Disease Risk Prediction System Based on End-Sleep Model and Sleep Monitoring Equipment

Abstract

This invention provides a disease risk prediction system based on an end-to-end sleep model and sleep monitoring devices, belonging to the field of intelligent technology. It identifies a list of candidate disease hypotheses based on user static information and pre-defines expected sleep fingerprint drift vectors, expected stimulus response vectors, recommended stimulus types, and stimulus parameters for each disease. Based on effective multimodal physiological signals, it constructs the sleep fingerprint and original disease list for the current night and determines whether the night is a "significantly abnormal night." For "significantly abnormal nights," it obtains the drift consistency degree, marks "diseases to be verified" based on the drift consistency degree, and implements active stimulation to obtain the final personalized prediction level for each disease in the original disease list for the current night. Based on the final personalized prediction level, it determines the final predicted disease level and generates a predicted disease report. This provides a feasible path for the paradigm shift of sleep monitoring devices from passive monitoring to proactive early health screening.

Description

Technical Field

[0001] This invention relates to the field of intelligent technology, specifically to a disease risk prediction system based on end-to-end sleep model and sleep monitoring equipment. Background Technology

[0002] With the rapid popularization of sleep monitoring devices and edge computing technology, and the increasing demand for continuous home health monitoring, traditional methods of disease detection relying on medical devices face significant challenges such as high equipment costs, inability to be used at home, and difficulty in long-term tracking. Existing consumer-grade sleep monitoring technology has gradually evolved from basic sleep staging, heart rate, and blood oxygen statistics to the extraction of local physiological features, but it still cannot achieve comprehensive screening of diseases across multiple systems. The limitation of computing and communication resources on the edge side has become a key bottleneck, seriously affecting the security, timeliness, and user experience of disease prediction.

[0003] Existing systems primarily rely on passively monitored physiological signals (such as heart rate, blood oxygen, and body movement) over a single night or multiple nights, outputting disease risk probabilities through statistical models or black-box neural networks. They lack individualized baseline calibration, and the significant differences in physiological characteristics among users mean that single-night data is highly susceptible to short-term fluctuations such as sleep environment, psychological state, and occasional body movement, leading to numerous false positives. For example, a healthy user might be misdiagnosed as having a high risk of sleep apnea due to frequent tossing and turning on a particular night, while subtle abnormalities in patients with early-stage diseases are overlooked due to individual differences. Passive monitoring can only capture the statistical correlation between physiological signals and diseases, failing to distinguish between causality and coincidence, and requires long-term trend accumulation over weeks or even months to confirm abnormalities, resulting in long cycles and low efficiency. For early or resting functional abnormalities that are not obvious (such as delayed motor response in early neurodegenerative diseases or autonomic nervous system dysregulation in early cardiovascular diseases), passive monitoring struggles to trigger identifiable signals, often missing the optimal intervention window. Therefore, existing passive monitoring models face significant bottlenecks in terms of individualized calibration, causal evidence acquisition, early anomaly detection, and interpretability. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: a disease risk prediction system based on an end-to-end sleep model and sleep monitoring equipment, the system comprising: Hypothesis baseline pre-set unit: Based on user static information, a list of candidate disease hypotheses is defined, and for each disease, the expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type and stimulus parameters are pre-set; Abnormal drift detection unit: Based on effective multimodal physiological signals, construct the sleep fingerprint and original disease list for the night, and determine whether the night is a "significantly abnormal night"; Hypothesis-deductive verification unit: For a “significantly abnormal night”, obtain the degree of drift matching for that night, mark the “disease to be verified” according to the degree of drift matching for that night, and implement active stimulation to obtain the final personalized prediction level of each disease in the original disease list for that night; Risk Decision Reporting Unit: Based on the final personalized prediction level, determine the final predicted disease level and generate a predicted disease report.

[0005] Furthermore, the step of defining a candidate disease hypothesis list based on user static information, and pre-setting an expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type, and stimulus parameters for each disease, includes: By using static user information, we obtain an initial list of candidate disease hypotheses and a list of diseases with the lowest probability. The candidate disease hypothesis list is obtained by fusing the lowest probability disease list and the initial candidate disease hypothesis list. The candidate disease hypothesis list includes the disease name, prior probability and prior risk level. Set up a preset disease pattern knowledge base, traverse the candidate disease hypothesis list, and retrieve the expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type, and stimulus parameters for each disease in the candidate disease hypothesis list.

[0006] Furthermore, the setting of a preset disease pattern knowledge base includes: Step ss1: Collect polysomnography data and electronic health records, including longitudinal diagnostic information, medication history, examination results and demographic information; Step ss2: Filter the clinically labeled patient set corresponding to each disease; Step ss3: Select healthy individuals whose demographic information matches the clinically labeled patient set and use them as the control group; Step ss4: Construct the original drift pattern vector of the clinically labeled patient set corresponding to each disease; and compare it with the background drift pattern vector constructed through the healthy control group to obtain the expected sleep fingerprint drift vector; Step 5: For each disease, set three stimulus types, extract the mean of the response features of each stimulus for each disease and the patient set, and compare it with the mean of the response features of the healthy control group to obtain the expected stimulus response vector. Select recommended stimuli and record stimulus parameters based on the magnitude of the expected stimulus response vector.

[0007] Furthermore, the determination of whether the night is a "significantly abnormal night" includes: For valid multimodal physiological signal data, obtain the sleep fingerprint and original disease list for the night; the original disease list for the night includes the disease name, the original disease probability vector for the night, and the original disease level for the night. Based on the normal sleep fingerprints of the night, a rolling baseline for the night is constructed; if the user is using the system for the first time or if the multimodal physiological signal data collected in the previous Nn nights are invalid, the fingerprint vector of a general healthy population is used to obtain the current rolling baseline; the nightly drift vector is obtained based on the current rolling baseline; and it is determined whether the night is a "significantly abnormal night".

[0008] Furthermore, for "significantly abnormal nights," obtaining the degree of drift matching for that night includes: The nighttime match is set based on the nighttime drift vector and the expected sleep fingerprint drift vector in the preset disease pattern knowledge base, including high match, partial match and no match. For each disease in the candidate disease hypothesis list, a short-term day window is set, and the number of days that each disease is highly consistent, partially consistent, and inconsistent within the short-term day window is recorded. The degree of consistency of the overnight drift for each disease is determined, and the degree of consistency of the overnight drift includes four levels: low, medium, medium-high, and high.

[0009] Furthermore, the labeling of "diseases to be verified" based on the degree of drift matching that night includes: If the drift match is "medium-high" or higher on the night, the corresponding disease will be marked as "positive evidence". It also checks whether the expected stimulus response vector corresponding to the disease marked as "positive evidence" retrieved from the preset disease pattern knowledge base is "no preset"; if it is not "no preset", then the disease is marked as "disease to be verified".

[0010] Furthermore, the implementation of active stimulation yields the final personalized prediction level for each disease in the original disease list for that night, including: For diseases marked as "diseases to be verified", natural stimuli are tried first. If the final verification result is unsuccessful within U consecutive attempts, it is converted to deception stimulation. Then, the stimulation parameters corresponding to the recommended stimulation type are applied to confirm the final verification result. For "diseases to be verified", if active stimulation verification is completed, the drift fit degree for the night is adjusted according to the final verification result to obtain a personalized disease prediction level: if the final verification result is passed, it is upgraded by one level; if the final verification result is failed, it is downgraded by one level; if natural stimulation and deceptive stimulation are performed in succession, the final verification result of the deceptive stimulation shall prevail. If active stimulus verification is not triggered, but the overnight drift concordance is "high" or "medium-high", then the overnight drift concordance will be used; otherwise, the prior risk level of the disease will be maintained. For diseases not in the candidate disease hypothesis list, the personalization level is set to "low"; Obtain the final personalized prediction level for each disease in the original disease list for that night.

[0011] Furthermore, the natural stimuli include: Acquire triaxial acceleration signals and remove noise to obtain the resultant acceleration vector within a fixed time window; when the resultant vector exceeds the upper limit of the resting standard threshold within the fixed time window starting from the resting state, the moment corresponding to the first time the static standard threshold is exceeded within the fixed time window is taken as the start time of turning over; the moment corresponding to the resultant vector being less than or equal to the lower limit of the resting standard threshold is taken as the end time of turning over. If the start and end times of the turn are within the range of QQ, the turn event will be analyzed as a turn event; otherwise, it will be discarded. The baseline period is set according to the start time of rolling over; the response period is set according to the end time of rolling over; during the baseline period and the response period, rolling over response features are extracted and dimensionlessized to obtain the rolling over response feature vector; the result of this natural validation is judged by comparing it with the expected stimulus response vector. If all the natural verification results corresponding to the cumulative U rollover stimuli are passed, the final verification result is passed; otherwise, the deception stimulus is activated.

[0012] Furthermore, the deceptive stimuli include: Recommended stimulus types and parameters are extracted from a pre-defined disease pattern knowledge base, and recommended stimuli are applied during light sleep. During the light sleep period (Fd), a random time point is selected, and stimulation is applied according to the recommended stimulation type and corresponding stimulation parameters based on the sleep monitoring device; stimulation is then applied again after an interval of Sd. The deception stimulus response signal is collected from the baseline and response periods corresponding to the stimulus application time points, and the deception feature is extracted and dimensionless processed, and then compared with the expected stimulus response vector. If the verification corresponding to the cumulative U deception stimuli passes, the final verification result is passed; otherwise, the final verification result is failed.

[0013] Furthermore, the step of determining the final predicted disease level based on the final predicted personalized level and generating a predicted disease report includes: Determine whether the final personalized prediction level matches each disease level in the disease level for that night; if the levels are the same, it is completely consistent; if they differ by one level, it is partially consistent; if they differ by two levels or more, it is seriously inconsistent. For diseases with partially identical disease levels, the higher level is used as the final predicted disease level. For diseases with completely identical disease levels, the higher level is used as the final predicted disease level. For diseases with significantly different disease levels, if the user is in the initial stage of use, the corresponding disease level of that night is used as the final predicted disease level. Otherwise, the corresponding final predicted personalized level is used as the final predicted disease level. The predicted disease levels for each disease are sorted in descending order. Diseases with the highest predicted disease levels are selected as output, and corresponding predicted disease reports are generated.

[0014] Disease risk prediction methods based on end-to-end sleep models and sleep monitoring devices include: Step S1: Based on the user's static information, a list of candidate disease hypotheses is defined, and for each disease, an expected sleep fingerprint drift vector, an expected stimulus response vector, a recommended stimulus type, and stimulus parameters are pre-defined. Step S2: Based on the effective multimodal physiological signals, construct the sleep fingerprint and original disease list for the night, and determine whether the night is a "significantly abnormal night"; Step S3: For “significantly abnormal nights”, obtain the drift matching degree for that night, mark “diseases to be verified” according to the drift matching degree for that night, and implement active stimulation to obtain the final personalized prediction level of each disease in the original disease list for that night; Step S4: Based on the final predicted personalized level, determine the final predicted disease level and generate a predicted disease report.

[0015] This invention provides a disease risk prediction system based on an end-to-end sleep model and sleep monitoring equipment. It has the following beneficial effects: By integrating prior static information, sleep fingerprint drift trend analysis, and active stimulus verification, a closed-loop disease risk prediction system from hypothesis generation to causal confirmation was constructed. First, based on user static information, a list of candidate disease hypotheses including conventional, anomalous, and contradictory pairs is generated, breaking the limitations of traditional reliance solely on statistical probability and enabling proactive capture of rare diseases or atypical manifestations. Second, a large-scale SleepFM model on the edge is used to extract sleep fingerprints and construct a personalized rolling baseline, combined with nightly drift consistency judgment, to stably identify disease-related physiological change trends. Third, based on a dual-mode active verification mechanism of natural challenge and deceptive stimulus, no additional data is required. By obtaining causal response evidence through user cooperation, the specificity of verification and sensitivity of early diagnosis can be significantly improved. At the same time, natural stimuli with zero interference are prioritized, and low-power deceptive stimuli are only activated when they are insufficient, so as to maximize the protection of user sleep quality and compliance. Finally, through the divergence analysis of dual-path prediction (original output of large model and personalized verification results), an interpretable risk level is output, which avoids the black box misjudgment of large model and overcomes the interference of individual baseline drift. It realizes home-based, non-invasive, privacy-preserving, low-power, and high-precision full-domain health early screening, and provides a feasible path for the paradigm shift of sleep monitoring devices from passive monitoring to active diagnosis. Attached Figure Description

[0016] Figure 1 This is an architecture diagram of the disease risk prediction system based on the end-to-end sleep model and sleep monitoring device of the present invention; Figure 2 This is a flowchart illustrating the active stimulation process of the present invention; Figure 3 This diagram illustrates the steps of the disease risk prediction method based on the end-to-end sleep model and sleep monitoring device of the present invention. Detailed Implementation

[0017] 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.

[0018] like Figures 1 to 2 As shown, a disease risk prediction system based on an end-to-end sleep model and sleep monitoring equipment includes: Hypothesis baseline pre-set unit: Based on user static information, a list of candidate disease hypotheses is defined, and for each disease, the expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type and stimulus parameters are pre-set; Abnormal drift detection unit: Based on effective multimodal physiological signals, construct the sleep fingerprint and original disease list for the night, and determine whether the night is a "significantly abnormal night"; Hypothesis-deductive verification unit: For a “significantly abnormal night”, obtain the degree of drift matching for that night, mark the “disease to be verified” according to the degree of drift matching for that night, and implement active stimulation to obtain the final personalized prediction level of each disease in the original disease list for that night; Risk Decision Reporting Unit: Based on the final personalized prediction level, determine the final predicted disease level and generate a predicted disease report.

[0019] Based on user static information, a list of candidate disease hypotheses is defined, and for each disease, a pre-defined expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type, and stimulus parameters are provided, including: When a user first launches the accompanying control terminal (APP), they are guided to complete a structured health information questionnaire. The questionnaire collects fields including: date of birth (for calculating age), gender, current height and weight (for calculating body mass index), ethnic background (e.g., Asian, Caucasian, African, etc.), known history of chronic diseases (e.g., hypertension, diabetes, coronary heart disease, etc.), family medical history of first-degree relatives (clinically labeled diseases in parents, siblings, etc.), a list of medications currently being taken (including prescription and over-the-counter drugs), and lifestyle parameters (smoking status, weekly frequency and amount of alcohol consumption, and weekly frequency and duration of moderate to high-intensity exercise). After the user confirms all the information, it is encrypted and stored in the protected storage area of ​​the accompanying control terminal, generating the user's static information. The "prior risk assessment module" is deployed on the device side by inputting static user information. This module is a lightweight neural network whose parameters have been pre-trained in the cloud using large-scale epidemiological data (such as NHANES, UK Biobank, and other public databases). The module outputs a prior disease probability vector of length 130, where each dimension corresponds to the prior risk probability of a disease (value between 0 and 1). The prior risk probabilities are sorted from high to low, and the top T (usually 20) diseases are selected as the initial candidate disease hypothesis list, with the prior probability of each disease recorded. The remaining Y (usually 5) diseases are selected as the lowest probability disease list, with the prior probability of each disease recorded. Based on the prior probabilities, a corresponding prior risk level is generated: if the prior probability < 0.2, the level is defined as low; if 0.2 ≤ prior probability < 0.4, the level is defined as medium; if 0.4 ≤ prior probability < 0.6, the level is defined as medium-high; and if 0.6 ≤ prior probability ≤ 1, the level is defined as high. In the 130-dimensional prior disease probability vector, the cumulative probability of the first 20 diseases often covers the vast majority (e.g., >80%) of potential risks, which is sufficient to encompass common and high-probability diseases, while avoiding excessive computation in subsequent steps due to an overly long list. The last 5 diseases with the lowest probabilities are selected as "abnormal hypotheses" to proactively introduce rare diseases or atypical manifestations without significantly increasing the list length, preventing the complete neglect of low-probability but high-harm diseases due to statistical bias. T and Y are obtained through multiple rounds of offline data testing, which can achieve a reasonable balance between coverage, computational efficiency, and representativeness of anomalous hypotheses.

[0020] Lightweight Neural Network: Static health information is output and standardized to obtain an encoded feature vector; this vector is then input into a lightweight fully connected feedforward neural network for processing; the network is typically designed with an architecture of 2-3 hidden layers, and after computation by the neural network, the final output is a prior disease probability vector of 130 diseases.

[0021] The candidate disease hypothesis list is obtained by fusing the lowest probability disease list and the initial candidate disease hypothesis list (e.g., [initial candidate disease hypothesis list, lowest probability disease list]). The candidate disease hypothesis list includes the disease name, prior probability, and prior risk level. The system iterates through the candidate disease hypothesis list, retrieves the expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type, and stimulus parameters for each disease in the candidate disease hypothesis list from the preset disease pattern knowledge base, and packages them with the corresponding diseases in the candidate disease hypothesis list and stores them in the matching control terminal.

[0022] The pre-defined disease pattern knowledge base provides a "standard template" for each disease (i.e., expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type and stimulus parameters). Among them, the expected sleep fingerprint drift vector refers to the difference in a patient's sleep physiology relative to their healthy baseline in the months before a clinically labeled disease; The expected stimulus response vector refers to the response characteristics extracted from physiological signals synchronously collected by sleep monitoring equipment when a person is in a specific sleep stage (such as light sleep) to a certain external stimulus (vibration, sound, or flash of light), and the difference between the response characteristics and those of a healthy person.

[0023] Both vectors were obtained through statistical analysis of differences between cases and controls. They are standardized templates at the population level and are used to compare similarity with individual measured data.

[0024] Set up a preset disease pattern knowledge base, including: Step 1: Collect polysomnography (PSG) data (with a collection time of 4 hours or more) and corresponding electronic health records from multiple sleep medicine centers, top-tier hospitals, and public databases (such as the National Sleep Research Resource Center in the United States). The polysomnography data includes physiological characteristics such as electroencephalogram (EEG), electrooculogram (EOG), electromyogram (EMG), electrocardiogram (ECG), respiratory airflow, chest and abdominal movements, blood oxygenation, and body position. Electronic health records contain at least 3 years of longitudinal diagnostic information (ICD code, i.e., diagnosis date), medication history, examination results, and demographic information (age, sex, body mass index). Step ss2: For each disease, select a clinically labeled patient set that meets the following criteria (at least 200 patients for each disease): The date of diagnosis has been determined; At least six consecutive months prior to diagnosis, polysomnography data (at least one night per month) must be available. Follow-up should be conducted for at least 6 months after diagnosis. Exclude patients who were already diagnosed with the disease before the initial diagnosis; Step ss3: Select healthy individuals (without any disease) whose demographic information matches the clinically labeled patient set, and use them as a healthy control group; for example: if a clinically labeled patient set for a certain disease is selected, whose age range is 46 to 50, whose gender is male, and whose body mass index is within the standard range; then the matched healthy individuals are also in the age range of 46 to 50, whose gender is also male, and whose body mass index is also within this standard range. Step 4: Using the trained SleepFM model, take the polysomnography of each patient in the clinically labeled patient set as input and output the polysomnography fingerprint of that night (i.e., the output multidimensional feature vector). Using the diagnosis date as the zero point, three time windows were divided backward (early window (6 to 3 months before clinically marked), mid-term window (3 to 1 month before clinically marked), and recent window (1 month before clinically marked to the diagnosis date)); for the healthy control group, a random date was used as the sham diagnosis date, and the same window division was applied. For each patient, the average value of polysomnography fingerprints within 6 months prior to diagnosis was taken as the patient's "health baseline fingerprint"; for the healthy control group, the polysomnography fingerprints within 6 months prior to the spurious diagnosis date were taken as the "pseudo-health baseline fingerprint". For each patient, the average polysomnography fingerprint (calculated by averaging) in the recent window, intermediate window, and early window is compared with the healthy baseline fingerprint (the difference is calculated) to obtain the drift vectors of the recent window, intermediate window, and early window. Each dimension of each drift vector represents the direction and magnitude of change of each physiological feature (e.g., positive value is an increase, negative value is a decrease). For each disease, calculate the dimension-wise mean of the recent window drift vectors of all patients in the disease patient set, and use it as the original drift pattern vector; Simultaneously, the mean value of the pseudo-drift vector of the healthy control group in the corresponding pseudo-diagnosis recent window is calculated as the background drift pattern vector. For each dimension, the difference between the means of the patient set and the healthy control group is compared to obtain the difference drift vector; a t-test is performed on each dimension of the difference drift vector, and only the dimensions with p < 0.05 are retained, while the rest are set to 0 (indicating that there is no drift in that dimension for the disease). The final difference drift vector is dimensionless (e.g., Z-Score normalization) and used as the expected sleep fingerprint drift vector; The purpose of performing the t-test is to statistically confirm that the differences between the patient group and the healthy control group in each dimension are significant, to avoid the interference of random fluctuations, and to ensure that the selected drift or response features are disease-related and generalizable. Step 5: For different diseases, set three types of stimulation, including vibration stimulation (frequency 20-100Hz, weak / medium / strong intensity, duration 1-5 seconds), auditory stimulation (500Hz or 1000Hz pure tone, volume 30-40 dB, duration 0.5-2 seconds), and light stimulation (green or red LED flashing, frequency 1-2Hz, duration 0.2-1 seconds). For each disease, clinically labeled patients and healthy controls were selected as subjects. Subjects used sleep monitoring devices (such as smartwatches) to fall asleep. When subjects entered a preset sleep stage (such as light sleep N1 / N2 stage), one of the above stimuli was randomly applied, and physiological signals (PPG, acceleration, skin conductance, etc.) were collected simultaneously 20 seconds before and after the stimulation. Each subject completed at least 3 stimulation tests, and different stimulation types were distributed across different sleep nights. After each stimulus test, response features are extracted. For example, using an existing lightweight neural network model, physiological signals collected 20 seconds before and after the stimulus are used as input for response feature extraction. Lightweight neural networks are an existing technology. The specific steps for extracting response features using a lightweight neural network model are as follows: After time alignment and preprocessing of the physiological signals 20 seconds before and after stimulation, they are used as input to the processing layer. The processing layer uses a lightweight temporal convolutional network (TCN) as a feature extractor; temporal features are extracted from the time-aligned and preprocessed physiological signals through causal convolution and dilated convolution to obtain a high-dimensional temporal feature map; The high-dimensional temporal feature map is input into the channel attention mechanism, and then weights are generated through two fully connected networks. The weights are multiplied by the high-dimensional temporal feature map to obtain the weighted feature map. Finally, the weighted feature map is compressed into a fixed-length feature, namely the response feature, through global average pooling.

[0025] For each type of stimulus corresponding to each disease, the mean of the response features extracted from the patient set and the mean of the response features extracted from the healthy control group are compared (the difference is calculated) to obtain the difference feature vector; that is, one disease corresponds to three types of stimulus, and finally three difference feature vectors are generated. A t-test was performed on each dimension of the differential feature vector, retaining only dimensions with p < 0.05 and setting the rest to 0; and dimensionless processing was performed to obtain the "expected stimulus response vector" for each disease under each stimulus type; if a certain disease has no significant difference under all stimulus types (i.e., a t-test was performed on each dimension of the differential feature vector, and p for each dimension was greater than or equal to 0.05), it was marked as "no pre-set". For each disease, the stimulus type with the largest expected stimulus response vector magnitude is selected as the recommended stimulus, and the corresponding stimulus parameters (frequency, intensity, duration) are recorded. The core purpose of implementing vibration, auditory, and light stimulation is to transform disease risk prediction from passively monitoring naturally occurring physiological signals during sleep to actively applying controllable physiological challenges. By measuring the response characteristics of the user's physiological system to these challenges, evidence of causality can be obtained. Passive monitoring can only observe correlations and requires long-term trend accumulation, while active stimulation can expose functional abnormalities that are not obvious in the resting state (such as early neuropathy, motor disorders, etc.), thereby significantly shortening the time from suspicion to confirmation, improving the accuracy and interpretability of predictions, and enabling sleep monitoring devices to achieve proactive early health screening capabilities similar to clinical functional tests.

[0026] SleepFM is an open-source sleep multimodal basic model developed by a team at Stanford University School of Medicine. It can predict the risk of developing more than 130 diseases in the future by analyzing a single night's sleep data. Its research results were officially published in the top journal Nature Medicine in January 2026.

[0027] The specific steps of the existing SleepFM model are as follows: The collected signal data (such as brain activity signals, electrocardiogram, respiratory signals, etc.) were standardized and preprocessed, segmented into 5-second standard segments, and resampled to a uniform frequency (e.g., 128Hz) as input features for the SleepFM model. Simultaneously, a pre-training task was constructed using a self-supervised learning framework, enabling the learning of essential physiological signal features without manual data labeling, thus establishing the foundation for model training. The preprocessed data was then input into the SleepFM model, processed by three 1D CNN encoders for different modalities, and then feature alignment and fusion were performed using a leave-one-out comparison learning framework. Finally, a high-dimensional feature vector was extracted from the penultimate fully connected layer of the model as a sleep fingerprint, ultimately predicting the probability of 130 diseases. This is a pre-trained model; you can simply input the required data later. No further details will be provided.

[0028] Determining whether a night is a "significantly abnormal night" includes: Users (who may or may not have a disease) use sleep monitoring devices during nighttime sleep to collect valid multimodal physiological signal data (e.g., cumulative sleep time of at least 4 hours; otherwise, the collected multimodal physiological signal data is considered "invalid" and will not be included in subsequent analysis). The valid multimodal physiological signal data collected during the valid time period is preprocessed and input into the large sleep model (SleepFM), which outputs the night's sleep fingerprint and the night's original disease list. The night's original disease list includes the disease name, the night's original disease probability vector (130-dimensional probability, with probabilities ranging from 0 to 1), and the night's original disease level. Among them, the nighttime sleep fingerprint vector refers to a high-dimensional feature vector extracted from the penultimate fully connected layer of the SleepFM model after preprocessing the effective multimodal physiological signals of the whole night and inputting them into the SleepFM model; The model directly outputs the disease name and its corresponding probability (which constitutes the original disease probability vector for that night). When the probability is [0, 0.2), the level is defined as low; when the probability is [0.2, 0.4), the level is defined as medium; when the probability is [0.4, 0.6), the level is defined as medium-high; and when the probability is [0.6, 1], the level is defined as high. Based on the normal sleep fingerprints of the night, a rolling baseline for the night is constructed, and the drift vector for the night is obtained. If the multimodal physiological signal data collected by the user for the first time or for the previous Nn nights (Nn is generally 4 to 7, and the tolerance range of 4 to 7 nights can tolerate short-term random data loss and avoid long-term lack of effective data causing the system to fail to work, which is an engineering experience value that is a trade-off between personalization and robustness) are all invalid, then the fingerprint vector of the general healthy population is used to obtain the current rolling baseline. The sleep fingerprint of the night is compared with the rolling baseline of the night. If the values ​​of H dimensions (10% of the total dimensions, which is set based on the expected number of multiple independent dimensions that randomly exceed 2 standard deviations in statistics) are not within the range of the dimensions corresponding to the current rolling baseline, then the sleep fingerprint of the night is judged to be abnormal and the night is marked as "significantly abnormal night". Otherwise, it is judged as "no significant abnormal night". The normal queue is a fixed-length (e.g., 7 nights) first-in-first-out queue specifically used to store the sleep fingerprints of the user that have been determined to be valid multimodal physiological signal data outputs in the recent period. Whenever the sleep fingerprint of a night is found to be normal compared with the current rolling baseline, the sleep fingerprint of that night is added to the tail of the queue. If the queue is full, the earliest sleep fingerprint of the night is automatically removed, so that the queue always contains the sleep fingerprints of the user's recent health status.

[0029] The specific steps for constructing the overnight rolling baseline are as follows: For sleep fingerprints in normal queues, the median (i.e., the mean of each dimension) of each dimension in the sleep fingerprints in normal queues is obtained as the baseline center vector; the standard deviation of each dimension in the sleep fingerprints in normal queues is obtained as the baseline standard deviation vector; a fluctuation coefficient is set, and the baseline fluctuation range is obtained based on the baseline center vector, the baseline standard deviation vector, and the fluctuation coefficient. The baseline fluctuation range is then defined as the rolling baseline for the night. For example, assuming the sleep fingerprint has only 3 dimensions (actually 512 dimensions), and the normal queue contains the fingerprints without anomalies from the last 3 nights: Night 1 [0.85, 0.62, 0.31], Night 2 [0.82, 0.59, 0.33], Night 3 [0.88, 0.64, 0.30]; then the baseline center vector is [0.85, 0.62, 0.31], and the baseline standard deviation vector is [0.03, 0.025, 0.015]; the fluctuation coefficient is 2, set according to empirical methods; then the rolling baseline for that night is [(0.79~0.91), (0.57~0.67), (0.28~0.34)]; If the sleep fingerprint for the night is [0.72, 0.50, 0.35], then the first dimension being outside the range indicates a deviation, the second dimension being outside the range indicates a deviation, and the third dimension being outside the range indicates a deviation. The specific steps for using the general healthy population fingerprint vector to obtain the current rolling baseline are the same as the steps for constructing the rolling baseline for the night, namely, obtaining the general healthy population fingerprint of known healthy users, and then calculating the baseline center vector, baseline standard deviation vector and fluctuation coefficient to obtain the rolling baseline for the healthy users for the night, which serves as the rolling baseline for the user (whether or not they have the disease). The difference between the nighttime sleep fingerprint and the baseline center vector is calculated dimension by dimension and then processed to obtain the nighttime drift vector. When the user's first collection or multimodal physiological signal data collected over several consecutive nights are all invalid, no sleep fingerprint has been obtained. At this time, in order to determine whether the user has significant abnormalities, a general reference benchmark is needed. The general healthy population fingerprint vector is the average sleep fingerprint obtained through statistical analysis of large-scale healthy population data, representing the sleep physiological characteristics of a typical healthy individual. Although this range does not perfectly match the individual differences of users, it can provide a relatively reasonable basis for judging abnormalities during the cold start phase when user data is missing. If the user's sleep fingerprint for that night deviates significantly from the normal range of this healthy population, it indicates that there may be a health risk.

[0030] Among them, the effective multimodal physiological signal data includes photoplethysmography raw waveforms, electrocardiogram signals, triaxial acceleration signals, gyroscope angular acceleration signals, successive heart rate values, successive heartbeat intervals, and blood oxygen saturation; all of which are collected through sleep monitoring equipment. For "significantly abnormal nights," the degree of drift matching for that night is obtained, including: The degree of match for the current night is determined based on the drift vector of the current night and the expected sleep fingerprint drift vector; then the degree of match for the current night is obtained based on the degree of match for the current night. Determine the degree of agreement between the nighttime drift vector and the expected sleep fingerprint drift vector; for example, using the cosine similarity of the two vectors (to measure directional consistency) and the magnitude ratio (the magnitude of the nighttime drift vector divided by the magnitude of the expected sleep fingerprint drift vector, to measure the sufficiency of the magnitude); then determine according to the following rules: If the cosine similarity is ≥0.8 and the modulus ratio is ≥0.6, then it is considered a high match. If the cosine similarity is ≥0.5 and the modulus ratio is ≥0.4, but the high-fit condition is not met, then it is a partial fit. Otherwise, it is considered a mismatch.

[0031] For each disease in the candidate disease hypothesis list, a short-term window R (the three most recent nights, as three consecutive sampling days are sufficient to identify monotonic trends in time series analysis, while avoiding excessively long windows that could delay responses to physiological changes) is set. The degree of fit is recorded each night, and the number of highly fitted days in the short-term window R is denoted as H, the number of partially fitted days as P, and the number of non-fit days as N. Based on the fit within the short-term window, the drift fit for that night is determined; H = R (i.e., the fit within the window). If all nights are highly consistent, then the drift consistency for that night is high; if H=R−1 and P=1 (i.e., only one night is partially consistent, and the rest are highly consistent), then the drift consistency for that night is medium-high; if H≥1 (at least one night is highly consistent) or P≥R / 2 (half or more of the days are partially consistent), and the above "high" or "medium-high" conditions are not met, then the drift consistency for that night is medium; otherwise (i.e., most of the days are not consistent, and the number of highly consistent days is 0, and the number of partially consistent days is less than half), then the drift consistency for that night is low.

[0032] The degree of overnight drift fit is the cumulative trend level of a certain disease in the candidate disease hypothesis list within a short-term window.

[0033] Diseases to be verified are categorized based on the degree of similarity in drift patterns over the night, including: If the drift match reaches "medium-high" or higher (medium-high and high) on the same night, the disease is marked as "positive evidence". It also checks whether the expected stimulus response vector corresponding to the disease marked as "positive evidence" retrieved from the preset disease pattern knowledge base is "no preset"; if it is not "no preset", then the disease is marked as "disease to be verified".

[0034] Active stimulation was implemented to obtain the final personalized prediction level for each disease in the original disease list for that night, including: For diseases marked as "diseases to be verified", natural stimuli are tried first. If the final verification result is unsuccessful within U consecutive attempts, it is converted to deception stimulation. Then, the stimulation parameters corresponding to the recommended stimulation type are applied to confirm the final verification result. Natural stimuli, including: Natural stimuli refer to physiological events that are likely to occur during sleep without actively applying external stimuli, such as turning over. The triaxial acceleration signal at the wrist is collected by the triaxial accelerometer in the smartwatch (sampling rate greater than or equal to 50Hz), and noise is removed (e.g., by sliding median filtering). The resultant acceleration vector is obtained in each fixed time window (e.g., 0.5 seconds). When the resultant vector exceeds the upper limit of the resting standard threshold within the fixed time window, the moment corresponding to the first time the static standard threshold is exceeded in the fixed time window is taken as the start time of turning over. The moment corresponding to the resultant vector being less than or equal to the lower limit of the resting standard threshold is taken as the end time of turning over. Among them, the upper and lower limits of the resting standard threshold are set by the mean and standard deviation of the sum vector within 10 seconds before the start of the rolling over (e.g., the upper limit is the sum of the mean and 3 times the standard deviation; the lower limit is the sum of the mean and 1 time the standard deviation). If the start and end times of rolling over are within the range of Qq (e.g., 2 to 10 seconds, a normal rolling over action usually lasts 2 to 10 seconds from start to finish; less than 2 seconds may be due to local limb twitching or noise interference, and more than 10 seconds may include multiple continuous body movements or abnormal rigidity; therefore, selecting this range can effectively filter out complete and typical rolling over events), then this rolling over event will be analyzed as a rolling over event; otherwise, it will be discarded. A baseline period is set based on the start time of rolling over (e.g., within 5 seconds before the start time); a response period is set based on the end time of rolling over (e.g., within 10 seconds after the start time); during the baseline and response periods, rolling over response features are extracted from multimodal physiological information data and dimensionlessized to obtain a rolling over response feature vector; the expected stimulus response vector is retrieved from a preset disease pattern knowledge base; if the numerical deviations of the corresponding dimensions of the rolling over response feature vector and the expected stimulus response vector are both within ±5%, then the natural validation result is considered passed. If the natural verification results corresponding to the rolling over stimulus are all passed within the set cumulative U times (usually set to 2 times, as a single rolling over may be misjudged due to accidental factors such as semi-awakeness or signal glitch, requiring two consecutive rolling over to pass the verification can significantly improve the confidence of the judgment, while avoiding excessive number of times that would lead to an excessively long verification cycle and delay subsequent decisions), then the final verification result is passed; otherwise, the deceptive stimulus is activated.

[0035] Deceptive stimulation, including: Recommended stimulus types and parameters are extracted from a pre-defined disease pattern knowledge base. Recommended stimuli are applied during light sleep (reason: users are less likely to be awakened during light sleep and their physiological responses are stable). When a sleep monitoring device (such as a smartwatch) determines that a user has entered a light sleep stage, a random time point is selected within the Fd time period of entering the light sleep stage (e.g., 2 or 3 minutes, to ensure that the user has entered a stable light sleep stage and to avoid applying stimulation at the boundary of sleep stages or a short transition period, which may lead to unstable responses or unexpected awakenings). The sleep monitoring device then applies stimulation according to the stimulation parameters corresponding to the recommended stimulation type. After this stimulation, another stimulation is applied at an interval of Sd time period (e.g., 5 or 10 minutes, because based on the recovery time of the human autonomic nervous system and cardiovascular system after a single stimulation (usually the heart rate, skin conductance, etc., can return to baseline within 2 to 5 minutes), which can avoid mutual interference between the previous and subsequent stimulations, ensure that multiple effective tests can be completed within the limited sleep duration, and minimize damage to the sleep structure). Multimodal physiological information data were collected from the baseline and response periods corresponding to the time points of stimulus application. Deception features were extracted and dimensionlessized, and then compared with the expected stimulus response vector. If the numerical deviations of the corresponding dimensions of the deception response feature vector and the expected stimulus response vector are within ±5%, the deception stimulus verification is considered successful; otherwise, the deception stimulus verification is considered unsuccessful. If all the verifications corresponding to the cumulative U deception stimuli (usually set to 2) pass, the final verification result is "pass"; otherwise, the final verification result is "fail".

[0036] The so-called deceptive stimulus is a low-interference verification method used when natural stimuli cannot complete the risk verification. It is called "deceptive" because after the user enters a light sleep stage, an extremely slight, almost imperceptible microstimuli are applied. Without disturbing the user's normal sleep or waking the user, it "tricks" the user's sleep perception state and actively triggers the corresponding physiological response, thereby completing the verification without the user's awareness.

[0037] The specific steps to obtain the final predicted personalization level are as follows: For "diseases to be verified", if active stimulation verification is completed, the drift fit degree for the night is adjusted according to the final verification result to obtain the personalized disease prediction level: if the final verification result is passed, it is increased by one level (if it is the highest level, it is not increased); if the final verification result is failed, it is decreased by one level (if it is the lowest level, it is not decreased); if natural stimulation and deceptive stimulation are performed in succession, the final verification result of the deceptive stimulation shall prevail. If active stimulus verification is not triggered, but the overnight drift concordance is "high" or "medium-high", then the overnight drift concordance will be used; otherwise, the prior risk level of the disease will be maintained. For diseases not in the candidate disease hypothesis list, the personalization level is set to "low"; Obtain the final personalized prediction level for each disease in the original disease list for that night.

[0038] The design advantage of prioritizing natural stimuli and switching to deceptive stimuli only when they are insufficient is as follows: Natural stimuli utilize physiological events that inevitably occur during sleep, such as turning over, without the need for active external stimulation. This results in zero interference with the user's sleep, zero energy consumption, and a response that more closely resembles the actual physiological state. It can obtain high-quality validation evidence with high ecological validity at low cost and high frequency. Only when natural events are insufficient or the response characteristics are not obvious will deceptive stimuli be used for active detection. This ensures the validity of the validation while minimizing active intervention in the user's sleep and energy consumption, thereby improving the user experience and compliance. The final personalized prediction level is a risk level that is personalized and calibrated from the original prediction results of the SleepFM large model. It integrates the user's long-term sleep drift trend, causal evidence verified by active stimulation, and static prior information. It can effectively correct the bias of the large model on individuals, filter out random fluctuations in a single night, and avoid false alarms or missed alarms. As a result, it outputs a more accurate, reliable, and clinically valuable disease risk assessment, enabling users to obtain customized health warnings based on their own physiological status and providing evidence-based priority guidance for subsequent health intervention decisions.

[0039] Based on the final personalized prediction level, the final predicted disease level is determined, and a predicted disease report is generated, including: Determine whether the final personalized prediction level matches each disease level in the disease level for that night; if the levels are the same, it is completely consistent; if they differ by one level (e.g., one is high and the other is medium), it is partially consistent; if they differ by two levels or more, it is seriously divergent. For diseases with partially identical disease levels, the higher level is used as the final predicted disease level. For diseases with completely identical disease levels, the higher level is used as the final predicted disease level. For diseases with significantly different disease levels, if the user is in the initial stage of use, the corresponding disease level of that night is used as the final predicted disease level. Otherwise, the corresponding final predicted personalized level is used as the final predicted disease level. The predicted disease level for each disease is sorted in descending order (high, medium-high, medium, low). Diseases with a high predicted disease level are selected as output, and a corresponding predicted disease report is generated for health intervention (such as going to the hospital for examination).

[0040] The sleep monitoring device transmits effective multimodal physiological signals to the matching control terminal in real time via Bluetooth; the control terminal stores the data and uploads it to the cloud via Wi-Fi / cellular network; the cloud-based sleep model performs inference and returns a sleep fingerprint and original disease list; the control terminal combines local rolling baseline and active stimulation verification to finally generate a predicted disease report; if deception stimulation is required, the control terminal controls the wearable device to execute it via Bluetooth command; throughout the entire process, the control terminal is the core hub connecting the sleep monitoring device and the cloud, and the data does not go directly from the sleep monitoring device to the cloud to ensure privacy and stability; Sleep monitoring devices include: smart wearable devices, smartwatches, smart bracelets, smart sleep rings, smart chest patches, smart ankle bracelets, smart sleep pillows, smart sleep headbands, smart sleep eye masks, smart sleep mattresses, smart sleep pads, sleep monitoring belts, sleep monitoring mats, smart beds, millimeter-wave radar sleep monitors, non-contact sleep monitoring devices, and sleep environment monitoring devices. One or more of these devices can be used to effectively collect multimodal physiological signals.

[0041] First, candidate disease hypotheses are pre-defined based on user static information, and a priori benchmark and pre-set disease pattern knowledge base matching individuals are built. Then, through the anomaly drift detection unit, "significantly abnormal nights" with potential abnormalities are quickly screened out, completing the first step of anomaly identification. Afterwards, for the anomalies screened out in the initial screening, the hypothesis deduction verification unit uses natural stimuli and low-interference deception stimuli for active verification, which solves the bottleneck of passive monitoring's inability to verify and high false alarm rate. Finally, through the risk decision reporting unit, an early screening report is generated. The entire process not only realizes the paradigm upgrade from passive monitoring to proactive health early screening, but also takes into account the privacy protection of the device and the user's sleep experience.

[0042] like Figure 3 As shown, a disease risk prediction method based on an end-to-end sleep model and sleep monitoring equipment is presented. The method includes: Step S1: Based on the user's static information, a list of candidate disease hypotheses is defined, and for each disease, an expected sleep fingerprint drift vector, an expected stimulus response vector, a recommended stimulus type, and stimulus parameters are pre-defined. Step S2: Based on the effective multimodal physiological signals, construct the sleep fingerprint and original disease list for the night, and determine whether the night is a "significantly abnormal night"; Step S3: For “significantly abnormal nights”, obtain the drift matching degree for that night, mark “diseases to be verified” according to the drift matching degree for that night, and implement active stimulation to obtain the final personalized prediction level of each disease in the original disease list for that night;

[0043] Step S4: Based on the final predicted personalized level, determine the final predicted disease level and generate a predicted disease report.

[0044] In this embodiment, a list of candidate disease hypotheses, including conventional, anomalous, and contradictory pairs, is generated using static user information. This breaks through the limitations of traditional methods that rely solely on statistical probability and can proactively capture rare diseases or atypical manifestations. Secondly, a sleep fingerprint is extracted using the edge-side SleepFM large model, and a personalized rolling baseline is constructed. Combined with the nightly drift consistency judgment, the trend of disease-related physiological changes is stably identified. Thirdly, based on the dual-mode active verification mechanism of natural challenge and deceptive stimulus, causal response evidence can be obtained without user cooperation, significantly improving verification specificity. At the same time, the use of zero-interference natural stimulus is prioritized, and low-power deceptive stimulus is only activated when it is insufficient, maximizing user sleep quality and compliance. Finally, through the divergence analysis of dual-path prediction (original output of the large model and personalized verification results), an interpretable risk level is output and health intervention is carried out. This avoids the black-box misjudgment of the large model and overcomes the interference of individual baseline drift, realizing home-based, non-invasive, privacy-secure, low-power, and high-precision full-domain health early screening. This provides a feasible path for the paradigm shift of intelligent sleep monitoring devices from passive monitoring to active health early screening.

[0045] This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code that, when executed by the one or more processors, can perform the disease risk prediction system based on the end-to-end sleep model and sleep monitoring device as described above.

[0046] The methods and systems according to the embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or hard disk, may store the disease risk prediction system based on an end-to-end sleep model and sleep monitoring device provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components of the electronic device shown in this application may be omitted according to actual needs.

[0047] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover 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 limitations, an element defined by the phrase "comprising a reference structure" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A disease risk prediction system based on an end-to-end sleep model and sleep monitoring equipment, characterized in that, The system includes: Hypothesis baseline pre-set unit: Based on user static information, a list of candidate disease hypotheses is defined, and for each disease, the expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type and stimulus parameters are pre-set; Abnormal drift detection unit: Based on effective multimodal physiological signals, construct the sleep fingerprint and original disease list for the night, and determine whether the night is a "significantly abnormal night"; Hypothesis-based verification unit: For a "significantly abnormal night", obtain the drift consistency of that night, mark the "disease to be verified" according to the drift consistency of that night, and implement active stimulation to obtain the final personalized prediction level of each disease in the original disease list for that night; Risk Decision Reporting Unit: Based on the final personalized prediction level, determine the final predicted disease level and generate a predicted disease report.

2. The disease risk prediction system based on end-to-end sleep model and sleep monitoring equipment according to claim 1, characterized in that, The process of defining a candidate disease hypothesis list based on user static information, and pre-assigning an expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type, and stimulus parameters for each disease, includes: By using static user information, we obtain an initial list of candidate disease hypotheses and a list of diseases with the lowest probability. The candidate disease hypothesis list is obtained by fusing the lowest probability disease list and the initial candidate disease hypothesis list. The candidate disease hypothesis list includes the disease name, prior probability and prior risk level. Set up a preset disease pattern knowledge base, traverse the candidate disease hypothesis list, and retrieve the expected sleep fingerprint drift vector, expected stimulus response vector, recommended stimulus type, and stimulus parameters for each disease in the candidate disease hypothesis list.

3. The disease risk prediction system based on end-to-end sleep model and sleep monitoring equipment according to claim 2, characterized in that, The setting of a preset disease pattern knowledge base includes: Step ss1: Collect polysomnography data and electronic health records, including longitudinal diagnostic information, medication history, examination results and demographic information; Step ss2: Filter the clinically labeled patient set corresponding to each disease; Step ss3: Select healthy individuals whose demographic information matches the clinically labeled patient set and use them as the control group; Step ss4: Construct the original drift pattern vector of the clinically labeled patient set corresponding to each disease; and compare it with the background drift pattern vector constructed through the healthy control group to obtain the expected sleep fingerprint drift vector; Step 5: For each disease, set three stimulus types, extract the mean of the response features of each stimulus for each disease and the patient set, and compare it with the mean of the response features of the healthy control group to obtain the expected stimulus response vector. Select recommended stimuli and record stimulus parameters based on the magnitude of the expected stimulus response vector.

4. The disease risk prediction system based on end-to-end sleep model and sleep monitoring equipment according to claim 3, characterized in that, The determination of whether a night is a "significantly abnormal night" includes: For valid multimodal physiological signal data, obtain the sleep fingerprint and original disease list for the night; the original disease list for the night includes the disease name, the original disease probability vector for the night, and the original disease level for the night. Based on the normal sleep fingerprints of the night, a rolling baseline for the night is constructed; if the user is using the system for the first time or if the multimodal physiological signal data collected in the previous Nn nights are invalid, the fingerprint vector of a general healthy population is used to obtain the current rolling baseline; the nightly drift vector is obtained based on the current rolling baseline; and it is determined whether the night is a "significantly abnormal night".

5. The disease risk prediction system based on end-to-end sleep model and sleep monitoring device according to claim 4, characterized in that, For "significantly abnormal nights," the degree of drift matching for that night is obtained, including: The nighttime match is set based on the nighttime drift vector and the expected sleep fingerprint drift vector in the preset disease pattern knowledge base, including high match, partial match and no match. For each disease in the candidate disease hypothesis list, a short-term day window is set, and the number of days that each disease is highly consistent, partially consistent, and inconsistent within the short-term day window is recorded. The degree of consistency of the overnight drift for each disease is determined, and the degree of consistency of the overnight drift includes four levels: low, medium, medium-high, and high.

6. The disease risk prediction system based on end-to-end sleep model and sleep monitoring device according to claim 5, characterized in that, The labeling of "diseases to be verified" based on the degree of drift matching that night includes: If the drift match is "moderate to high" or higher on the night of the event, the corresponding disease will be marked as "positive evidence". It also checks whether the expected stimulus response vector corresponding to the disease marked as "positive evidence" retrieved from the preset disease pattern knowledge base is "no preset"; if it is not "no preset", then the disease is marked as "disease to be verified".

7. The disease risk prediction system based on end-to-end sleep model and sleep monitoring equipment according to claim 6, characterized in that, The implementation of active stimulation yields the final personalized prediction level for each disease in the original disease list for that night, including: For diseases marked as "diseases to be verified", natural stimuli are tried first. If the final verification result is unsuccessful within U consecutive attempts, it is converted to deception stimulation. Then, the stimulation parameters corresponding to the recommended stimulation type are applied to confirm the final verification result. For "diseases to be verified", if active stimulation verification is completed, the drift fit degree for the night is adjusted according to the final verification result to obtain a personalized disease prediction level: if the final verification result is passed, it is upgraded by one level; if the final verification result is failed, it is downgraded by one level; if natural stimulation and deceptive stimulation are performed in succession, the final verification result of the deceptive stimulation shall prevail. If active stimulus verification is not triggered, but the overnight drift concordance is "high" or "medium-high", then the overnight drift concordance will be used; otherwise, the prior risk level of the disease will be maintained. For diseases not in the candidate disease hypothesis list, the personalization level is set to "low"; Obtain the final personalized prediction level for each disease in the original disease list for that night.

8. The disease risk prediction system based on end-to-end sleep model and sleep monitoring equipment according to claim 7, characterized in that, The natural stimuli include: Acquire triaxial acceleration signals and remove noise to obtain the resultant acceleration vector within a fixed time window; when the resultant vector exceeds the upper limit of the resting standard threshold within the fixed time window starting from the resting state, the moment corresponding to the first time the static standard threshold is exceeded within the fixed time window is taken as the start time of turning over; the moment corresponding to the resultant vector being less than or equal to the lower limit of the resting standard threshold is taken as the end time of turning over. If the start and end times of the turn are within the range of QQ, the turn event will be analyzed as a turn event; otherwise, it will be discarded. The baseline period is set according to the start time of rolling over; the response period is set according to the end time of rolling over; during the baseline period and the response period, rolling over response features are extracted and dimensionlessized to obtain the rolling over response feature vector; the result of this natural validation is judged by comparing it with the expected stimulus response vector. If all the natural verification results corresponding to the cumulative U rollover stimuli are passed, the final verification result is passed; otherwise, the deception stimulus is activated.

9. The disease risk prediction system based on end-to-end sleep model and sleep monitoring device according to claim 8, characterized in that, The deceptive stimuli include: Recommended stimulus types and parameters are extracted from a pre-defined disease pattern knowledge base, and recommended stimuli are applied during light sleep. During the light sleep period (Fd), a random time point is selected, and stimulation is applied according to the recommended stimulation type and corresponding stimulation parameters based on the sleep monitoring device; stimulation is then applied again after an interval of Sd. The deception stimulus response signal is collected from the baseline and response periods corresponding to the stimulus application time points, and the deception feature is extracted and dimensionless processed, and then compared with the expected stimulus response vector. If the verification corresponding to the cumulative U deception stimuli passes, the final verification result is passed; otherwise, the final verification result is failed.

10. The disease risk prediction system based on end-to-end sleep model and sleep monitoring device according to claim 9, characterized in that, The process of determining the final predicted disease level based on the final predicted personalized level and generating a predicted disease report includes: Determine whether the final personalized prediction level matches each disease level in the disease level for that night; if the levels are the same, it is completely consistent; if they differ by one level, it is partially consistent; if they differ by two levels or more, it is seriously inconsistent. For diseases with partially identical disease levels, the higher level is used as the final predicted disease level. For diseases with completely identical disease levels, the higher level is used as the final predicted disease level. For diseases with significantly different disease levels, if the user is in the initial stage of use, the corresponding disease level of that night is used as the final predicted disease level. Otherwise, the corresponding final predicted personalized level is used as the final predicted disease level. The predicted disease levels for each disease are sorted in descending order. Diseases with the highest predicted disease levels are selected as output, and corresponding predicted disease reports are generated.

Images