A dynamic early warning method and system for bipolar disorder risk
By collecting multi-source time-series data to construct an individualized baseline set and decoupling multi-scale features, combined with a pre-trained model for early warning, and calibrated through clinical feedback, the problems of individual differences and closed-loop feedback in existing technologies have been solved, achieving early and accurate early warning of bipolar disorder.
Patent Information
- Application Number
- CN202611279507.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-21
- Publication Date
- 2026-11-17
AI Technical Summary
Existing bipolar disorder early warning methods fail to effectively adapt to individual differences, ignore physiological rhythms, cannot distinguish between short-term transient disturbances and medium- to long-term trend drifts, lack closed-loop feedback between early warning results and actual clinical outcomes, and cannot adapt to changes in patients' brain neurons.
Multi-source time-series monitoring data are collected to construct an individualized baseline set. Short-term transient disturbances and medium- to long-term trend drift features are decoupled and separated through multi-scale features. Early warning is provided by combining a pre-trained biphasic emotion phase transition probability model and adaptive calibration is performed through clinical feedback.
It enables personalized early and accurate warnings, reduces the rate of missed and false alarms, improves the accuracy and adaptability of warnings, and supports long-term adaptive iteration.
Smart Images

Figure CN123120300A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical information technology, specifically to a method and system for dynamic early warning of bipolar disorder risk. Background Technology
[0002] Bipolar disorder is a severe mental illness characterized by recurrent episodes of alternating manic and depressive phases. Its neurobiological basis involves abnormal functional connectivity of the prefrontal cortex-limbic system emotion regulation circuit and amygdala overactivation. These abnormalities in brain structure and function lead to instability of the emotion homeostasis regulation mechanism, making the "normal-onset" phase transition significantly unpredictable. Therefore, early identification and timely intervention are key clinical needs for reducing relapse rates and suicide risk.
[0003] With the development of digital psychiatry, existing monitoring programs are mainly divided into two categories: one is online follow-up tools that rely solely on patients' self-rating mood scales, and the other is physiological and behavioral time-series monitoring systems based on wearable devices. Some programs attempt to integrate the two types of data to conduct emotional risk assessment. These technical approaches attempt to indirectly infer the central emotional state from peripheral physiological signals, but the dynamic changes in the brain's emotion regulation network have extremely strong individual differences.
[0004] Therefore, existing early warning methods still have some shortcomings. For example, existing early warning systems uniformly use fixed thresholds based on population statistics to determine abnormalities, ignoring the strong individual heterogeneity of the physiological rhythms and emotional baselines of bipolar patients; existing technologies directly model the raw time-series data without performing multi-scale decoupling processing on the time-series signals, making it impossible to distinguish between short-term transient disturbances and medium- to long-term trend drifts in the brain's emotional functional state that are about to transition phases; most existing mental and emotional monitoring models borrow from the monitoring approach for depression, without conducting targeted modeling for phase transitions in bipolar disorder, resulting in a lack of closed-loop feedback between early warning results and actual clinical outcomes, and an inability to adapt to long-term physiological baseline drifts caused by changes in the patient's brain neural processes. Therefore, there is an urgent need for a dynamic early warning method and system for relapse risk that adapts to the pathogenesis of phase alternation in bipolar disorder, can accommodate individual patient differences, suppress subjective self-evaluation noise, distinguish between transient stress and prodromal pathological signals, and supports long-term adaptive iteration. Summary of the Invention
[0005] In order to overcome the above-mentioned deficiencies of the prior art, embodiments of the present invention provide a dynamic early warning method and system for bipolar disorder risk, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a dynamic early warning method for the risk of bipolar disorder, comprising: S1: Collect multi-source time-series monitoring datasets of target patients, including physiological time-series data, behavioral time-series data and emotional self-rating time-series data, and perform time alignment of multi-source data. Calculate data quality scores based on the data time alignment results. S2: Construct an individualized baseline set based on the historical time-series monitoring data aligned with the target patient, including the homeostatic physiological behavior baseline and the emotional phase time-series baseline, and update the individualized baseline set by sliding according to a preset window mechanism; S3: Calculate the deviation between the aligned multi-source time-series monitoring data and the current individualized baseline set, and obtain the deviation sequence by combining the data quality score. Decouple the deviation sequence by multi-scale features, separate the short-term transient disturbance features and medium-to-long-term trend drift features of the multi-source data, and obtain the global feature vector. S4: Input the global feature vector into the pre-trained biphasic emotion phase transition probability model and output the target patient's current emotion recognition state and the dynamic risk index of transitioning to the manic or depressive phase; S5: Based on the dynamic risk index and the trend of risk changes, it performs graded early warning according to threshold rules and outputs the early warning information to the terminal; S6: Based on the early warning information, clinical feedback is confirmed, and the parameters of the individualized baseline set and the bipolar emotion phase transition probability model are updated in a closed loop according to the labeled data confirmed by the feedback, and adaptive calibration is performed.
[0007] Preferably, a dynamic early warning system for the risk of bipolar disorder includes: Multi-source temporal monitoring data acquisition module: Collects multi-source temporal monitoring datasets of target patients, including physiological temporal data, behavioral temporal data and emotional self-rating temporal data, and performs time alignment of multi-source data, and calculates data quality scores based on the data time alignment results; Asynchronous dual-baseline architecture construction module: Based on the historical time-series monitoring data aligned with the target patient, an individualized baseline set is constructed, including the steady-state physiological behavior baseline and the emotional phase time-series baseline, and the individualized baseline set is updated by sliding according to a preset window mechanism; Multi-scale sentiment feature decoupling module: The alignment of multi-source time-series monitoring data and the current individualized baseline set are used to calculate the deviation, and the deviation sequence is obtained by combining the data quality score. Multi-scale feature decoupling is performed on the deviation sequence to separate the short-term transient disturbance features and medium- and long-term trend drift features of multi-source data and obtain the global feature vector. Emotional phase transition risk inference module: Input the global feature vector into the pre-trained biphasic emotional phase transition probability model, and output the target patient's current emotional recognition state and the dynamic risk index of transitioning to manic or depressive phases; Dynamic risk warning module: Based on the dynamic risk index and the trend of risk changes, it provides graded warnings according to threshold rules and outputs the warning information to the terminal; Event-triggered closed-loop calibration module: Based on early warning information, clinical feedback is confirmed, and the parameters of the individualized baseline set and the bipolar emotion phase transition probability model are updated in a closed loop according to the confirmed annotation data, and adaptive calibration is performed.
[0008] The technical effects and advantages of this invention are as follows: 1. This invention collects stable time series samples of patients during remission through an individualized baseline set construction step, and independently fits parameters such as the mean, standard deviation, and daily cosine rhythm of emotions of the specific baseline. It abandons the universal fixed threshold for the whole population and avoids the problem of using a unified population statistical threshold to judge time series abnormalities. It uses the patient's own stable state as the benchmark for judging abnormalities, thereby reducing the false alarms and missed warnings caused by individual differences. 2. This invention decouples and separates two independent time-series components—short-term perturbation high-frequency reconstruction signal and medium-to-long-term trend low-frequency reconstruction signal—through a multi-scale emotional feature decoupling step. These components are then input into the subsequent model in a layered manner, solving the problem of existing solutions directly using unlayered original time-series modeling. This achieves the isolation of short-term random perturbation noise from medium-to-long-term rhythmic drift signals that reflect the prodromal characteristics of disease onset. 3. This invention constructs a five-layer emotional latent state space of the bipolar disease course through the emotional phase transition risk output step and a pre-trained bipolar emotional phase transition probability model. This solves the problem of not conducting targeted modeling for phase transition in bipolar disorder, improves the early weak rhythm shift in the transition from remission to manic and depressive prodromal states, and achieves earlier and more accurate prospective early warning of bipolar relapse. 4. This invention uses a feedback standard data adaptive calibration step and a hierarchical early warning event push mechanism to receive clinical doctors' follow-up visit labels. It adopts a lightweight incremental iteration method to hierarchically and adaptively update the complete set of parameters of the individualized steady-state baseline and the parameters of the biphasic phase-transfer HMM model, thereby solving the problem of the lack of closed-loop feedback between early warning results and actual clinical outcomes and improving the early warning accuracy in long-term operation. Attached Figure Description
[0009] Figure 1 This is a schematic diagram of the overall process of the present invention.
[0010] Figure 2 This is a schematic diagram of the method flow of the present invention.
[0011] Figure 3 This is a schematic diagram of the global feature vector output process of the present invention. Detailed Implementation
[0012] 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.
[0013] Please see Figure 1 As shown, the present invention provides a dynamic early warning system for the risk of bipolar disorder, including a multi-source time-series monitoring data acquisition module, an asynchronous dual-baseline architecture construction module, a multi-scale emotional feature decoupling module, an emotional phase shift risk inference module, a dynamic risk early warning module, and an event-triggered closed-loop calibration module.
[0014] The multi-source time-series monitoring data acquisition module is connected to the asynchronous dual-baseline architecture construction module. The multi-scale emotion feature decoupling module is connected to both the asynchronous dual-baseline architecture construction module and the emotion phase transfer risk inference module. The dynamic risk warning module is connected to both the emotion phase transfer risk inference module and the event-triggered closed-loop calibration module. The event-triggered closed-loop calibration module is connected to both the asynchronous dual-baseline architecture construction module and the emotion phase transfer risk inference module.
[0015] Multi-source temporal monitoring data acquisition module: Collects multi-source temporal monitoring datasets of target patients, including physiological temporal data, behavioral temporal data and emotional self-rating temporal data, and performs time alignment of multi-source data, and calculates data quality scores based on the data time alignment results; Asynchronous dual-baseline architecture construction module: Based on the historical time-series monitoring data aligned with the target patient, an individualized baseline set is constructed, including the steady-state physiological behavior baseline and the emotional phase time-series baseline, and the individualized baseline set is updated by sliding according to a preset window mechanism; Multi-scale sentiment feature decoupling module: The alignment of multi-source time-series monitoring data and the current individualized baseline set are used to calculate the deviation, and the deviation sequence is obtained by combining the data quality score. Multi-scale feature decoupling is performed on the deviation sequence to separate the short-term transient disturbance features and medium- and long-term trend drift features of multi-source data and obtain the global feature vector. Emotional phase transition risk inference module: Input the global feature vector into the pre-trained biphasic emotional phase transition probability model, and output the target patient's current emotional recognition state and the dynamic risk index of transitioning to manic or depressive phases; Dynamic risk warning module: Based on the dynamic risk index and the trend of risk changes, it provides graded warnings according to threshold rules and outputs the warning information to the terminal; Event-triggered closed-loop calibration module: Based on early warning information, clinical feedback is confirmed, and the parameters of the individualized baseline set and the bipolar emotion phase transition probability model are updated in a closed loop according to the labeled data confirmed by the feedback, and adaptive calibration is performed; Please see Figure 2 As shown, a dynamic early warning method for bipolar disorder risk includes: S1: collecting a multi-source time-series monitoring dataset of the target patient, including physiological time-series data, behavioral time-series data, and emotional self-rating time-series data, and performing time alignment of the multi-source data, calculating a data quality score based on the data time alignment result; S2: constructing an individualized baseline set based on the aligned historical time-series monitoring data of the target patient, including a steady-state physiological and behavioral baseline and an emotional phase time-series baseline, and updating the individualized baseline set by sliding according to a preset window mechanism; S3: calculating the deviation between the aligned multi-source time-series monitoring data and the current individualized baseline set, and obtaining a deviation sequence by combining the data quality score, and then... The deviation sequence is decoupled using multi-scale features to separate short-term transient perturbation features from medium- and long-term trend drift features of multi-source data, resulting in a global feature vector; S4: The global feature vector is input into a pre-trained biphasic emotion phase transition probability model, which outputs the current emotion recognition state of the target patient and a dynamic risk index for transitioning to manic or depressive phases; S5: Based on the dynamic risk index and the trend of risk changes, a graded warning is issued according to threshold rules, and the warning information is output to the terminal; S6: Clinical feedback confirmation is performed based on the warning information, and the individualized baseline set and the parameters of the biphasic emotion phase transition probability model are updated in a closed loop according to the confirmed labeled data, and adaptive calibration is performed.
[0016] S1: Collect multi-source time-series monitoring datasets of the target patients, including physiological time-series data, behavioral time-series data, and emotional self-rating time-series data, and perform time alignment of the multi-source data. Calculate a data quality score based on the data time alignment results, including: S1.1: Collect multi-source time-series monitoring dataset X(t)=[x1(t),x2(t),...,x D (t)] T D represents the dimension of the monitoring data at time t, and x represents the dimension of the data. d X(t) represents the d-th monitoring data at time t, where d = 1, 2, ..., D; the multi-source time-series monitoring data X(t) includes objectively collected physiological time-series data, behavioral time-series data, and patient emotional self-assessment time-series data input as an auxiliary input; the physiological time-series data reflects the autonomic nervous function and physiological arousal level of the target user, including but not limited to heart rate variability (HRV), skin conductance, sleep time series, and body surface temperature; the behavioral time-series data reflects the continuous monitoring time-series data related to the target user's daily movement and circadian rhythm, including but not limited to exercise intensity, activity rhythm, and diurnal activity data; the emotional self-assessment time-series data x... MOOD(t), used to reflect the target user's subjective perception of their own emotional state, with daily synchronous collection of patients' self-rating emotion scale scores as auxiliary temporal input (international biphasic follow-up digital tools (BDTrack, moodfit, etc.) all incorporate self-rating emotion scales, which are standard auxiliary input modalities); heart rate variability characteristics x HRV (t) The RR interval sequence between consecutive heartbeats is calculated using a preset sliding window Δt1 (5-minute sliding window, 1-minute sliding step), and the standard deviation is calculated to obtain the skin electrical activity characteristics (EDA). t Using an electrical conductance analysis (EDA) sensor, a preset sliding window Δt2 is used to extract the average SCL (baseline level of electrical conductance) and the number of SCR peaks per unit time (electrical conductance response peaks), thus obtaining x. EDA (t); Sleep time sequence includes the effective sleep duration time sequence SD at night. t (Total effective sleep duration per night (excluding time spent awake in bed at night), in minutes), frequency and timing of nighttime awakenings (WC) t (Total number of times you wake up during the night after falling asleep), Sleep Efficiency Time Sequence (SE) t (Effective sleep duration ÷ Total nighttime bed rest); Body surface temperature data x TEM (t) Continuously collect body surface temperature x by placing a body surface temperature sensor in close contact with the skin of the wrist. TEM (t); Exercise intensity data x AIt (t) A continuous time-series index characterizing the intensity of human limb activity per unit time, calculated from the root mean square amplitude of triaxial acceleration within a unit sliding window collected by an accelerometer from a wearable device; activity rhythm data x ARh (t) is constructed by time series of exercise intensity throughout the day, and the mean of all exercise intensity in each time period is taken as the segmented vector dimension to characterize the periodic distribution characteristics of activity levels in different time periods of the day; diurnal activity data x SSc (t) includes the first time to fall asleep, the last time to wake up in the morning, and the timestamp of the sleep-wake cycle, representing the patient's sleep-wake cycle pattern; the value range of the self-rating time series data is [-10, +10], with negative values representing a depressed state, positive values representing a manic state, and 0 points representing a stable and relieved state. The self-rating data is processed as follows: In the data of the manic prodrome and manic episode, the self-rating is unreliable due to the lack of insight and is marked as "missing in the episode", and risk assessment is based solely on objective physiological and behavioral data; the self-rating data of the depressive episode has relatively complete insight and is retained and used in subsequent data processing; In this embodiment, it should be specifically noted that the target patient refers to a patient who has been clinically diagnosed with bipolar disorder or a high-risk group assessed as having the risk of an episode. The sampling frequency of each dimension of data can be independently configured according to the capabilities of the data source device and clinical monitoring needs. In a preferred embodiment, the sampling frequency of heart rate data is 1 / 60Hz, skin conductance is 0.1Hz, and self-assessment data is collected by the mobile terminal application at preset time points (such as 08:00 and 20:00 daily) through push reminders. Physiological time-series data and behavioral time-series data can be flexibly selected according to the hardware configuration of the wearable device. Those skilled in the art can select any continuously collected time-series data that reflects the patient's autonomic nervous system state and daily activity rhythm. For example, physiological time-series indicators may also include data such as blood oxygen saturation, respiratory rate, and amount of sweating on the skin surface. The behavioral activity time-series indicators also include data such as duration of outdoor activities, duration of mobile phone use, and daily cumulative steps.
[0017] S1.2: Imputation processing for missing values in objective time-series data: When the missing duration is less than or equal to the first preset threshold (e.g., consecutive missing ≤3 sampling points), linear interpolation is used for imputation; when the missing duration is greater than the first preset threshold and less than or equal to the second preset threshold (e.g., consecutive missing 4-10 sampling points), forward imputation is used, i.e., the last valid value before the missing value appears, until a new valid value appears; when the missing duration is greater than the second preset threshold (e.g., consecutive missing >10 sampling points), the data for that period is marked as unusable and not included in subsequent processing; for data with diurnal periodicity (e.g., daily variations in sleep time series, activity rhythms, etc., usually with a preset diurnal period of 24 hours), the average of adjacent days is used for imputation first. If adjacent day data is also missing, it reverts to the aforementioned linear interpolation or forward imputation; since the patient's emotional state may affect their willingness to respond to self-rating (lack of motivation during depression, inattention during mania), when x at a certain moment t MOOD When (t) is missing, a surrogate estimate of x is constructed using existing objective physiological and behavioral data. MOOD g (t)=f pr (X obj (t),Θ pr ), X obj (t) is an objective data vector, f pr (·) represents the surrogate estimation model trained during the patient's stable period (used to map objective data to the emotion rating space; if the patient is not in a stable period, the surrogate model is not enabled, and missing emotion self-rating data in this window is directly marked as "unknown" and not involved in subsequent data processing). Θ pr The parameter set of the surrogate estimation model; the surrogate estimation model f pr Construction and updating of (·): First, training data: based on the patient's stable period T ref The data within is used to construct the training sample set D.pr tr ={X obj (t),x MOOD (t)|t∈T ref ,x MOOD (t) Not missing}, and Z-score standardize each dimension of the objective data to x. d,st Then, the model is constructed: Lasso regression (minimum absolute shrinkage and selection operator) is used first to obtain the surrogate estimate. D1 represents the dimension of the objective data, and the model parameter Θ is... pr ={β0,β1,...,β D1 ,λ} is obtained by minimizing the objective function minβ: N is the number of training samples, λ is the regularization coefficient, and the optimal value is selected through 5-fold cross-validation. X i obj,st The value is the standardized objective data of the i-th sample; then the confidence level is assessed: the proxy inference value x. MOOD g Confidence level η of (t) pr Defined as: η pr (t)=exp(-|x MOOD g (t)-x MOOD la (t)| / σ re )×min[1,max(0,(1-(σ pr ) 2 / (σ re ) 2 )],x MOOD la (t) represents the most recent self-rated true emotion value, σ re Let σ be the standard deviation of the residuals of the surrogate model on the training set. pr ) 2 The prediction variance estimated by the Bootstrap method (the Bootstrap method is a resampling statistical technique that randomly divides the training data into multiple different subsets (randomly sampling 80% of the samples each time), trains a surrogate model on each subset, thereby obtaining multiple different models, resulting in multiple prediction results, and calculating the prediction variance), (σ re ) 2 Let η be the variance of the self-reported variance during the stable period. pr When (t) > the corresponding preset value, the proxy inference is available; otherwise, the self-evaluation data of emotion at time t is marked as "unknown" and does not participate in subsequent data processing. S1.3: Unify all data to the same baseline sampling interval Δt. The rule for selecting the baseline interval is as follows: traverse all original sampling frequencies of physiological, behavioral, and emotional dimensions, select the lowest frequency dimension that is clinically effective as the baseline dimension, and set the inherent sampling period of this dimension as the baseline interval Δt (in this embodiment, the sleep / activity rhythm diurnal scale derived hourly interval is selected, and Δt = 1 hour is set); perform downsampling processing on high-frequency data (heart rate, skin conductance, HRV, etc.) with sampling frequencies higher than the baseline interval, and calculate the mean of high-frequency data within the baseline sampling interval Δt; For low-frequency data with sampling frequencies lower than the baseline interval (such as daily cumulative steps, daily effective sleep duration, etc.), time-limited backtracking and upsampling are performed: starting from the low-frequency data update time (the update time usually refers to the timestamp when the raw sensor data for the whole day is summarized and the daily aggregated data is calculated), values are only assigned to the earlier continuous baseline sampling intervals that do not exceed 24 hours. Baseline sampling intervals that exceed 24 hours are uniformly marked as missing data (low-frequency data subdivision filling rules: when the daily effective sleep duration is filled to the baseline interval △t sampling point, it is divided into daily averages to obtain each baseline interval). The equivalent sleep duration at intervals of Δt, the frequency of nighttime awakenings, and sleep efficiency are copied in their entirety and assigned to the baseline sampling point without equal division. The activity rhythm data is copied in its entirety as a segmented vector for the entire day and assigned to the baseline sampling point. The rules for filling circadian rhythm data are as follows: if the baseline sampling interval falls within the sleep period of the day, the corresponding sleep / awake state is marked and a corresponding timestamp is bound; if the sampling interval does not fall within the sleep period of the day (sleep period refers to the time from the first time one falls asleep to the last time one wakes up in the morning), the circadian rhythm timestamp field is marked as missing, and missing data is not included in the subsequent baseline mean. (Calculation of standard deviation and other data). If the sampling time is earlier than the first low-frequency record, it is uniformly filled with null values initially. Sentiment self-report data is mapped to the corresponding sampling interval according to the actual reporting time. When there are multiple self-report records in the same window, the average value is taken. If there is no self-report data, it is filled in according to the surrogate estimation model. If the surrogate estimation is unavailable, it is marked as "unknown" for the self-report data in that sampling interval. Finally, all dimensions are aligned on the same time axis, and each dimension generates one data point every Δt, forming a unified multidimensional time series dataset X(Δt)=[x1(Δt),x2(Δt),...,x D (△t)] T ; S1.4: Define the data quality score Q(t)∈[0,1]: Q(t) = scores of all dimensions SC d The sum / total number of dimensions D, if within the baseline sampling interval Δt, the proportion of missing original sampling points in dimension d is greater than or equal to the missing threshold (e.g., 50%), then the labeled data is missing, and SC is used. d =0, if the missing percentage is less than the missing threshold within the baseline sampling interval Δt, it indicates that data exists, SC d =max(0, 1 - the proportion of missing original sampling points in the d-th dimension); S2: Construct an individualized baseline set based on the historical time-series monitoring data aligned with the target patient, including a homeostatic physiological behavior baseline and an emotional phase time-series baseline, and update the individualized baseline set by sliding according to a preset window mechanism, including: S2.1: In the initial stage after system startup (e.g., the first 6 weeks), collect time-aligned multidimensional time-series data D_in={X(t)} of a set length T_in. t=1 T_in For each time t, calculate the multidimensional stable period score SI(t). , z k (t) represents the k-th stability index, including the standard deviation of the self-rating emotion score, the coefficient of variation obtained by combining the standard deviation and mean of exercise intensity, the coefficient of variation obtained by combining the standard deviation and mean of total effective sleep duration, and the sliding coefficient of variation obtained by combining the standard deviation and mean of heart rate variability, median(z) k ) represents the median of the initial stage, MAD(z) k )=median(|z k -median(z k )|) represents the absolute deviation of the median, K SI The total number of stability indices (4 in this scheme) indicates greater stability; the larger the value, the more stable the system. Candidate stable periods are identified based on the following condition: SI(t) > corresponding threshold SI. th (SI) th For adjustable thresholds, such as SI th =0.3), duration △T ≥ the set minimum stable period T min,sta (e.g., 14 days) and the duration ΔT has no warning trigger records (no warning in the initial stage, the condition is automatically met). If there are multiple candidate stable periods, the one with the largest average SI among the multiple candidate stable periods is selected as the reference stable period T. ref In the initial stage after system startup, the surrogate estimation model f will not be enabled. pr (·), missing emotional self-rating values are directly marked as "unknown". After the initial screening of T_ref, f is immediately trained. pr (·) and backtrack to fill in the missing emotional values from the initial stage, recalculating SI(t) to confirm T. ref stability; S2.2: Based on the reference stable period T ref Based on the data, an individualized baseline set is constructed, starting with the homeostatic physiological behavior baseline: for each objective data point x d (t) Calculate the average value μ d,ref Standard deviation σ d,ref And the correlation coefficient matrix between indicators ∑ ref ∈R D1×D1D1 is the dimension of the objective data, and its ij-th element ∑ i,j ref Obtained through Pearson correlation coefficient; then the emotional phase timeline baseline: based on the reference stable period T. ref Self-rated emotion data x MOOD (t), t∈T ref Cosine fitting was used to extract daily rhythm parameters: The rhythm parameters are obtained by fitting with the least squares method: rhythm amplitude A (reflecting the daily fluctuation of emotions), peak phase φ (reflecting the daily occurrence time of the emotional peak), and daily average level C of emotion score, and the random noise residual term ε(t) of the emotion time series at time t. This embodiment specifically explains that for single-value time-series features (effective sleep duration, body surface temperature, exercise intensity, etc.), the window mean and standard deviation are directly calculated using all scalar sampling points within the sliding window; for activity rhythm features, the mean of exercise intensity within the baseline interval Δt is used as the time-series segmented vector, and during statistics, the time-series mean and standard deviation are independently calculated for each segment component, generating the activity rhythm mean vector and standard deviation vector respectively; the sample size is T. ref / △t, round up.
[0018] S2.3: The individualized baseline set includes prospective baseline B pr and review baseline B re Forward baseline B pr Divided into long-term stable baseline B pr,lo and short-term dynamic baseline B pr,sh Both are used for real-time risk prediction, B pr,lo The data inclusion criteria are: only data that has been clinically confirmed as being in a non-acute phase and whose confirmation time is no later than the current time △ cf Historical data for 7 days (usually 7 days, ensuring that all data used for baseline calculations are fully clinically validated and that unvalidated recent data are not included in the baseline) are defined by the following data admission rules: B pr,lo (t)=Base({X(s):s≤t-△ cf Label(s) = non-episode}), Base(·) is the baseline statistical calculation function of S2.2, X(s) = [x1(s), x2(s), ..., x D (s)] T Let D be the data vector of all dimensions collected at time s, where D is the dimension of the data vector, and Label(s) = non-episode indicates that time s has been clinically confirmed not to be within the acute phase; B pr,sh The data admission rule is: only data up to the current time △1 will be included. cf Historical data that is 1 day in length (usually 1 day) and has been clinically confirmed as being from a non-acute phase, is defined as follows: Bpr,sh (t)=Base({X(s):s≤t-△1 cf Label(s) = non-episode} (used to capture chronic prodromal symptoms of slow disease progression and gradual phase transitions over several months); review baseline B re For offline analysis and model training, the data admission rules are as follows: all historical data are included, without excluding epileptic phases, but all data are accompanied by clinically labeled epileptic / non-episodic phase tags, used only for model evaluation and post-hoc analysis, not for real-time prediction; based on the two-branch data admission rules of the prospective baseline, a long-term stable baseline B is triggered once a day according to a preset time window (e.g., 03:00 every day). pr,lo and short-term dynamic baseline B pr,sh The statistical updates are uniformly numbered K (representing the daily update cycle number of the daily baseline); the B pr,lo Statistical econometric updates: up to K-△ cf The source data is the confirmed non-acute period data from 2000 days ago. The mean μ of each dimension is updated using an exponentially weighted moving average strategy. d,pr,lo (K) and standard deviation σ d,pr,lo (K), μ d,pr,lo (K)=α×μ d,pr,lo (K-1)+(1-α)×μ d,re,lo , μ d,re,lo σ d,re,lo The mean and standard deviation of confirmed non-acute phase data within the most recent update window (usually 7 days), where the long-term forgetting factor α for historical data is a preset value (range 0.9–0.95); The B... pr,sh Statistical econometric update: up to K-△1 cf The data from the previous day, which were confirmed to be in a non-acute phase, were used as the source data. Similarly, the mean μ of each dimension was updated. d,pr,sh (K) and standard deviation σ d,pr,sh (K), where the short-term forgetting factor α1 in the exponentially weighted moving average strategy is a preset value (range 0.8 to 0.88). Please see Figure 3 As shown, S3: The deviation between the aligned multi-source time-series monitoring data and the current individualized baseline set is calculated, and the deviation sequence is obtained by combining the data quality score. Multi-scale feature decoupling is performed on the deviation sequence to separate the short-term transient disturbance features and medium-to-long-term trend drift features of the multi-source data, resulting in a global feature vector, including: S3.1: Calculate the deviation between the aligned multi-source time-series monitoring data and the current individualized baseline set, and combine this with the data quality score to obtain the deviation sequence. For the physiological behavior data dimension x... d At time t, Based on long-term stable baseline B pr,lo and short-term dynamic baseline B pr,sh Calculate the two sets of standardized deviations: long-run standardized deviation z d,lo (t)=(x d (t)-μ d,pr,lo ) / (σ d,pr,lo +ε) and short-term standardized deviation z d,sh (t)=(x d (t)-μ d,pr,sh ) / (σ d,pr,sh +ε), μ d,pr σ d,pr These are the mean and standard deviation of physiological behavioral data in prospective baseline data, ε=10. -6 To prevent division by zero by small constants; for the self-rating emotion data x MOOD (t), based on rhythmic parameters in the sentiment phase baseline in the long-term and short-term prospective baselines, including rhythmic amplitude A. pr Peak phase φ pr and rhythm termination C pr Construct long-term predicted normal values x for patients at any time t during the day. MOOD,lo c (t) and short-term predicted normal value x MOOD,sh c (t), corresponding to the standardized deviation of the two sets of emotion self-rating data: z MOOD,lo (t)=(x MOOD (t)-x MOOD,lo c (t)) / (σ MOOD,lo +ε) and z MOOD,sh (t)=(x MOOD (t)-x MOOD,sh c (t)) / (σ MOOD,sh +ε), σ MOOD,lo σ MOOD,sh These are the standard deviations of the residuals between the actual self-assessed values and the baseline predicted values during the stable period corresponding to the long-term stable baseline and the short-term dynamic baseline, respectively. For each data dimension, the maximum of the long-term and short-term standardized deviations is taken as the final standardized deviation z for each dimension. d (t)=max(z d,lo (t),z d,sh (t)); and simultaneously incorporate the weighted bias z of the data quality score Q(t). d j (t)=Q(t)×z d (t); finally, the deviation vector Z(t) = [z1] at time t is obtained. j (t),z2 j (t),...,zD j (t)] T D represents the dimension of the data; S3.2: The bias sequence {z} for each dimension d d j (t)} t=t0-W1 t0 W1 represents a continuous time window (e.g., the past 14 days), and t0 represents the current time. Multi-scale decomposition is performed using Discrete Wavelet Transform (DWT) to obtain... (Preferably, the db4 wavelet basis function is selected), A d,M W represents the low-frequency approximation component of layer M (corresponding to medium- to long-term trend drift characteristics (scales from several days to weeks, related to the evolution of disease precursors)). d,m The m-th layer represents the high-frequency detail components (corresponding to short-term transient disturbance characteristics (hourly to daily scale, external stress, temporary activity noise)), and M represents the wavelet decomposition layer number (based on the sampling frequency f). s Preset, such as M=[log2(f s / f tar )], round down, f tar (Classification frequency of short-term disturbances and medium-to-long-term trends); the coefficient sequences of the first M1 layers of the high-frequency detail components obtained from the decomposition are statistically calculated and aggregated into short-term transient disturbance features F. d,ST This includes standard deviation (characterizing intraday volatility), zero-crossing rate (characterizing the frequency of signal changes), power spectral density ratio (characterizing the proportion of diurnal rhythm components), and lag-1 autocorrelation coefficient (characterizing intraday signal persistence). The remaining high-frequency detail components and low-frequency approximation components are then aggregated into a medium- to long-term trend drift characteristic F. d,LT This includes the linear trend slope (representing the daily direction and rate of change), monotonicity index (representing the degree of trend consistency), lag 1 / 3 / 7 autocorrelation coefficients (representing the persistence of the trend at different time delays), and low-frequency energy proportion (representing the proportion of the trend component in the total signal). M1 is based on the sampling frequency f of dimension d. s With the preset minimum characteristic period T of the medium-to-long-term trend LT Determine: M1=[log2(f s ×T LT )], round down, f tar =1 / T LT In this embodiment, if the heart rate data is uniformly resampled to f s =1 time / hour, T LT=24 hours, J1=[log2(1×24)], rounded down to 4, the high-frequency detail components of layers 1 to 4 are aggregated as short-term perturbation features, and the components of layers 5 and above and low-frequency approximation are aggregated as medium- and long-term trend features; take J=4, use db4 wavelet to decompose the heart rate deviation sequence of the past 14 days into 4 layers to obtain the high-frequency detail components W1 to W3 aggregated as short-term perturbation features; the high-frequency detail component W4 and the low-frequency approximation A4 aggregated as medium- and long-term trend features; In this embodiment, it should be noted that standard deviation, zero-crossing rate, autocorrelation coefficient, slope, mutual information calculation, etc. are all basic definitions in the fields of statistics and signal processing, and will not be elaborated here.
[0019] S3.3: A key signal in the prodromal phase of bipolar disorder is the weakening or loss of cross-system synergy—that is, the normal coupling between different physiological and behavioral indicators is disrupted. Based on the discrete wavelet transform decomposition results, the appropriate time-scale signal is selected according to the coupling type to calculate the dynamic mutual information (DMI) between any two dimensions (i,j). ij cu The real-time dynamic mutual information is compared with the mean mutual information μ in the look-ahead baseline. ψ,ij pr Standard deviation σ ψ,ij pr Performing standardized calculations, we obtain the rate of change of cross-modal coupling strength Δψ. ij If i < j, quantify the degree of decoupling in the multimodal system; finally, all features are concatenated into a global feature vector F(t) = [{F d,ST} d=1 D ,{F d,LT} d=1 D ,{△ψ ij} i<j ] T The selection of appropriate time-scale signals based on coupling type: For each pair of dimensions (i,j), based on the physiological coupling type they reflect, appropriate time-scale signals are selected from the discrete wavelet transform decomposition results to calculate mutual information, including: autonomic nervous system-behavioral coupling (such as heart rate and exercise intensity): the correlation between the two is most significant on a minute to hourly scale, and a short-term perturbation reconstruction signal z is selected. i ST (t) and z j ST (t), Emotion-Autonomic Coupling (e.g., HRV and Emotional Self-Rating): The association between emotional fluctuations and autonomic regulation is significant on an intraday scale; short-term perturbation signals are used to reconstruct the signal. Rest-Activity Rhythm Coupling (e.g., Effective Sleep Duration and Activity Rhythm): The compensatory relationship between sleep and exercise is significant on a trans-day scale; medium- to long-term trend signals are used to reconstruct the signal. i LT (t) and zj LT (t), for other pairings not explicitly defined above, the short-term perturbation reconstructed signal is used by default; taking the current time t as the end point of the window, a sliding window of fixed length L is taken (in this embodiment, L=7 days, i.e., data points from the past 7 days), and the signal sequences of the two dimensions within the window are extracted respectively: Z i ={z i (t-L+1),z i (t-L+2),...,z i Similarly, Z(t)} is obtained. j Sort the values in the two sequences in ascending order, and discretize them into M1 levels (M1=5 in this embodiment) with equal frequency, to obtain the discretized symbol sequence Z. i L and Z j L Calculate the dynamic mutual information (DMI) between any two dimensions (i,j). ij cu This leads to the cross-modal coupling strength variation rate Δψ ij =(DMI ij cu -μ ψ,ij pr ) / (σ ψ,ij pr +ε), △ψ ij ≈0 indicates that the current coupling strength is consistent with the stable period, and the two-dimensional cooperative relationship is normal, △ψ ij A value greater than 0 indicates that the current coupling strength is higher than normal, potentially suggesting abnormal over-cooperation. Δψ ij A value less than 0 indicates that the current coupling strength is below normal, suggesting that the normal cooperative relationship between the two dimensions has been disrupted, which is an important signal for judging the precursor stage of biphasic disorder; in a preferred embodiment, since the total number of dimensions is D, the number of pairwise pairings is The full computation on embedded devices is computationally intensive. This invention only calculates the rate of change of cross-modal coupling strength for three clinically validated physiological pairs: heart rate and exercise intensity, HRV and mood self-rating, and effective sleep duration and activity rhythm. Other pairs are not included in the global feature vector. S4: Input the global feature vector into the pre-trained biphasic emotion phase transition probability model, and output the target patient's current emotion recognition state and a dynamic risk index for transitioning to the manic or depressive phase, including: S4.1: The biphasic emotion phase transition probability model is a five-state hidden Markov model, whose hidden states include a stable phase, a manic pre-manic phase, a manic episode phase, a depressive pre-depressive phase, and a depressive episode phase. The model describes the transition rules between each phase through a state transition probability matrix, and outputs the posterior probability of each state at the current moment through a forward algorithm, taking the state corresponding to the maximum posterior probability as the current emotion recognition state. The dynamic risk index is obtained by multiplying the current state posterior probability distribution by the Δth power of the state transition probability matrix, and the dynamic risk index for the manic phase and the depressive phase transition is calculated respectively. S4.2: A pre-trained biphasic emotional phase transition probability model based on a five-state HMM (Hidden Markov Model) framework is used to divide the target patient's emotional state into five implicit states: S={S0,S1,S2,S3,S4}, corresponding to the stable phase, manic prodrome, manic episode, depressive prodrome, and depressive episode, respectively. The global feature vector F(t) is input into the pre-trained biphasic emotional phase transition probability model to form the monitoring vector O for the time window t. t =F(t); S4.3: The five-state HMM model contains three sets of parameters: initial state distribution State transition probability matrix A and monitoring probability matrix B; =P(q1=S k ), k=0,1,2,3,4, q1 is the patient's implicit state at the first time point of the monitoring sequence (the initial state distribution is set according to the patient's historical epileptic pattern (e.g., patients with a history of many depressive episodes, and (It can be appropriately increased), or set to a uniform distribution when there is no clear prior knowledge. =0.2); A=[a IJ ], a IJ =P(q t+1 =S J |q t =S I ), IJ∈{0,1,2,3,4}, a IJ At time t, the state is S I Under the given conditions, the state transitions to state S at the next time step t+1. J The probability; Monitoring matrix B: B J (O t )=P(O t |q t =S J )=N(O t |μ J ,∑ J That is, the monitoring vector in each hidden state follows a multidimensional Gaussian distribution, with parameter μ. J ,∑J These are the mean and covariance matrix of the J-th state, respectively (learned from labeled training data using the standard training method of HMM). In this embodiment, it is important to specify that the stable phase is characterized by normal mood, energy, and activity levels, with no significant symptoms; the manic prodromal phase is characterized by mildly elevated mood, increased activity, and decreased sleep needs, but has not yet reached the criteria for an episode; the manic episode is characterized by meeting the DSM-5 (Diagnostic and Statistical Manual of Mental Disorders) criteria for mania or an episode, requiring clinical intervention; the depressive prodromal phase is characterized by mildly depressed mood, decreased interest, and increased fatigue, but has not yet reached the criteria for an episode; the depressive episode is characterized by meeting the DSM-5 criteria for a depressive episode, requiring clinical intervention; the transitions between the five states are not entirely free, but follow the natural course of bipolar disorder: the probability of transitions between adjacent states is relatively high (e.g., S0→S1 or S0→S3); leapfrogging is prohibited, and the probability of such prohibitions is constrained to 0 in this embodiment (e.g., S0→S2 or S0→S4), constrained by clinicopathology; the stable phase S0 is the "pivot" for state transitions, with most transitions starting or ending at S0.
[0020] S4.4: Let the monitoring sequence O 1:t ={O1,O2,...,O t} represents the complete sequence of monitored feature vectors from system startup to the current time t, combined with model parameters θ={ For each state (A, B) at time t, the posterior probability of each state at time t is calculated using the model forward algorithm: At the first time t=1, the initial forward probability a1(J) of each state is calculated. ×B J (O1); For times t=2,3,...,T, calculate sequentially. , for all times t, a t (J) Normalization yields the patient's state S at time t. J posterior probability γ t (J)=a t (J) / ∑a t (k); Finally, the current emotion recognition state is determined by the state S with the highest posterior probability. t =argmaxγ t (J); Based on the state posterior probability distribution γ at the current time t t Given the state transition probability matrix A, predict the state probability distribution after a future time Δ (usually 72 hours): A △ Let A be the Δ power of the transition probability matrix, (A △ ) IJLet Δ be the probability of transitioning from state I to state J; finally, the sum of the probabilities of being in either the premanic phase S1 or the manic episode S2 after a future time Δ is defined as the manic phase transition risk index R. M (t)=β △ (1)+β △ (2) Define the sum of probabilities of being in the prodromal phase S3 or the depressive episode S4 after the future time Δ as the depression phase transition risk index R. D (t)=β △ (3)+β △ (4); This embodiment needs to specifically explain the pre-trained bipolar mood phase transition probability model based on the five-state HMM model (five-state hidden Markov model). The training method is as follows: (1) Training data: Collect retrospective data of no less than 200 patients with bipolar disorder, and provide no less than 6 months of continuous multimodal monitoring data and corresponding mood phase annotations for each patient; (2) Based on the retrospective baseline B re The extracted global feature vector F(t) is used as the monitoring value O at each time step. t (3) Through the standard parameter estimation method of HMM (specifically used to train Hidden Markov Model Baum-Welch algorithm), the expected value is maximized iteratively updated. The algorithm iterates through three parameters, A, B, until convergence. Elements in matrix A constrained by clinicopathology are set to 0 after each iteration and do not participate in the parameter update iterations of the Baum-Welch algorithm. The remaining elements are renormalized. For the monitoring probability B... J (O t If the clinical sample is complex and has a multimodal feature distribution with multiple peaks, the monitoring probability can be extended to a Gaussian mixture model (GMM). The mixture model is composed of a weighted superposition of C multidimensional Gaussian sub-distributions. The parameters include the mean, covariance, and mixture weight coefficient of each sub-distribution. The mean and covariance matrix of each state are estimated by the expectation-maximization algorithm.
[0021] S5: Based on the dynamic risk index and risk trends, tiered early warnings are issued according to threshold rules, and the warning information is output to the terminal, including: S5.1: Based on the manic phase transition risk index R M (t) and the depression phase transition risk index R D (t), calculate the comprehensive risk score R for bipolar disorder. co (t)=w M ×R M (t)+w D ×R D (t), w M w D These are the corresponding weights (e.g., w) M=0.5、w D =0.5 (can be personalized based on the historical seizure type distribution of the target patient); while also considering the trend of risk changes. , △ st The set average warning period (e.g., 24 hours). S5.2: Implement tiered early warning based on threshold rules: If R co (t) < the first risk threshold R_th1 (e.g., 0.3) and Risk Trend First Threshold (e.g., 0.15), safe and green with no early warning, routine monitoring is performed, if R_th1≤R co (t) < the second risk threshold R_th2 (e.g., 0.55) or Risk Trend Second Threshold (e.g., 0.4), Level 1 Yellow Alert, push behavior suggestions, if R_th2≤R co If (t) < the third risk threshold R_th3 (e.g., 0.75) and the duration is ≥ the preset duration (e.g., 2 days), a Level 2 Orange Alert is issued. The caregiver is notified, and a telephone follow-up is recommended. If R... co (t)≥R_th3 or A Level 3 Red Alert is issued, recommending immediate clinical medical attention. Finally, the alert information is output to the terminal, including the alert level, corresponding color code, and the manic phase transition risk index R. M (t), Depression Phase Transition Risk Index R D (t), the current emotion recognition state judgment result, and the corresponding clinical suggestions; the terminal includes one or more of the following: the target patient's mobile terminal, the clinician's terminal device, the caregiver's terminal device, and the hospital information system's terminal device, used to integrate the early warning information into the electronic health record system; each threshold is a configurable initial value, and after actual deployment, it will be automatically fine-tuned periodically (e.g., quarterly) based on the S6.3 calibration result, using grid search to maximize the F1-score; S6: Based on the early warning information, clinical feedback is confirmed. The individualized baseline set and the parameters of the bipolar emotion phase transition probability model are updated in a closed loop according to the confirmed labeled data, and adaptive calibration is performed, including: S6.1: The system automatically generates a pending confirmation event and pushes it to the clinician's terminal when any of the following conditions are met: the first system issues a level 2 or higher warning, and the second system issues a level 1 warning for L consecutive days without downgrading (e.g., L=7 days); the pending confirmation event is sent via push notification through the clinician's terminal application. Confirmation by the clinician within L days is considered valid feedback; if not confirmed within L days, the event expires, and the system automatically marks the data during the warning period as "pending," excluding it from subsequent baseline updates and model calibration; the clinician confirms the event on the terminal device, and the system receives: labeled data: event confirmation label. ep ∈{0,1}, 1 indicates a confirmed seizure, 0 indicates no seizure; seizure type (when Label) ep =1): Type∈{manic prodrome, manic episode, depressive prodrome, depressive episode}; onset and end time of episode; S6.2: For time periods confirmed by clinicians as episodic, all time-series data within those periods are retrospectively removed from the prospective baseline calculation data source to prevent extreme values from episodic periods from contaminating the normal baseline; this retrospective removal is triggered during the next individualized baseline update after the physician's confirmation of the event, and is applied to data that is outdated (exceeding Δ). cf Window and more than △1 cf If the data (for days) is confirmed to be in an attack phase, the system will recalculate the baseline and remove the data. However, data already used for historical prediction cannot be retrospectively modified. After each baseline calibration, the complete monitoring timeline of patients and clinically confirmed labeled events for a set number of consecutive months (e.g., 3 months) are used as a unified validation dataset. The old baseline parameters before calibration are used as the baseline period, and the new baseline parameters after calibration iteration are used as the evaluation period. The precision (proportion of actual attacks in the warning events) and recall (proportion of warnings in actual attacks) of the warning are calculated respectively. If the precision or recall is improved after calibration compared to before calibration, the calibration is retained; otherwise, it reverts to the baseline parameters before calibration. If the precision or recall does not improve after a set number of consecutive calibrations (e.g., 2 times), the system will trigger a prompt, suggesting manual review of the baseline construction logic or warning threshold. S6.3: In the initial stage after model deployment (e.g., the first 30 days), incremental updates are not triggered; only labeled events are accumulated. The triggering conditions for incremental updates must be met simultaneously: First, since the last update, the newly accumulated labeled events must include both acute and non-acute phase samples, and the total number of samples must be no less than the set number (e.g., 5 items). Second, the time elapsed since the last incremental update has exceeded the set time ΔT2 (e.g., 72 hours), and the time elapsed since the most recent warning event has exceeded the set time ΔT3 (e.g., 24 hours, ensuring clinical confirmation of the event has been completed). The newly acquired clinical feedback labeled data is organized into an incremental training set D_new. Using the current model parameters θ_old as initial values, the standard parameter estimation method of HMM is re-executed to update the model parameters to θ_new. =(1-η)×θ_old+η×θ_D_new, where η∈(0,1) is the learning rate (usually η=0.3); when the cumulative number of labeled events reaches a preset threshold (e.g., 50), a full retraining is triggered; after each model calibration, the model's accuracy or F1-score is recalculated on the labeled data of the most recent set number of months (e.g., 3 months); if either of the two metrics is higher after calibration than before calibration, the updated model parameters are retained; otherwise, the model reverts to the version before calibration and the calibration is recorded as "invalid calibration"; the calibration effect is summarized and statistically analyzed on a weekly basis. If the model performance (F1-score) does not show a positive improvement within a set number of consecutive weeks (e.g., 4 weeks), a system alarm is triggered, prompting manual intervention for review; In a preferred embodiment of this invention, the experimental dataset consists of 236 patients diagnosed with bipolar disorder, and wearable multi-source time-series monitoring data from six consecutive months, including clinical annotations for manic / depressive episodes. The dataset is divided into a training set of 166 cases and a test set of 70 cases. Baseline solutions were compared: Solution 1: Traditional single physiological indicator early warning (HRV + fixed sleep duration threshold only); Solution 2: Single-scale machine learning early warning (Lasso only, without wavelet multi-scale decoupling or cross-modal coupling features); Solution 3: Conventional HMM early warning (without individualized dual baselines or clinical closed-loop calibration); Solution 4: The solution of this invention. Evaluation metrics included early warning accuracy, recall, F1 score, 72-hour early warning duration for episodes, and average number of false alarms per month. The overall early warning performance of different solutions was compared as follows:
[0022] Compared to existing single-threshold, single-scale machine learning, and HMM schemes without individualized baselines, the present invention improves the F1 score by up to 0.466, increases the early warning time by nearly 4 times, and significantly reduces false alarms by using asynchronous dual baselines, multi-scale wavelet decoupling, cross-modal coupling features, clinically constrained five-state HMM, and event closed-loop calibration.
[0023] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A dynamic early warning method for the risk of bipolar disorder, characterized in that: include: S1: Collect multi-source time-series monitoring datasets of target patients, including physiological time-series data, behavioral time-series data and emotional self-rating time-series data, and perform time alignment of multi-source data. Calculate data quality scores based on the data time alignment results. S2: Construct an individualized baseline set based on the historical time-series monitoring data aligned with the target patient, including the homeostatic physiological behavior baseline and the emotional phase time-series baseline, and update the individualized baseline set by sliding according to a preset window mechanism; S3: Calculate the deviation between the aligned multi-source time-series monitoring data and the current individualized baseline set, and obtain the deviation sequence by combining the data quality score. Decouple the deviation sequence by multi-scale features, separate the short-term transient disturbance features and medium-to-long-term trend drift features of the multi-source data, and obtain the global feature vector. S4: Input the global feature vector into the pre-trained biphasic emotion phase transition probability model and output the target patient's current emotion recognition state and the dynamic risk index of transitioning to the manic or depressive phase; S5: Based on the dynamic risk index and the trend of risk changes, it performs graded early warning according to threshold rules and outputs the early warning information to the terminal; S6: Based on the early warning information, clinical feedback is confirmed, and the parameters of the individualized baseline set and the bipolar emotion phase transition probability model are updated in a closed loop according to the labeled data confirmed by the feedback, and adaptive calibration is performed.
2. The method for dynamic early warning of bipolar disorder risk according to claim 1, characterized in that: In S1, the data quality score is defined as follows: Q(t)∈[0,1]: Q(t) = scores of all dimensions SC d Sum / Total Dimensions D, if within the baseline sampling interval Δt, the proportion of missing original sampling points in dimension d is greater than or equal to the missing threshold, indicating missing labeled data, then SC d =0, if the missing percentage is less than the missing threshold within the baseline sampling interval Δt, it indicates that data exists, SC d =max(0,1-the proportion of missing original sampling points in the d-th dimension).
3. The method for dynamic early warning of bipolar disorder risk according to claim 1, characterized in that: The implementation of S2 includes: S2.1: In the initial stage after system startup, collect multidimensional time-series data D_in={X(t)} of a set length T_in that has been aligned with the time axis. t=1 T_in For each time t, calculate the multidimensional stability period score SI(t); identify time periods that meet the following condition as candidate stability periods: SI(t) > corresponding threshold SI th Duration ΔT ≥ the set minimum duration of the stabilization period T min,sta If there are no warning trigger records for the duration △T, and there are multiple candidate stable periods, the one with the largest average SI value among the multiple candidate stable periods is selected as the reference stable period T. ref .
4. The method for dynamic early warning of bipolar disorder risk according to claim 3, characterized in that: The implementation of S2 further includes: S2.2: based on the reference stability period T ref Based on the data, an individualized baseline set is constructed, starting with the homeostatic physiological behavior baseline: for each objective data point x d (t) Calculate the average value μ d,ref and standard deviation σ d,ref And the correlation coefficient matrix between indicators ∑ ref ∈R D×D D1 represents the dimension of objective data; then the emotional phase timeline baseline is based on the reference stable period T. ref Self-rated emotion data x MOOD (t), cosine fitting is used to extract the daily rhythm parameters, and the least squares method is used to fit the rhythm parameters: rhythm amplitude A, peak phase φ and rhythm termination C; S2.3: The individualized baseline set includes prospective baseline B pr and review baseline B re Forward baseline B pr Divided into long-term stable baseline B pr,lo and short-term dynamic baseline B pr,sh Based on the two types of data admission rules for the forward-looking baseline, a long-term stable baseline B is triggered once a day according to a preset time window. pr,lo and short-term dynamic baseline B pr,sh The statistics are updated, with the update sequence number being K; the B pr,lo Statistical econometric update: The mean μ of each dimension is updated using an exponentially weighted moving average strategy. d,pr,lo (K) and standard deviation σ d,pr,lo (K); the B pr,sh Statistical econometric update: Similarly, update the mean μ of each dimension. d,pr,sh (K) and standard deviation σ d,pr,sh (K).
5. The method for dynamic early warning of bipolar disorder risk according to claim 1, characterized in that: The implementation of S3 includes: S3.1: Calculating the deviation between the aligned multi-source time-series monitoring data and the current individualized baseline set, and obtaining the deviation sequence by combining the data quality score, for the physiological behavior data dimension x. d At time t, based on the long-term stable baseline B pr,lo and short-term dynamic baseline B pr,sh Calculate the two sets of standardized deviations: long-run standardized deviation z d,lo (t) and short-run standardized deviation z d,sh (t); for the self-rating emotion data x MOOD (t), based on rhythmic parameters in the sentiment phase baseline in the long-term and short-term prospective baselines, including rhythmic amplitude A. pr Peak phase φ pr and rhythm termination C pr Construct long-term predicted normal values x for patients at any time t during the day. MOOD,lo c (t) and short-term predicted normal value x MOOD,sh c (t), corresponding to the standardized deviation of the two sets of emotion self-rating data: z MOOD,lo (t) and z MOOD,sh (t); For each data dimension, the maximum of the long-term and short-term standardized deviations is taken as the final standardized deviation z for each dimension. d (t); simultaneously, the weighted bias z of the data quality score Q(t) is incorporated. d j (t)=Q(t)×z d (t); finally, the deviation vector Z(t) at time t is obtained.
6. The method for dynamic early warning of bipolar disorder risk according to claim 5, characterized in that: The implementation of S3 further includes: S3.2: the deviation sequence {z} for each dimension d. d j (t)} t=t0-W t0 W represents the continuous time window, t0 represents the current time, and discrete wavelet transform is used for multi-scale decomposition. The coefficient sequences of the first J1 layers of the high-frequency detail components obtained from the decomposition are statistically calculated and aggregated into short-term transient perturbation features F. d,ST The remaining high-frequency detail components and low-frequency approximate components are aggregated into a medium- to long-term trend drift feature F. d,LT ; S3.3: Based on the discrete wavelet transform decomposition results, select the appropriate time-scale signal according to the coupling type, and calculate the dynamic mutual information (DMI) between any two dimensions (i,j). ij cu The real-time dynamic mutual information is compared with the mean mutual information μ in the look-ahead baseline. ψ,ij pr Standard deviation σ ψ,ij pr Performing standardized calculations, we obtain the rate of change of cross-modal coupling strength Δψ. ij , i < j; finally, all features are concatenated into a global feature vector F(t) = [{F d,ST } d=1 D ,{F d,LT } d=1 D ,{△ψ ij } i<j ] T .
7. The method for dynamic early warning of bipolar disorder risk according to claim 1, characterized in that: The implementation of S4 includes: the biphasic emotion phase transition probability model is a five-state hidden Markov model, whose hidden states include a stable phase, a manic pre-manic phase, a manic episode phase, a depressive pre-depressive phase, and a depressive episode phase; the model describes the transition rules between each phase through a state transition probability matrix, and outputs the posterior probability of each state at the current moment through a forward algorithm, and uses the state corresponding to the maximum posterior probability as the current emotion recognition state judgment result; the dynamic risk index is obtained by multiplying the current state posterior probability distribution by the state transition probability matrix to the power of Δ, and the dynamic risk index for the manic phase and the depressive phase transition is calculated respectively.
8. The method for dynamic early warning of bipolar disorder risk according to claim 1, characterized in that: The implementation of S5 includes: S5.1: Based on the manic phase transition risk index R M (t) and the depression phase transition risk index R D (t), calculate the comprehensive risk score R for bipolar disorder. co (t); while also considering the changing trend of risk. , △ st The set average warning period; S5.2: Implement tiered early warning based on threshold rules: If R co (t) < the first risk threshold R_th1 and Risk Trend First Threshold Safe, green, and without early warning; routine monitoring is performed. If R_th1≤R co (t) < the second risk threshold R_th2 or Risk Trend Second Threshold Level 1 Yellow Alert, push notification with behavioral suggestions, if R_th2≤R co (t) < risk threshold R_th3 and duration ≥ preset duration, Level 2 Orange Alert, notify caregiver, recommend telephone follow-up, if R co (t)≥R_th3 or A Level 3 red alert is issued, and immediate clinical medical attention is recommended; finally, the alert information is output to the terminal.
9. A dynamic early warning system for bipolar disorder risk, used in conjunction with the dynamic early warning method for bipolar disorder risk as described in any one of claims 1-8, characterized in that, include: Multi-source temporal monitoring data acquisition module: Collects multi-source temporal monitoring datasets of target patients, including physiological temporal data, behavioral temporal data and emotional self-rating temporal data, and performs time alignment of multi-source data, and calculates data quality scores based on the data time alignment results; Asynchronous dual-baseline architecture construction module: Based on the historical time-series monitoring data aligned with the target patient, an individualized baseline set is constructed, including the steady-state physiological behavior baseline and the emotional phase time-series baseline, and the individualized baseline set is updated by sliding according to a preset window mechanism; Multi-scale sentiment feature decoupling module: The alignment of multi-source time-series monitoring data and the current individualized baseline set are used to calculate the deviation, and the deviation sequence is obtained by combining the data quality score. Multi-scale feature decoupling is performed on the deviation sequence to separate the short-term transient disturbance features and medium- and long-term trend drift features of multi-source data and obtain the global feature vector. Emotional phase transition risk inference module: Input the global feature vector into the pre-trained biphasic emotional phase transition probability model, and output the target patient's current emotional recognition state and the dynamic risk index of transitioning to manic or depressive phases; Dynamic risk warning module: Based on the dynamic risk index and the trend of risk changes, it provides graded warnings according to threshold rules and outputs the warning information to the terminal; Event-triggered closed-loop calibration module: Based on early warning information, clinical feedback is confirmed, and the parameters of the individualized baseline set and the bipolar emotion phase transition probability model are updated in a closed loop according to the confirmed annotation data, and adaptive calibration is performed.
