recorded at 3 pm" contributes to predicting "reduced heart rate variability at 2 am"; and what the correlation pattern is between "continuously rising indoor carbon dioxide concentration at night" and "decreased respiratory rate stability".

[0075] This capability enables risk prediction models to not only see data at a single point in time, but also to understand long-term, complex causal relationships, thereby making more accurate forward-looking risk predictions.

[0076] Process A3: The decoder in the risk prediction model decodes the context representation in an autoregressive manner to obtain the probability time series prediction results and outputs the probability time series prediction results.

[0077] Understandably, the decoder of this risk prediction model also adopts a multi-layered stacked structure. The core task of the decoder is to receive the contextual representation from the encoder, which has been integrated with historical information, and to gradually deduce the prediction for a future period of time (such as the next hour) in an autoregressive manner.

[0078] Specifically, at each time step, the decoder predicts the state of the next time step based on the generated partial sequence and the encoder's output, and so on in a cyclical manner until a complete future sequence is generated.

[0079] The final output layer of the risk prediction model consists of a fully connected network followed by a sigmoid activation function, which maps the decoding results to probability values ​​between 0 and 1. Therefore, the final output of the model is a probability sequence P = {pt+1, pt+2, ..., pt+n}, which is the probability time series prediction result.

[0080] In other words, probabilistic time series prediction results are used to characterize the likelihood of a user experiencing a sleep disorder risk event within a preset time window in the future. For example, in a sequence of length 60, the Nth point represents the predicted probability of a specific risk event (such as an AHI index exceeding 15) occurring in the Nth minute in the future.

[0081] Step S103: Generate intervention instructions based on the probability time series prediction results and execute the intervention instructions.

[0082] It should be noted that the process of generating intervention instructions based on probability time series prediction results specifically includes (processes B1 to B3):

[0083] Process B1: Calculate the comprehensive risk score corresponding to the probability time series prediction results based on the multi-criteria decision fusion logic.

[0084] In process B1, based on the probabilistic time series prediction results P = {pt+1 , pt+2 , ... , pt+n}, a comprehensive risk score is calculated through multi-criteria decision fusion logic.

[0085] Among them, multiple criteria include at least:

[0086] Peak intensity analysis: Extracting the highest instantaneous predicted probability of a risk event occurring within a future prediction time window T (e.g., 60 minutes). ;

[0087] Persistence analysis: Calculate the cumulative duration or area under the curve for which the probability exceeds the clinical threshold;

[0088] Evolution slope analysis: Analyzing the rate of risk escalation using the first derivative;

[0089] Volatility stability analysis: assess the confidence level of the prediction using standard deviation or information entropy.

[0090] The formula for calculating the comprehensive risk score is as follows:

[0091] (1)

[0092] In formula (1), the weights The weight of each calculation item (peak, average load, slope, fluctuation) in the final comprehensive score is determined using a dual-track system of "static benchmark + dynamic evolution". This represents the area under the predicted probability curve within a future time window T, divided by time, which is the average predicted probability within that window. It is equivalent to calculating the arithmetic mean of the probability sequences. It represents the absolute maximum value of the first derivative of the predicted probability sequence over time, characterizing the fastest rate at which the probability increases or decreases. It represents the standard deviation of the predicted probability sequence within the time window T.

[0093] Process B2: Map the comprehensive risk score to a preset risk level system to obtain the risk level.

[0094] Understandably, the comprehensive risk score is mapped to a predefined risk level system, which is divided into three levels, each with a clearly defined judgment logic:

[0095] Low risk (Risk_Score < 0.3): This corresponds to physiological parameters fluctuating within the baseline range; the predicted probability is determined to be consistently below 30% within the next hour. This means the system considers the likelihood of a risk event occurring in the short term to be extremely low.

[0096] Medium risk (0.3 ≤ Risk_Score < 0.7): Corresponds to mild to moderate clinical sleep disorder; judged as having a predicted probability of exceeding the preset threshold within the next 30 minutes with a greater than 50% likelihood. This indicates that the risk is accumulating, requiring attention and potential preventative intervention.

[0097] High risk (Risk_Score ≥ 0.7): Corresponds to severe sleep apnea event or acute cardiovascular risk; it is determined that there is a greater than 80% probability that the predicted probability will exceed the preset threshold within the next 10 minutes, or the system has detected a risk event occurring in real time. This means that the risk is imminent or has already occurred, requiring immediate intervention.

[0098] Process B3: Retrieve the intervention strategy corresponding to the risk level from the intervention strategy library to generate intervention instructions.

[0099] It should be noted that intervention instructions include low-risk intervention instructions, medium-risk intervention instructions, and high-risk intervention instructions.

[0100] Among them, low-risk intervention instructions mainly include information prompts and environmental fine-tuning, providing gentle guidance in a non-invasive manner.

[0101] Information prompt: A low-intrusion reminder (such as vibration mode) will be pushed to the user's mobile app, for example: "The system predicts that your respiratory load may increase in the second half of tonight. It is recommended that you adjust to a side-lying sleeping position now."

[0102] Environmental fine-tuning: Link with smart home devices to perform minor adjustments, such as controlling a smart speaker to play soft, soothing music, or adjusting a night light to a warm, low brightness.

[0103] The medium-risk intervention instructions mainly include significant environmental adjustments and changes in the user's physiological posture, in order to prevent the risk from escalating through proactive intervention.

[0104] Significant environmental regulation: Send instructions to the environmental control system to execute specific adjustments, such as: slightly lowering the room temperature by 1-2 degrees Celsius; or activating the fresh air system to increase ventilation and reduce carbon dioxide concentration.

[0105] User physiological posture adjustment: If the user is using a smart bed, a command will be sent to gradually raise the head of the bed by 10-15 degrees to improve upper airway patency by adjusting the body position.

[0106] High-risk intervention commands mainly include user wake-up alarms and emergency contact notifications, which interrupt risk events in an immediate and effective manner.

[0107] User wake-up alarm: Immediately sends a strong vibration and visual alarm through the user's personal device (such as a smart bracelet / watch) to remind the user to take timely countermeasures (such as wearing a ventilator).

[0108] Emergency contact notification: Simultaneously send alarm information to preset emergency contacts (such as family members) via cloud service. The content includes a summary of the user's current status and real-time location, so as to initiate manual assistance when necessary.

[0109] Understandably, this tiered intervention system follows the principle of "no disturbance for low-risk cases, early prevention for medium-risk cases, and emergency response for high-risk cases," ensuring the continuity of users' sleep while effectively responding to different levels of risk.

[0110] In some specific embodiments, a closed-loop optimization method based on the coupling of safety-constrained reinforcement learning and lightweight adaptive fine-tuning is employed to continuously and safely personalize the risk prediction model and intervention strategies. This process is triggered after the intervention command is executed and specifically includes: collecting the user's physiological data after the intervention and constructing formatted empirical data based on the intervention command and physiological data; updating the strategies in the intervention strategy library based on an offline reinforcement learning framework according to the empirical data, and outputting the updated intervention strategy library. Furthermore, based on the empirical data, the parameters of the risk prediction model are updated using a low-rank adaptive method to obtain the latest risk prediction model.

[0111] It should be noted that the system continuously records changes in key physiological indicators for a period of time (e.g., 30 minutes) after the intervention. For example, after performing a medium-risk intervention such as "raising the head of the bed by 15 degrees," the system will focus on analyzing whether the user's respiratory rate stability has improved and whether the model-predicted AHI risk probability curve shows a downward trend. From this, the system generates a structured empirical triplet, which is then encrypted and stored.

[0112] Once such experiential data accumulates to a certain scale, the system can analyze it within a reinforcement learning framework. Within this framework, through algorithms (such as offline reinforcement learning), the system continuously learns from historical experience, autonomously discovering and iteratively generating the most effective personalized intervention strategy sequence for the current user, thus updating the strategies in the intervention strategy library. Simultaneously, the accumulated "state-action-outcome" triplet data provides valuable supervisory signals for the risk prediction model itself. This post-intervention physiological data of users can be used for periodic offline fine-tuning of the risk prediction model, making its prediction logic more closely aligned with the user's individual physiological response patterns, thereby continuously improving the foresight and accuracy of risk warnings.

[0113] It is understandable that the above content includes the following four core mechanisms:

[0114] 1. To ensure absolute safety of human intervention and avoid random trial and error on users, a conservative Q-learning framework is adopted for strategy optimization.

[0115] Basic strategy pre-training: An initial intervention strategy model is trained using a historical offline dataset (containing multimodal data, intervention instructions, and subsequent physiological responses). This strategy learns which intervention will maximize long-term benefits given a risk prediction.

[0116] Safety action screening: In the real-time closed loop, the system only allows an action to be executed if the expected benefit of the candidate intervention is significantly higher than the preset safety threshold and the potential risk is below an acceptable range. This establishes a safe trial-and-error boundary that conforms to medical ethics.

[0117] 2. The optimization process aims to maximize overall benefits, rather than simply improving a single physiological indicator. The physiological-psychological-environmental multidimensional reward function R is defined as:

[0118] (2)

[0119] In formula (2), (Improved blood oxygen saturation) and (Improved heart rate variability) is a positive reward item, reflecting a positive physiological response. (Intervention-induced sleep fragmentation) and (Noise from the actuator interfering with sleep) is a negative penalty to ensure the comfort of the intervention and overall sleep quality.

[0120] 3. To address the impracticality of full parameter fine-tuning for large Transformer models on resource-constrained devices (such as mobile phones and embedded hardware), a low-rank adaptive technique is employed. This technique injects a few attention weight matrices from the Transformer layers of the original risk prediction model into a low-rank decomposition matrix for fine-tuning, while freezing all other parameters. Each optimization requires only generating and uploading parameter increments of a few hundred KB, eliminating the need to upload raw physiological data, significantly saving bandwidth and enhancing privacy protection.

[0121] 4. To address the "cold start" problem during the period of sparse new user data, a meta-learning mechanism is introduced. Specifically, the system pre-trains meta-models covering various typical population groups (such as "obese elderly with OSA" and "stress-induced insomnia in young and middle-aged adults"). In the initial stage of new user usage (e.g., the first 3 hours), the system analyzes their preliminary physiological characteristics and uses meta-learning algorithms to quickly adapt the general model to the most closely related user profile sub-model, achieving "second-level personalization" of intervention strategies and significantly improving the initial experience.

[0122] To better explain the sleep disorder risk prediction and intervention method provided in the embodiments of the present invention, two specific embodiments will be described in detail below.

[0123] Example 1: Specific system implementation scheme for users who have been diagnosed with moderate to severe obstructive sleep apnea or have high-risk factors (such as obesity, hypertension, daytime sleepiness, etc.).

[0124] 1. System hardware configuration and data acquisition:

[0125] The system hardware configuration in this embodiment is designed to achieve comprehensive and seamless data acquisition.

[0126] Non-contact sensor: An industrial-grade millimeter-wave radar sensor is used. This sensor integrates the antenna into the package, simplifying hardware design. It is installed on the ceiling directly above the user's bed, approximately 2.0-2.5 meters from the mattress surface. The radar operates in FMCW mode, with the following configuration parameters:

[0127] Starting frequency: 60 GHz. Sweep bandwidth: 1.75 GHz. Chirp period: 40 μs. Frame period (FramePeriodicity): 50 ms (i.e., sampling rate 20 Hz). Processing flow: The onboard DSP performs range FFT and Doppler FFT on the raw ADC data to generate a range-Doppler heatmap. Phase information is extracted from the region of interest defined at the user's chest cavity location. The phase information is then unwound and bandpass filtered (e.g., 0.1-0.5 Hz for respiratory frequency band and 0.8-2.0 Hz for heart rate frequency band) to finally obtain high signal-to-noise ratio respiratory and heart rate waveforms.

[0128] Wearable Device: The system connects to the user's smart bracelet via a Bluetooth gateway. This bracelet was chosen because it offers a relatively open data export interface and continuous nighttime blood oxygen monitoring. The system retrieves the following data periodically via API (e.g., every 15 minutes): Daytime data: Heart rate per minute, body battery (a comprehensive indicator reflecting stress and recovery), total daily stress score, steps, and intensity minutes. Nighttime data: Average and minimum blood oxygen saturation during sleep, sleep stages (provided by the Garmin algorithm for reference), and resting heart rate.

[0129] Environmental Sensors: Construct a standalone, battery-powered wireless environmental monitoring node, placed on the user's bedside table. Used to measure carbon dioxide concentration, temperature, and humidity. Microphone: Used to collect audio signals, but does not upload the raw audio to protect privacy; only calculates the A-weighted sound pressure level (dBA) locally on the node, and uploads the average noise level once per minute.

[0130] In the accompanying terminal app, a "Daily Log" function is designed. Users can spend one minute before bed recording key events of the day by selecting tags, such as: {"Dinner Time": "21:00", "Did you drink alcohol?": "Yes", "Alcohol Content": "500ml beer", "Did you engage in strenuous exercise?": "No", "Subjective Stress Level": "High"}. These tags will be converted into quantifiable feature vectors.

[0131] Data processing and risk prediction model; the core of this embodiment lies in a finely designed Transformer-based risk prediction model.

[0132] 2. Data preprocessing and feature engineering:

[0133] (1) Time alignment and resampling: All data were resampled to a 1-minute time resolution. High-frequency data (such as respiratory and heart rate) were taken as the mean, standard deviation, maximum / minimum values ​​within 1 minute; low-frequency data (such as data pulled from API) were forward-filled.

[0134] (2) Feature extraction:

[0135] Nighttime physiology: Extract respiratory rate and respiratory amplitude variability from respiratory waveforms; extract time-domain and frequency-domain indicators of heart rate and heart rate variability from heartbeat waveforms.

[0136] Daytime physiology and behavior: Body electrical charge and stress score are used as direct features. Steps and intensity minutes are mapped to daytime activity levels (low, medium, high).

[0137] Environment: directly use temperature, humidity, carbon dioxide concentration, and noise level in decibels.

[0138] User tags: One-hot encoding of tags.

[0139] (3) Sequence construction: All features are concatenated into a multi-dimensional feature vector. Based on the current time point, the feature vector sequence of the past 24 hours (i.e., 1440 time points) is extracted to form the input X_in of the model.

[0140] Risk prediction model:

[0141] Model objective: To predict the probability of a "severe sleep apnea event" occurring every minute within the next 60 minutes. Here, a "severe sleep apnea event" is defined as: apnea / hypopnea lasting more than 30 seconds, as determined by millimeter-wave radar signals, accompanied by a decrease in blood oxygen saturation exceeding 4% as monitored by the wristband.

[0142] Model structure:

[0143] Embedding layer: The input X_in, which is 1440 x D_features (feature dimension), is mapped to a high-dimensional representation of 1440 x D_model (model dimension, e.g., 128) through a fully connected layer.

[0144] Encoder: A 6-layer stacked encoder is used. Each layer contains an 8-head self-attention mechanism. This structure enables the model to learn extremely complex temporal dependencies. For example, the model may find a strong correlation between the two events, {"dinner time later than 9 pm"} and {"drinking alcohol"}, which occurred 8-10 hours ago, and {"significantly increased risk of sleep apnea during deep sleep at 3-4 am"}.

[0145] Decoder: Also employs a 6-layer stacked decoder. The decoder's input, in addition to the encoder's output, includes a "start token" and the generated prediction sequence (used during training by the Teacher Forcing mechanism). It generates a 60 x 1 probability sequence P_out in an autoregressive manner.

[0146] Model Training: The model needs to be pre-trained on a large-scale, precisely labeled dataset. This dataset should contain multimodal data from thousands of obstructive sleep apnea patients, simultaneously collected from PSG, millimeter-wave radar, wearable devices, environmental and behavioral logs. The training loss function can be a binary cross-entropy loss.

[0147] 3. Closed-loop intervention strategy: The core of the intervention strategy is forward-looking, graded and personalized.

[0148] Risk level assessment:

[0149] Low risk (early warning level): The peak probability of the prediction in P_out for the next 30-60 minutes is between 30% and 50%.

[0150] Medium risk (intervention level): The peak probability of prediction in P_out for the next 10-30 minutes exceeds 50%.

[0151] High risk (alarm level): The peak predicted probability in P_out for the next 10 minutes exceeds 80%, or the real-time monitoring module has detected the event.

[0152] Tiered intervention implementation: Low-risk intervention: The system pushes a gentle notification via mobile APP: "Health Tip: Based on your daytime activities and current sleep status, the system predicts that you will enter a high-risk period for sleep apnea in 1 hour. It is recommended that you try to adjust to a side-lying position now."

[0153] Intervention for medium-risk situations: The system sends instructions to the user's smart adjustable bed frame via the Matter protocol or cloud-to-cloud API to slowly and silently raise the head position by 12 degrees. At the same time, it sends instructions to the air purifier to turn on the maximum fan speed mode to enhance indoor air circulation and reduce carbon dioxide concentration.

[0154] High-risk intervention: Immediately emit the strongest continuous vibration via the user's wristband. Simultaneously, make a voice call to the user's pre-set emergency contacts (such as spouse or children) via cloud communication services (such as Twilio), broadcasting a pre-set alarm message: "Warning, [User Name]'s sleep monitoring system has detected a high-risk breathing event. Please check immediately."

[0155] Closed-loop feedback and optimization: The system records each intervention. For example, after implementing "elevating the head of the bed by 12 degrees," the system analyzes whether the respiratory waveform monitored by the radar becomes more regular and the amplitude increases, and whether the blood oxygen saturation monitored by the wristband recovers within the next 30 minutes. This data, along with the intervention actions, is used in an intervention strategy optimization model based on a multi-armed slot machine algorithm. This model models the "success rate" of each intervention measure under different states (such as different sleep stages and different risk levels), thereby dynamically adjusting the priority of intervention strategies and achieving personalized optimization.

[0156] Example 2: A specific system implementation scheme for users with chronic insomnia who may also have cardiovascular risks such as nocturnal palpitations and tachycardia.

[0157] 1. System hardware configuration and data acquisition:

[0158] Non-contact monitoring unit: The configuration is the same as in Example 1, but the algorithm focuses on the accurate extraction of heart rate and the number of heartbeats per unit time. It extracts beat-to-beat intervals (RR intervals) through higher temporal resolution and more refined signal processing algorithms.

[0159] Wearable device interface: This interface connects the system to the user's smart bracelet. It obtains the following key data:

[0160] Daytime data: atrial fibrillation history, high / low heart rate notifications, walking heart rate variability, and aerobic fitness estimates.

[0161] Nighttime data: wrist temperature changes during sleep, heart rate, and the number of heartbeats per unit time.

[0162] Environmental sensors: configured as in Example 1.

[0163] User data recording module: The "Daily Log" function of the APP, in addition to the tags in Example 1, also adds tags related to emotions and cognition, such as: {"Daytime mood": "Anxiety", "Whether to meditate": "Yes", "Duration": "15 minutes", "Screen time before bed": ">1 hour"}.

[0164] 2. Data Processing and Risk Prediction Model:

[0165] Data preprocessing and feature engineering:

[0166] Feature Extraction: Nighttime Physiology: Based on the beat-by-beat RR interval (i.e., the time interval between adjacent heartbeats) sequence, the following parameters are calculated: heart rate per minute, standard deviation of normal sinus RR interval, root mean square of the difference between adjacent RR intervals, percentage of heartbeats with an adjacent RR interval difference greater than 50 milliseconds, and frequency domain parameters including very low frequency power, low frequency power, high frequency power, and the ratio of low frequency to high frequency power. These indicators can precisely reflect the balance state of the autonomic nervous system.

[0167] Daytime physiology and behavior: The AFib history and high / low heart rate notifications from smart bracelets were identified as strong risk characteristics. Wrist temperature variation trends were used as an indicator of circadian rhythm stability.

[0168] Other data processing is the same as in Example 1.

[0169] Risk prediction model:

[0170] Model objective: To predict the probability of a "nocturnal heart rate abnormality event" occurring within the next 30 minutes. Here, an "event" is defined as: a heart rate exceeding 100 bpm or falling below 40 bpm for more than 1 minute, or a sharp decrease in the standard deviation of the normal sinus RR interval (e.g., below 2 standard deviations from the normal baseline).

[0171] Model capabilities: The model can learn, for example, multimodal features such as {"daytime emotional anxiety"} + {"prolonged screen use before bed"} + {"disrupted wrist temperature rhythm at night"}, which collectively point to the complex pattern of {"difficulty transitioning from light sleep to deep sleep between 1-2 am, accompanied by excessive activation of the sympathetic nervous system, leading to an increased risk of tachycardia"}.

[0172] 3. Closed-loop intervention strategy:

[0173] The intervention strategy in this embodiment focuses more on soothing and regulating the autonomic nervous system.

[0174] Risk level assessment:

[0175] Low risk (soothing level): The probability of abnormal heart rate in the next 30 minutes is predicted to be between 40% and 60%.

[0176] Medium risk (adjusted): The probability of an abnormal heart rate in the next 15 minutes exceeds 60%.

[0177] High risk (wake-up / alarm level): Predicts an abnormal heart rate probability exceeding 85% within the next 5 minutes, or detects an event in real time.

[0178] Implementation of tiered intervention:

[0179] Low-risk intervention: The system connects to the user's smart speaker via Wi-Fi and plays clinically validated pink noise or binaural beats audio at a very low volume (e.g., 15-20 dB). Studies have shown that this type of audio helps induce relaxation and promote slow-wave sleep.

[0180] Medium-risk intervention: The system activates an air purifier fan, switching to a gentle, natural wind-simulating mode while slightly lowering the room temperature by 0.5-1°C. This slight temperature drop and airflow have been shown to help lower core body temperature and promote sleep. Simultaneously, the intelligent lighting system adjusts the color temperature of all light sources in the bedroom (if on) to the warmest red light (<2000K) and reduces brightness to the lowest possible level to minimize the suppression of melatonin secretion by blue light.

[0181] High-risk intervention: The smart bracelet provides gentle but clear tactile feedback and a soft alert tone, while the app pushes a notification: "Your heart rate seems a little high. We suggest you try the 4-7-8 deep breathing exercise." If the heart rate remains abnormal, the high-risk alarm process is upgraded to the same as in Example 1, and emergency contacts are notified.

[0182] Closed-loop feedback and optimization:

[0183] The system evaluates the effectiveness of each soothing or modal intervention. For example, does the ratio of low-frequency to high-frequency heart rate power decrease after playing pink noise (indicating reduced sympathetic activity)? Does the user fall into deep sleep more quickly after lowering the room temperature? This feedback is used to optimize intervention strategies; for instance, the system might find that, for a particular user, playing binaural beats is more effective than pink noise in lowering their nighttime heart rate.

[0184] In this invention, significant innovations have been achieved in the field of sleep health management. First, by transforming from a "recorder" to a "predictor," the core function can predict the probability of sleep disorder risk events occurring within a future period, providing a valuable time window for proactive intervention and risk prevention. Second, a systematic, full-cycle health view is constructed, integrating multimodal data such as nighttime sleep, daytime activities, environmental factors, and user subjective feelings to understand the root causes of individual sleep problems from a more comprehensive perspective. Furthermore, intelligent and personalized proactive closed-loop intervention is achieved, adopting different intervention measures based on changes in risk level and continuously optimizing strategies through a closed-loop feedback mechanism to ensure personalized management. Finally, a temporal deep learning model based on advanced architectures such as Transformer is employed to effectively capture the complex temporal dependencies between multimodal data, significantly improving the accuracy and reliability of predictions. In summary, this invention provides a forward-looking risk prediction and automated intervention method and system for sleep health management, possessing personalized self-optimization capabilities, and offering a new solution for improving sleep health.

[0185] Corresponding to the sleep disorder risk prediction and intervention method provided in the above embodiments of the present invention, see also... Figure 2 The diagram shows a structural block diagram of a sleep disorder risk prediction and intervention device provided in an embodiment of the present invention.

[0186] The device includes: a data acquisition unit 201, a prediction unit 202, and an intervention unit 203.

[0187] The acquisition unit 201 is used to acquire the user's multimodal time series data.

[0188] The prediction unit 202 is used to process multimodal time series data and input it into the risk prediction model, and output the probability time series prediction results.

[0189] Intervention unit 203 is used to generate intervention instructions based on probability time series prediction results and execute the intervention instructions.

[0190] In this invention, significant innovations have been achieved in the field of sleep health management. First, by transforming from a "recorder" to a "predictor," the core function can predict the probability of sleep disorder risk events occurring within a future period, providing a valuable time window for proactive intervention and risk prevention. Second, a systematic, full-cycle health view is constructed, integrating multimodal data such as nighttime sleep, daytime activities, environmental factors, and user subjective feelings to understand the root causes of individual sleep problems from a more comprehensive perspective. Furthermore, intelligent and personalized proactive closed-loop intervention is achieved, adopting different intervention measures based on changes in risk level and continuously optimizing strategies through a closed-loop feedback mechanism to ensure personalized management. Finally, a temporal deep learning model based on advanced architectures such as Transformer is employed to effectively capture the complex temporal dependencies between multimodal data, significantly improving the accuracy and reliability of predictions. In summary, this invention provides a forward-looking risk prediction and automated intervention method and system for sleep health management, possessing personalized self-optimization capabilities, and offering a new solution for improving sleep health.

[0191] Combination Figure 2 The content shown includes multimodal time-series data such as sleep physiological data, environmental data, wakefulness physiological data and behavioral data, as well as health logs;

[0192] The acquisition unit 201 includes: a first acquisition module, a second acquisition module, and an acquisition module.

[0193] The first acquisition module is used to collect sleep physiological data of the user during sleep through non-contact sensors and to collect environmental data of the user during sleep through environmental sensors.

[0194] The second data acquisition module is used to collect the user's physiological and behavioral data when the user is awake using wearable devices.

[0195] The acquisition module is used to acquire the health logs entered by the user through the terminal.

[0196] Combination Figure 2 The content shown, prediction unit 202, includes: processing module, encoding module and decoding module.

[0197] The processing module is used to process multimodal time series data into a multidimensional feature vector sequence and add position encoding to obtain the target multidimensional feature vector sequence.

[0198] The encoding module is used to input the target multidimensional feature vector sequence into the risk prediction model. In the risk prediction model, the encoder encodes the target multidimensional feature vector sequence through a multi-head self-attention mechanism to generate a contextual representation.

[0199] The decoding module is used in the risk prediction model to decode the context representation in an autoregressive manner, obtain the probability time series prediction results, and output the probability time series prediction results.

[0200] Combination Figure 2 The content shown, intervention unit 203, includes: a calculation module, a mapping module and a generation module.

[0201] The calculation module is used to calculate the comprehensive risk score corresponding to the probability time series prediction results based on the multi-criteria decision fusion logic.

[0202] The mapping module is used to map the comprehensive risk score to a preset risk level system to obtain the risk level.

[0203] The generation module is used to obtain the intervention strategies corresponding to the risk level from the intervention strategy library in order to generate intervention instructions.

[0204] Combination Figure 2 The device, as shown, also includes: a construction unit, a first update unit, and a second update unit.

[0205] The building blocks are used to collect physiological data from users after intervention and to build formatted experience data based on intervention instructions and physiological data.

[0206] The first update unit is used by the offline reinforcement learning-based learning framework to update the policies in the intervention policy library based on empirical data and output the updated intervention policy library.

[0207] The second update unit is used to update the parameters of the risk prediction model based on empirical data using a low-rank adaptive method, so as to obtain the latest risk prediction model.

[0208] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0209] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0210] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for predicting and intervening in the risk of sleep disorders, characterized in that, The method includes: Collect users' multimodal time-series data; The multimodal time series data is processed and input into the risk prediction model, and the probabilistic time series prediction results are output. An intervention instruction is generated based on the probability time series prediction results, and the intervention instruction is executed.

2. The method according to claim 1, characterized in that, The multimodal time-series data includes sleep physiological data, environmental data, wakefulness physiological data and behavioral data, as well as health logs; The collected multimodal time-series data of users includes: The system collects sleep physiological data of users during sleep using non-contact sensors and environmental data of users during sleep using environmental sensors. Use wearable devices to collect users' physiological and behavioral data when they are awake; Obtain health logs entered by the user through the terminal.

3. The method according to claim 1, characterized in that, The process of processing the multimodal time series data and inputting it into the risk prediction model to output probabilistic time series prediction results includes: The multimodal time series data is processed into a multidimensional feature vector sequence and position encoding is added to obtain the target multidimensional feature vector sequence; The target multidimensional feature vector sequence is input into the risk prediction model. In the risk prediction model, the encoder encodes the target multidimensional feature vector sequence through a multi-head self-attention mechanism to generate a contextual representation. The decoder in the risk prediction model decodes the context representation in an autoregressive manner to obtain the probability time series prediction result, and outputs the probability time series prediction result.

4. The method according to claim 1, characterized in that, The step of generating intervention instructions based on the probability time series prediction results includes: The comprehensive risk score corresponding to the probability time series prediction result is calculated based on the multi-criteria decision fusion logic. The comprehensive risk score is mapped to a preset risk level system to obtain the risk level; The intervention strategy corresponding to the risk level is obtained from the intervention strategy library to generate an intervention instruction.

5. The method according to claim 4, characterized in that, The method further includes: Collect physiological data of the user after intervention, and construct formatted experience data based on the intervention instructions and the physiological data; Based on the offline reinforcement learning learning framework, the strategies in the intervention strategy library are updated according to the empirical data, and the updated intervention strategy library is output. Based on the empirical data, the parameters of the risk prediction model are updated using a low-rank adaptive method to obtain the latest risk prediction model.

6. A device for predicting and intervening in the risk of sleep disorders, characterized in that, The device includes: The acquisition unit is used to acquire users' multimodal time-series data; The prediction unit is used to process the multimodal time series data and input it into the risk prediction model, and output the probability time series prediction result. An intervention unit is used to generate intervention instructions based on the probability time series prediction results and to execute the intervention instructions.

7. The apparatus according to claim 6, characterized in that, The multimodal time-series data includes sleep physiological data, environmental data, wakefulness physiological data and behavioral data, as well as health logs; The acquisition unit includes: The first acquisition module is used to acquire sleep physiological data of the user during sleep through non-contact sensors and to acquire environmental data of the user during sleep through environmental sensors. The second data acquisition module is used to collect the user's physiological and behavioral data when the user is awake using wearable devices; The acquisition module is used to acquire the health logs entered by the user through the terminal.

8. The apparatus according to claim 6, characterized in that, The prediction unit includes: The processing module is used to process the multimodal time series data into a multidimensional feature vector sequence and add position encoding to obtain the target multidimensional feature vector sequence; The encoding module is used to input the target multidimensional feature vector sequence into the risk prediction model. In the risk prediction model, the encoder encodes the target multidimensional feature vector sequence through a multi-head self-attention mechanism to generate a contextual representation. The decoding module is used by the decoder in the risk prediction model to decode the context representation in an autoregressive manner, obtain the probability time series prediction result, and output the probability time series prediction result.

9. The apparatus according to claim 6, characterized in that, The intervention unit includes: The calculation module is used to calculate the comprehensive risk score corresponding to the probability time series prediction result based on the multi-criteria decision fusion logic; The mapping module is used to map the comprehensive risk score to a preset risk level system to obtain a risk level; The generation module is used to obtain the intervention strategy corresponding to the risk level from the intervention strategy library in order to generate intervention instructions.

10. The apparatus according to claim 9, characterized in that, The device further includes: A construction unit is used to collect the user's physiological data after intervention and to construct formatted experience data based on the intervention instructions and the physiological data; The first update unit is used to update the strategies in the intervention strategy library based on the offline reinforcement learning learning framework, according to the empirical data, and output the updated intervention strategy library. The second update unit is used to update the parameters of the risk prediction model based on the empirical data using a low-rank adaptive method, so as to obtain the latest risk prediction model.

Images