A Data Processing Method for Health-Conversational Intelligent Agents Based on Emotion Perception

By fusing multimodal physiological signals with emotional text and employing dual-granularity memory, this technology solves the problems of lacking psychological contextual semantics and long-term time-series tracing in existing technologies. It achieves accurate processing of emotional understanding and health data, and provides a low-intrusion intelligent reminder mechanism.

CN123117364APending Publication Date: 2026-11-17NANJING JIANXIN MEDICAL TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611618838.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing health monitoring equipment cannot effectively distinguish the psychological contextual semantics behind physiological signals, lacks emotional sampling, and is difficult to trace health evolution over long time scales. General large language models have low accuracy in health data processing, and proactive reminder mechanisms lack comprehensive evaluation, leading to excessive disturbance or omission of important risks.

Method used

By collecting multimodal physiological signals in parallel and aligning them with ultra-short emotional texts using timestamps and cross-attention fusion, dual-granularity rolling memory processing is performed. Combining emotional-physiological dual-stream alignment, dual-granularity rolling memory, separate computation, and semantic reasoning, psychological context understanding and low-intrusion intelligent reminders are achieved.

Benefits of technology

It achieves psychological context understanding, long-term accurate memory, highly reliable numerical response, and low-intrusion intelligent reminders, providing health conversational intelligent agent data processing that combines semantic understanding and numerical accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123117364A_ABST
    Figure CN123117364A_ABST
Patent Text Reader

Abstract

This invention discloses a data processing method for a health-conversational intelligent agent based on emotion perception, comprising: performing timestamp alignment and cross-attention fusion on multimodal physiological signals from wearable devices and ultra-short emotional text actively input by the user to generate a calibrated physiological feature sequence; performing dual-granularity rolling memory processing on the calibrated physiological feature sequence to obtain short-term cached data and long-term compressed memory data, and recursively updating the hidden health state; receiving a natural language health query and decomposing it into a numerical calculation subtask and a semantic generation subtask through routing decision, which are then processed by a graph neural network computation channel and a medical fine-tuning large language model semantic channel, respectively, and integrated to output a dialogue response; performing a three-factor weighted score based on real-time sensor data, current emotional state, and hidden health state, and outputting graded intervention instructions based on the score and threshold comparison results. This invention achieves a deep integration of wearable health monitoring and emotional semantic understanding.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of health monitoring and artificial intelligence interaction technology, and in particular to a data processing method for a health conversational intelligent agent based on emotion perception. Background Technology

[0002] In recent years, with the rapid development of wearable sensing technology and artificial intelligence, smartwatches, wristbands, and other devices integrating sensors such as heart rate, heart rate variability, skin conductance, and acceleration have been widely used for daily health monitoring. These devices can continuously capture users' physiological signals and issue alerts when abnormalities are detected. At the same time, conversational agents based on large language models are gradually being introduced into the field of health management, attempting to provide users with natural language interactive health consultations and guidance.

[0003] However, existing technologies still face several unresolved challenges in the aforementioned integration directions. First, the physiological signals passively collected by wearable devices lack contextual semantics. For example, an elevated heart rate may stem from exercise, pain, anxiety, or a drug response; sensor data alone cannot distinguish the underlying driving factors. Existing solutions either completely ignore this semantic gap or rely on users actively completing lengthy emotional questionnaires, resulting in sparse data and a heavy burden on users, making them unsuitable for sustained operation in daily health management. Second, existing health inference models mostly employ pattern recognition paradigms within short time windows, making it difficult to trace, compare, and predict users' health evolution over timescales of weeks to months. Continuous physiological data storage is extremely expensive, while simple downsampling can lead to the loss of crucial event information. Third, general-purpose large language models perform poorly when processing health time-series numerical queries. Research shows that existing models have accuracy rates far below practical requirements in tasks involving heart rate trend calculations, sleep duration statistics, and indicator correlation analysis, frequently producing numerical errors or logical contradictions. Fourth, existing proactive reminder mechanisms are usually based on fixed thresholds or simple schedules, lacking a comprehensive assessment of the user's current activity context, emotional state, and health risks, resulting in either excessive disturbance to the user or omission of important risks.

[0004] Therefore, how to organically combine the multimodal physiological perception, low-burden emotional sampling, long-term health memory storage, and precise numerical reasoning capabilities of wearable devices to construct an intelligent agent method that can both understand user emotions and accurately process health data has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a data processing method for a health-oriented conversational intelligent agent based on emotion perception, in order to solve at least some of the problems existing in the prior art.

[0006] Technical solution: A data processing method for a health-oriented conversational intelligent agent based on emotion perception, comprising the following steps:

[0007] Based on the multimodal physiological signals collected in parallel by wearable devices and the ultra-short emotional text actively input by users, timestamp alignment and cross-attention fusion processing are performed to obtain a calibrated physiological feature sequence with emotional labels.

[0008] The calibration physiological feature sequence is processed by dual-granularity rolling memory to obtain short-term cache data and long-term compressed memory data. The current health hidden state is then recursively updated based on the calibration physiological feature sequence and long-term compressed memory data.

[0009] Receive user natural language health queries, perform routing decision processing on user natural language health queries, and obtain at least one numerical computation subtask and at least one semantic generation subtask;

[0010] The computation channel is invoked to retrieve and compute from short-term cached data and / or long-term compressed memory data according to the numerical computation subtask, so as to obtain accurate numerical results;

[0011] Invoke the semantic channel, generate natural language response fragments based on semantic subtasks and combine them with precise numerical results, and output the complete dialogue response;

[0012] Based on the sensor data collected in real time by the wearable device, the current emotional state extracted from the calibrated physiological feature sequence, and the current hidden health status obtained by recursion, a three-factor weighted scoring process is performed to obtain the trigger score.

[0013] Based on the comparison between the trigger score and the preset threshold, the system outputs a first-level intervention instruction, a second-level intervention instruction, or a third-level intervention instruction.

[0014] Beneficial effects: This invention achieves healthy conversational intelligent agent data processing that combines psychological context understanding, long-term accurate memory, highly reliable numerical response, and low-intrusion intelligent reminders through emotional-physiological dual-flow alignment, dual-granularity rolling memory, separate computing and semantic reasoning, and three-factor adaptive intervention. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall solution of the present invention.

[0016] Figure 2 This is a schematic diagram of the system hardware architecture and edge-cloud collaborative data flow of the present invention.

[0017] Figure 3 This is a statistical chart of multi-dimensional health data during the postoperative recovery period of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0019] Example 1

[0020] like Figures 1 to 2 As shown, a data processing method for a health-oriented conversational intelligent agent based on emotion perception specifically includes:

[0021] S1: Based on the multimodal physiological signals collected in parallel by wearable devices and the ultra-short emotional text actively input by the user, timestamp alignment and cross-attention fusion processing are performed to obtain a calibrated physiological feature sequence with emotional labels.

[0022] In this application, two data acquisition streams are executed in parallel via a wearable device: one for capturing the user's physiological state and the other for acquiring minimalist emotional feedback proactively provided by the user. Specifically, multiple sensors integrated into the wearable device operate continuously at a sampling rate of no less than 100Hz. A PPG / ECG sensor acquires photoplethysmography (PPG) and electrocardiogram (ECG) signals, an IMU accelerometer captures triaxial acceleration signals, a skin conductance sensor records EDA signals, and a temperature sensor measures body surface temperature. These multi-channel raw physiological signals are fed into a signal conditioning module in real time. For example, this module first applies a bandpass filter with a cutoff frequency of 0.5Hz to 5Hz to preserve heart rate and related harmonic components while suppressing DC drift and power frequency interference. Subsequently, a motion artifact suppression algorithm based on acceleration signals is employed: by detecting abrupt changes in acceleration amplitude, adaptive filtering or interpolation replacement is performed on the PPG / ECG segments within the corresponding time window to obtain a clean, filtered physiological time series.

[0023] Preferably, after signal conditioning, feature extraction is performed on the filtered physiological time series. The HRV time-domain features are calculated as follows: root mean square of continuous differences, standard deviation of normal sinus intervals, and percentage of adjacent RR intervals with a difference greater than 50 ms. When calculating the HRV frequency-domain features, a Fast Fourier Transform is performed on the RR interval sequence to extract the ratio of low-frequency power to high-frequency power. Simultaneously, the amplitude and decay time constant of the skin conductance response are extracted from the EDA signal, the first-order differential rate of change is calculated from the body surface temperature signal, and the sample entropy value is calculated from the acceleration signal. These multidimensional features are concatenated into an original feature vector, which is then mapped to a 128-dimensional physiological embedding vector through a fully connected network. This mapping process not only achieves dimensionality reduction but also enables the embedding vector to better express discriminative information related to emotion and health through end-to-end training.

[0024] Preferably, this solution employs a low-burden sentiment sampling mechanism. When time-triggered conditions (e.g., 8 AM and 8 PM daily), event-triggered conditions (e.g., HRV decreasing by more than 20% of the personalized dynamic threshold within 5 consecutive minutes), or context-triggered conditions (e.g., after a user actively mentions keywords like fatigue or anxiety in a conversation), the wearable device pushes a minimal sentiment sampling prompt to the user via vibration or screen notification, such as "The thing that matters most to me right now, in three words or less." The user-inputted ultra-short sentiment text (typically 1 to 5 words in length) is fed into a lightweight sentiment encoder deployed on the device. This encoder is based on a 5-layer Transformer distillation model, pre-trained on a large-scale sentiment corpus and then pruned, with only about 10% of the parameters of the standard BERT, allowing it to operate at microcontroller-level power consumption. The encoder outputs a 128-dimensional sentiment embedding vector, and the original text is immediately erased from memory after encoding, thus protecting user privacy.

[0025] Furthermore, the sentiment embedding vectors generated within the same time window (e.g., 2.5 minutes before and after, for a total window length of 5 minutes) are used as the query matrix, and the corresponding physiological embedding vectors are used as the key and value matrices, respectively, and input into a multi-head cross-attention layer. In this layer, the dot product of the query matrix and the key matrix is ​​calculated, and after Softmax normalization, the sentiment-physiological association weights are obtained. These weights are then multiplied by the value matrix to output the fused vector. This vector serves as the sentiment anchor attached to the corresponding physiological feature, thereby generating a calibrated physiological feature sequence with sentiment labels. Each sample point corresponds to a 5-minute time slice, containing the original physiological feature values ​​and the added sentiment semantic labels. Through this processing, the originally isolated physiological signals are given a psychological context, providing semantic priors for subsequent health inferences.

[0026] S2: Perform dual-granularity rolling memory processing on the calibration physiological feature sequence to obtain short-term cached data and long-term compressed memory data, and recursively update the current health hidden state based on the calibration physiological feature sequence and long-term compressed memory data.

[0027] In this embodiment, the calibrated physiological characteristic sequences are sent to the edge gateway layer for dual-granularity storage. Preferably, the edge gateway can be a user's smartphone or a dedicated home health center. A circular buffer is established in the edge gateway layer specifically for writing the calibrated physiological characteristic sequences generated every minute within the last 72 hours. This layer, called short-term cache data, is stored without any compression or transformation, at its original sampling precision. The purpose is to support users' precise retrospective queries of recent health events. For example, if a user asks, "What was the trend of my heart rate during the period when I was feeling down last night?", the system can directly retrieve the fine-grained data for the corresponding time period from the short-term cache.

[0028] For historical data exceeding 72 hours, this solution employs a health event primitive compression algorithm to reduce storage overhead. A predefined library of health event primitive types is used, including: "Sleep Phase": characterized by a gradual decrease in heart rate and a slow increase in the ratio of low-frequency to high-frequency HRV power; "Deep Sleep Phase": characterized by extremely low heart rate and high RMSSD; "Morning Blood Pressure Rush Phase": characterized by a rapid increase in systolic blood pressure accompanied by a slight increase in heart rate; "Postprandial Blood Glucose Fluctuation Phase": characterized by a short-term decrease in heart rate variability and a slight increase in skin conductance; "HRV Response Pattern During Exercise": characterized by a sharp increase in heart rate accompanied by a sharp decrease in HRV; and "Stress Recovery Curve": a waveform template showing the slow recovery of HRV to baseline after high-intensity stress, among other standard physiological patterns. For historical calibrated physiological feature sequences exceeding 72 hours, a dynamic time warping algorithm is used to match each segment with each primitive in the template library. The shortest path distance between each data segment and the template is calculated, and the primitive with the smallest distance is selected as the matching result, outputting a matching confidence score between 0 and 1.

[0029] For example, when the confidence level is greater than a preset threshold (e.g., 0.7), the historical data segment is determined to match a certain health event primitive. In this case, only the primitive's identifier (e.g., an integer code), the start and end timestamps of the event, the matching confidence level, and the emotional anchor association identifier corresponding to that period are stored. In this way, physiological signals that might otherwise require thousands of floating-point numbers are compressed into records of tens of bytes. Conversely, if the confidence level is less than or equal to the threshold, the data segment is marked as an "abnormal event" and downsampled at one-tenth of the original sampling rate. The downsampled key feature fragments are stored and also included in long-term compressed memory data. In this way, 90 days of continuous health data can be compressed to less than 3% of the original data volume.

[0030] In addition to storage management, this solution also maintains a health hidden state vector that is recursively updated over time. This state vector is a multidimensional variable, specifically including: cardiovascular load index (reflecting the balance between the sympathetic and parasympathetic nervous systems); autonomic nervous system regulatory efficacy (calculated based on HRV frequency domain indicators); sleep recovery quality (integrating parameters such as deep sleep duration, sleep latency, and nocturnal heart rate drift); and daytime fatigue accumulation rate (assessed through accelerometer-based body movement changes and daytime heart rate patterns). Its update process employs a time-aware gated recursive unit. Specifically, it reads the health hidden state vector S from the previous historical moment. t-1 Obtain the calibration physiological feature sequence X at the current moment. t (Including the latest physiological embeddings and emotional anchors), the current emotional embedding vector E t The inputs, along with the actual time interval Δt between the current moment and the previous historical moment, are fed into the recursive unit according to formula S. t = f(S t-1 , X t E t The updated current health hidden state vector is calculated using the function f(Δt). The core of the function f is an improved gated recurrent unit, which adds a time interval embedding layer to the traditional update and reset gates: Δt is mapped to a time modulation vector through a learnable sine-cosine position code, and then multiplied with the hidden state. This design enables the model to perceive the true time span between two updates, thus more accurately modeling long-term dependencies such as circadian rhythms and work-rest cycles.

[0031] S3: Receive user natural language health query, perform routing decision processing on user natural language health query, and obtain at least one numerical calculation subtask and at least one semantic generation subtask.

[0032] In this application, users input natural language health queries via voice or text, such as "How many days in the past week did I sleep less than 7 hours?" or "Does my deep sleep duration decrease after each night I report anxiety?" The query is first fed into a semantic channel: a large language model fine-tuned from a medical dialogue corpus, with approximately 7 to 13 bytes of parameters. The semantic channel parses the query and outputs a query type label. The labels are divided into three categories: numerical / statistical, indicating that the core request is to obtain a specific number or statistic; semantic interpretation, indicating that the user hopes to receive conceptual explanations or health advice; and multi-step reasoning, indicating that the query requires multiple logical steps, involves multiple data sources, or conditional judgments to be answered.

[0033] In a further improvement to this embodiment, the routing controller makes decisions based on the query type label. When the type is numerical statistics, the routing controller generates a numerical calculation subtask and a semantic generation subtask. The numerical calculation subtask includes the time range to be retrieved (e.g., "past week"), the metric name (e.g., "sleep duration"), the filtering conditions (e.g., "less than 7 hours"), and the required statistical operations (e.g., "counting"). The semantic generation subtask is a placeholder that will be used later to convert the calculation results into natural language.

[0034] When the type is semantic interpretation, the routing controller generates only one or more semantic generation subtasks, and does not generate numerical computation subtasks. For example, if a user asks "Why do I feel palpitations after drinking coffee in the morning?", the semantic generation subtask will instruct the semantic channel to retrieve explanatory content about the relationship between caffeine and heart rate from the medical knowledge base.

[0035] When the type is multi-step reasoning, the routing controller's processing becomes more complex. Taking the query "In the past month, on the night after each time I reported 'anxiety,' was my deep sleep duration shorter than usual?" as an example, the controller breaks it down into the following subtasks arranged in sequence: The first subtask is a numerical calculation, extracting all dates marked "anxiety" from the emotional anchors over the past 30 days; the second subtask is a numerical calculation, extracting the deep sleep duration value for the following night for each anxiety day; the third subtask is a numerical calculation, randomly sampling the same number of non-anxiety days as a control group and extracting the corresponding deep sleep duration; the fourth subtask is a numerical calculation, calculating the mean difference in deep sleep duration between the two groups, the p-value of the independent samples t-test, and the effect size; finally, a semantic generation subtask synthesizes the above numerical results and organizes them into a natural language answer with empathetic expression and conclusive description. All subtasks are encapsulated into a task queue, arranged in execution order, and passed to subsequent modules.

[0036] S4: Invoke the computation channel to retrieve and compute from short-term cached data and / or long-term compressed memory data according to the numerical computation subtask, and obtain accurate numerical results.

[0037] Preferably, the computation channel is a dedicated module independent of the semantic channel, a health data query engine built on a graph neural network. This engine does not understand natural language; it is only responsible for performing precise numerical retrieval and computation operations. Upon receiving a numerical computation subtask, the computation channel first parses the specified query time range and query metrics in the subtask. For example, if the subtask requires extracting "sleep duration" for the "past week," the computation channel will map this time range to specific start and end timestamps and determine the data source to be accessed: short-term cached data (if the time range is within the last 72 hours) or long-term compressed memory data (if the time range is earlier).

[0038] In one alternative implementation, short-term cached data is organized as a time-series database, supporting fast indexing by timestamp and feature name. Long-term compressed memory data is stored as a sequence of event primitives, which the query engine needs to filter based on the primitive type. For example, "deep sleep duration" can be directly read from the "deep sleep period" primitive, and the total duration is obtained by summing multiple primitive fragments. For abnormal event fragments that are not compressed into primitives, the query engine reads the downsampled feature values ​​and performs approximate calculations.

[0039] The computation channel includes a rich set of numerical operators: numerical statistics include summation, counting, mean, median, standard deviation, and percentiles; trend calculations include linear regression slope, moving average, and cumulative sum; pattern matching includes time series similarity (dynamic time-warped distance) and waveform matching; hypothesis testing includes t-test, chi-square test, and Wilcoxon signed-rank test, outputting p-values ​​and effect sizes. All operations return precise numerical results in floating-point or integer form, along with a confidence interval: for approximations recovered from compressed data, the confidence interval is estimated based on the compression loss model; for results calculated directly from the original short-term cache, the confidence interval is set at a 99% confidence level.

[0040] For example, when the query is "number of days in the past week with less than 7 hours of sleep", the calculation channel first retrieves the total sleep duration for each day from the short-term cache (if all of the past week was within 72 hours) or the duration of the "sleep period" primitive from the long-term compressed memory. Then, it compares each day to see if it is less than 7 hours, and finally returns the count value "3 days" and the confidence interval "±0 days". All intermediate processes do not involve any language model generation, ensuring the accuracy and repeatability of the numerical results.

[0041] S5: Invoke the semantic channel, generate natural language response fragments based on semantic subtasks and combine them with precise numerical results, and output the complete dialogue response.

[0042] After receiving the precise numerical results from the semantic generation subtask and the computation channel, the semantic channel begins to organize the response content. Unlike end-to-end models that directly generate answers, the semantic channel in this scheme is designed for "fill-in-the-blank" generation: first, a predefined response template is selected based on the type of the subtask; then, the numerical results are filled into the placeholders in the template; finally, the template is refined with rhythm and empathy. For example, for a numerical statistical query, the basic template is: "In the past [time range], the number of days you [condition] is [numerical] days." After filling, it becomes "In the past week, the number of days you slept less than 7 hours is 3 days." Then, the semantic channel identifies the user's emotional tendency (e.g., anxiety or calmness) in the current conversation history through a lightweight emotion-aware module and automatically adds an empathy prefix: "It sounds like you are concerned about your sleep quality. In the past week, the number of days you slept less than 7 hours is 3 days." If the user is in a low mood, a suggestion is added: "This may indicate that your stress level is high; trying some breathing exercises before bed might help."

[0043] For multi-step reasoning queries, the semantic channel needs to integrate multiple numerical results into a coherent narrative according to the order of subtasks. Taking the example above regarding the association between anxiety and deep sleep duration, the semantic channel receives four numerical results: number of anxious days = 9 days, average deep sleep duration the day after an anxious day = 83 minutes, average deep sleep duration on non-anxious days = 111 minutes, mean difference = -28 minutes, p-value = 0.03, and effect size = 0.62. The semantic channel feeds this data into a comprehensive template: “In the past month, you marked anxiety on [number of anxiety days] days in your emotional sampling; on nights following these days, your average deep sleep duration was [deep sleep duration on anxiety days] minutes, while your average deep sleep duration on normal (non-anxiety days) was [deep sleep duration on non-anxiety days] minutes. The difference is [mean difference] minutes, and statistical tests show that this difference is significant (p=[p value], effect size=[effect size]). This may suggest that stress has a cascading effect on your nighttime recovery quality.” The final output is a complete dialogue response that includes both precise figures and medical interpretation.

[0044] S6: Based on the sensor data collected in real time by the wearable device, the current emotional state extracted from the calibrated physiological feature sequence, and the current hidden health status obtained by recursion, a three-factor weighted scoring process is performed to obtain the trigger score.

[0045] This solution not only passively responds to user queries but also possesses proactive intervention capabilities. In the background, the system continuously calculates a three-factor weighted score. First, it acquires real-time sensor data from the wearable device's accelerometer, gyroscope, ambient light sensor, and optional GPS. This data is then fed into a lightweight random forest classifier running on the edge device. This classifier outputs the current activity context (one of rest, sleep, exercise, commuting, or social) and the spatial context (one of home, office, outdoors, or transportation). Based on an experience mapping table, the context combination is converted into a context weight w. C For example, during sleep... C Take 0.1 (indicating that it is not advisable to disturb at this time), when resting and at home, w C A value of 0.9 (indicating that users can accept a moderate level of interaction) is used during exercise or commuting. C Take 0.3, and 0.2 for social occasions.

[0046] Secondly, the current emotional state is extracted from the most recent calibrated physiological feature sequence. Specifically, the emotional anchor embedding vector from the past hour is taken, and its correlation with preset emotional dimensions (pleasure, activation) is calculated to obtain an emotional load value e, ranging from 0 to 1; a higher value indicates greater negative emotion or stress. Simultaneously, instantaneous emotions are inferred from real-time physiological signals using a lighter-weight Transformer emotion recognition model. The two are then weighted and fused, and then subjected to an S-shaped function transformation: w E (e) = sigmoid(α·e + β), where α and β are personalized parameters calibrated based on the user's baseline. This transformation makes w corresponding to moderate emotional load (e.g., e=0.5)... E Approximately 0.5, while at extremely low loads (e=0.1) w E Approaching 0, extremely high load (e=0.9) w E Approximately 1. The output is a sentiment score.

[0047] Next, obtain the expected baseline distribution (e.g., the normal range for the same period in the past 30 days) in the current health hidden state vector, and the calibrated physiological characteristic sequence at the current moment (heart rate, HRV, EDA, etc. in the most recent minute). Calculate the Mahalanobis distance between the two, which considers the covariance relationship between each dimension and reflects the degree of multivariate deviation better than the Euclidean distance. Transform the Mahalanobis distance into the prior probability h of the acute health event using a pre-trained logistic regression model, with a value ranging from 0 to 1. Then, apply an exponential weighting function to h: w H (h) = h γγ takes a value greater than 1 (e.g., 2), and its function is to further suppress the score in the low-probability area and retain or even amplify the signal in the high-probability area, avoiding a large number of invalid alerts caused by small fluctuations. The output is a health risk dimension score.

[0048] Finally, calculate the overall trigger score: Score = α·w C + β·w E + γ·w H Here, α, β, and γ are learnable weight coefficients, and their sum is 1. These coefficients are not fixed but are dynamically adjusted based on the user's historical feedback through reinforcement learning: if the user frequently manually turns off reminders or expresses aversion to reminders, the system will reduce the weights of β and γ and increase α (contextual weight) to reduce disturbances in noisy or inconvenient scenarios; conversely, if the user has experienced a health event that was not detected in time, the system will correspondingly increase the weight of the health risk dimension.

[0049] S7: Based on the comparison between the trigger score and the preset threshold, output the first-level intervention instruction, the second-level intervention instruction, or the third-level intervention instruction.

[0050] In this embodiment, three configurable thresholds are preset: a first low threshold T1 (typically set to 0.25), a second medium threshold T2 (0.55), and a third high threshold T3 (0.85). These thresholds can be adjusted according to the user's sensitivity and security needs. The Score calculated by the previous module is compared with these three thresholds.

[0051] If Score < T1, the current user's state is determined to be stable and the situation is relatively sensitive (such as sleeping or socializing), so no intervention instructions are output and the system remains silent and listens.

[0052] If T1 ≤ Score < T2, a Level 1 intervention command is triggered. This command is sent wirelessly to the wearable device, which then alerts the user with a gentle vibration (at the lowest preset intensity, lasting 0.5 seconds). Simultaneously, a line of empathetic text (e.g., "We've noticed your heart rate fluctuating a lot lately, would you like to talk?") and a minimalist emotion sampling input box pop up on the device's miniature screen. The user can ignore this alert, and the system will not repeat it. This level aims to obtain the user's emotional feedback with minimal intrusion while providing an emotional outlet.

[0053] If T2 ≤ Score < T3, a second-level intervention instruction is triggered. Before executing this instruction, the system first checks the user's state: it checks whether the wearable device's accelerometer has detected large-range arm movements (e.g., waving arms) within the last 30 seconds, and whether the ambient light sensor indicates dim indoor lighting (which may indicate the user is resting). If the system detects that the user is at rest and not moving significantly, it considers the current situation suitable for voice interaction. At this time, the wearable device plays a pre-recorded breathing guidance audio (1-2 minutes long) through a bone conduction speaker, or initiates a short voice conversation, such as asking, "You sound a bit stressed. Would you like me to do a two-minute relaxation exercise with you?" If the system detects that the user is in a dynamic or social situation, it does not execute voice prompts but instead uses vibration.

[0054] If Score ≥ T3, and context weight w C Indicates that the user has the ability to respond (i.e., w) C If the score is >0.3 (excluding sleep and social situations), a Level 3 intervention command is triggered, representing the highest level of emergency response. This command simultaneously performs multiple actions: controlling the wearable device to vibrate continuously three times at maximum intensity, each time for 0.3 seconds, with a 0.2-second interval; emitting a rapid alarm sound (volume adjustable) through a bone conduction speaker; automatically generating an emergency message on the associated mobile terminal, containing the user's current geographical location (obtained from GPS or network location), a summary of key physiological indicators for the past 5 minutes (e.g., heart rate, HRV, blood pressure estimates), and a pre-set list of emergency contacts. This message is pushed to pre-set contacts (e.g., family members or emergency rescue services) via cellular network or Wi-Fi. Simultaneously, a prominent "Seek immediate medical attention" prompt is displayed on the device screen, along with a one-click emergency call button. After each intervention command is issued, the system records the time of the triggering event, the scores for each dimension, and whether the user responded, serving as feedback data for subsequent reinforcement learning adjustments to the weight coefficients.

[0055] This solution realizes a complete closed loop from emotional-physiological perception, long and short-term memory management, intelligent routing parsing, accurate numerical calculation, empathic response generation to hierarchical active intervention, providing wearable health devices with an intelligent agent data processing method that combines semantic understanding and numerical accuracy.

[0056] Example 2

[0057] like Figure 3 As shown, this embodiment uses a patient XX (65 years old, female) who underwent left total knee replacement surgery as an example to demonstrate the practical application of the above-mentioned health conversational intelligent agent data processing method based on emotion perception during her home rehabilitation period 8 weeks after surgery.

[0058] The patient was discharged home on the 3rd day after surgery. Medical instructions included performing ankle pump exercises 4 times daily, passive knee flexion exercises 2 times daily, wearing compression stockings, and recording pain scores (0-10). The patient also had mild hypertension and anxiety tendencies. Wearable devices included a smartwatch (integrating PPG, ECG, IMU, EDA, and temperature sensors) and a lumbar fixed posture sensor, both connected to the patient's mobile phone (edge ​​gateway) via Bluetooth, which in turn communicated with cloud services.

[0059] 1. Emotional-physiological dual-flow alignment: Early postoperative pain and mood monitoring

[0060] On the first day after the patient's discharge at 9:30 AM, the wearable device detected the following multimodal physiological signals: heart rate increased from 78 bpm to 98 bpm, HRV RMSSD decreased from 32 ms to 18 ms, skin conductance increased from 2.5 microsiemens to 5.2 microsiemens, and body surface temperature was 1.2°C higher around the left knee than the right. After bandpass filtering (0.5-5 Hz) and IMU-based motion artifact suppression, the feature extraction module calculated the HRV frequency domain LF / HF ratio to be 2.8 (baseline 1.2), the EDA response decay time constant to be extended to 4.5 seconds (baseline 2.1 seconds), and the acceleration signal sample entropy value to decrease from 0.6 to 0.3. These features were mapped into a 128-dimensional physiological embedding vector through a fully connected layer.

[0061] Simultaneously, the system detected a time-triggered condition (9:30 AM being the preset "morning mood sampling window") and displayed a prompt on the patient's watch screen: "Please describe your current feelings in three words." The patient entered "pain, annoyance, fatigue." This extremely short text was encoded into an emotional embedding vector by a lightweight emotion encoder (a 5-layer Transformer distillation model), and the original text was then deleted. Using the emotional embedding vectors generated within the same time window (9:28-9:32) as the query matrix and the physiological embedding vectors as the keys and values, a weighted fused feature vector was calculated through a multi-head cross-attention layer. This vector was appended to the physiological features to generate a calibrated physiological feature sequence with emotional labels. This sequence clearly indicated that "the physiological stress response at 9:30 AM is highly correlated with negative emotions," rather than simply a physiological abnormality. This semantic prior provided crucial information for subsequent inference: the patient's pain at this time was not only a signal of tissue damage but also accompanied by significant emotional distress.

[0062] 2. Dual-granularity rolling memory and health hidden state update: Postoperative rehabilitation trajectory modeling

[0063] From day 1 to day 3 post-surgery, all calibrated physiological characteristic sequences were stored in a 72-hour circular buffer on the mobile phone (edge ​​gateway) at a granularity of one sample point per minute. For example, during the ankle pump exercise at 2:00 PM on day 1, the data stored included a 15-minute curve showing the heart rate changes from 82 to 112 to 96, the waveform showing the knee joint angle changes, and the patient's actively input emotional anchors of "soreness" and "swelling." This fine-grained data allows patients or doctors to review physiological responses at precise time points.

[0064] On the 4th day post-surgery, the data from the 1st day exceeded the 72-hour window and entered the compression process. The system dynamically time-warped the ankle pump exercise data from 2:00 PM on the 1st day against the "HRV response pattern during exercise" in the predefined health event primitive type library, calculating a match confidence of 0.89 (greater than the threshold of 0.7). Therefore, only one record was stored: Health event primitive identifier = 0x03 (corresponding to "HRV response pattern during exercise"), duration stamp = 2025-05-20 14:00 to 14:15, match confidence = 0.89, and emotional anchor association identifier pointing to "soreness, swelling". The original data, which originally required 1800 floating-point numbers (15 minutes × 120 sample points), was compressed to less than 50 bytes. Meanwhile, two brief drops in heart rate that occurred that afternoon (from 92 beats / min to 58 beats / min, lasting 10 seconds) failed to match any primitives (with a maximum confidence of only 0.32), were downsampled and stored as abnormal event fragments, marked as "suspected arrhythmia", and included in long-term compressed memory.

[0065] The recursive update of the health hidden state is performed at 6:00 AM on day 4. The system reads the health hidden state vector S at the end of day 3. t-1 = [Cardiovascular load index = 0.42, Autonomic nervous system regulation efficiency = 0.38, Sleep recovery quality = 0.51, Daytime fatigue accumulation rate = 0.67], obtain the calibrated physiological characteristic sequence X for the current time (4:00-6:00 AM). t (Including deep sleep heart rate of 55 beats / min, HRV RMSSD=45ms, no emotional anchor), emotional embedding vector E t The input time-aware gated recursive units are empty (because there is no active sampling during this period), and the time interval Δt = 14 hours. These input time-aware gated recursive units are then processed according to formula S. t = f(S t-1 , X t E tThe updated hidden health status was calculated using Δt: the cardiovascular load index decreased to 0.35 (due to good night's rest), the autonomic nervous system regulation efficiency increased to 0.52, the sleep recovery quality increased to 0.68, and the daytime fatigue accumulation rate decreased to 0.44. This status indicates that the patient's postoperative recovery trend is positive, but the autonomic nervous system regulation is still weak, suggesting the need to be vigilant for orthostatic hypotension.

[0066] 3. Routing Decision: Decomposition of User Queries

[0067] On the morning of the 5th day after surgery, the patient said to his phone: "Since the surgery, am I more tired in the afternoon than in the morning? How does my heart rate change when I feel tired?" This natural language query was sent to the semantic channel (medical fine-tuning large language model). After model parsing, the output query type label was "multi-step reasoning type" because the question requires: (a) comparing the degree of fatigue in the morning and afternoon; (b) extracting the heart rate change pattern at the time of fatigue.

[0068] The routing controller breaks down the query into the following subtasks and arranges them in order:

[0069] Numerical calculation subtask 1: Extract the number of times "tired" is marked in the emotional anchor points each morning (8:00-12:00) from the data of the past 5 days, and the corresponding heart rate value at that time.

[0070] Numerical calculation subtask 2: Extract the number of times "tired" is marked in the emotional anchor points every afternoon (13:00-17:00) and the corresponding heart rate value at that time.

[0071] Numerical calculation subtask 3: Perform a paired t-test on the cumulative counts for the morning and afternoon (to compare whether the difference between the morning and afternoon counts each day is significant).

[0072] Numerical computation subtask 4: For all moments marked "tired", extract the percentage change in heart rate relative to the baseline heart rate 10 minutes prior to that moment, and calculate the mean and standard deviation.

[0073] Semantic generation subtask: Summarize the above numerical results to form a natural language response.

[0074] 4. Calculation Channel Retrieval and Precise Calculation

[0075] After receiving the above subtasks, the computation channel (a data query engine based on graph neural networks) performs the following operations:

[0076] First, a query was performed on the short-term cache data (the most recent 72 hours, i.e., days 2-4 post-surgery): Day 2, marked "tired" once in the morning (10:30, heart rate 92 bpm), and twice in the afternoon (14:20, heart rate 104 bpm, 16:10, heart rate 98 bpm); Day 3, 0 times in the morning, and 2 times in the afternoon (15:00, heart rate 110 bpm, 16:30, heart rate 102 bpm); Day 4, 1 time in the morning (11:00, heart rate 88 bpm), and 3 times in the afternoon (13:40, heart rate 106 bpm, 15:20, heart rate 100 bpm, 17:00, heart rate 96 bpm). For the data from the morning of Day 5 (the current day), a real-time retrieval from the real-time cache was performed: 0 times in the morning. No data was found in the afternoon.

[0077] Next, data from the long-term compressed memory data were retrieved for the first day after surgery: On the morning of the first day (8:00-12:00), the patient was not fully awake from anesthesia and there was no emotional sampling record, so it was recorded as invalid; in the afternoon (13:00-17:00), "tired" was marked once (heart rate 118 beats / min at 15:30).

[0078] Calculate the channel to perform statistical operations:

[0079] Total number of "accumulated" times in the morning: 1 time on day 2 + 0 times on day 3 + 1 time on day 4 + 0 times on day 5 = 2 times, effective days: 4 days, average 0.5 times / day.

[0080] The total number of "accumulated" times in the afternoon is: 1 time on day 1 + 2 times on day 2 + 2 times on day 3 + 3 times on day 4 = 8 times. The effective number of days is 4, with an average of 2 times per day.

[0081] Paired t-test: Input the difference between morning and afternoon on the same day (day 2 -1, day 3 -2, day 4 -2, day 5 0, day 1 missing) into the t-tester and get the t-value = -3.24, degrees of freedom = 3, p-value = 0.048, effect size Cohen's d = 1.87.

[0082] Heart rate variability: Changes in heart rate relative to baseline (average heart rate over the previous 10 minutes) at all "tired" moments: Day 1 afternoon +22%, Day 2 morning +15%, afternoon +28% and +21%, Day 3 afternoon +30% and +19%, Day 4 morning +12%, afternoon +25%, +18% and +20%, Day 5 none. The calculated mean is +21.2%, and the standard deviation is 5.7%.

[0083] The calculation channel returned precise numerical results: the average number of "tired" times in the afternoon was 2.0 times / day, and in the morning it was 0.5 times / day, p=0.048, the average increase in heart rate was 21.2% (±5.7%), and the confidence interval (estimated based on the reconstruction error of compressed data) was ±2%.

[0084] 5. Semantic channel generates response

[0085] The semantic channel receives the above numerical results and generates a response by combining a preset template and empathy adjustment. Since the patient currently has no acute emotional abnormalities (the emotional anchor point for the last 30 minutes is "okay"), the empathy module outputs a neutral tone:

[0086] "Based on your rehabilitation data over the past 5 days, I noticed that you reported feeling 'tired' a significantly higher number of times each afternoon (average 2 times) than in the morning (average 0.5 times), and statistical tests showed this difference to be statistically significant (p=0.048). When you feel tired, your heart rate is on average about 21% higher than at rest, with a typical fluctuation range between 15% and 30%. This is common in the early post-knee replacement surgery period because afternoon rehabilitation training (such as knee flexion exercises) and cumulative pain can lead to sympathetic nerve activation. One suggestion: you can break your high-intensity afternoon training into two shorter sessions with a 20-minute rest in between, which may help reduce fatigue."

[0087] The watch then displays a simplified version of the text, which is simultaneously read aloud through a bone conduction speaker. The patient listens and replies, "Okay, I'll try."

[0088] 6. Three-factor weighted scoring: proactive intervention decision-making

[0089] At 3:20 PM on the 7th day after surgery, the real-time data from the wearable device sensors is as follows:

[0090] Contextual dimensions: Accelerometer readings show the patient sitting on a sofa (activity context = resting), GPS displays the home address (spatial context = home), and the ambient light sensor reading is 120 lux (normal indoor lighting). The context weights w are mapped from the output of the random forest classifier. C = 0.85 (resting + home is a highly disturbable state).

[0091] Emotional Dimension: Within the past hour, the patient input "annoyed, pain" in an emotional sampling at 2:45 PM, with an emotional load value (e) calculated to be 0.78. The transient emotion recognition model inferred a high activation dimension (tension) from the recent 5-minute HRV (RMSSD = 22ms) and EDA (level 4.1μS). Weighting function w E (e) = sigmoid(α·e + β), where α = 2.5, β = -1.2 (personally calibrated based on the patient's postoperative baseline), and w is calculated. E = sigmoid(2.5×0.78 -1.2) = sigmoid(0.75) = 0.68.

[0092] Health risk dimension: The expected baseline for the cardiovascular load index in the current hidden health state is 0.40 (normal range for 3:00-4:00 PM), and the current actual value is 0.71; the Mahalanobis distance is calculated to be 2.4. The logistic regression model (parameters trained with preoperative data) converts the Mahalanobis distance into a prior probability of acute events, h = 0.32 (high, as the postoperative recovery period threshold is more sensitive). The weighting function is wH(h) = h γ Taking γ=1.5, we get w H = 0.32 1.5 = 0.181.

[0093] The current values ​​of the weighting coefficients α, β, and γ are 0.2, 0.5, and 0.3, respectively (dynamically adjusted based on patient feedback to past reminders: the patient previously reported "Don't disturb me while I'm watching TV," so the contextual weight was reduced; the patient responded positively to the heart rate abnormality reminder, so the health risk weight was appropriately retained). The trigger score is calculated as: Score = 0.2 × 0.85 + 0.5 × 0.68 + 0.3 × 0.181 = 0.17 + 0.34 + 0.0543 = 0.5643.

[0094] The preset thresholds are T1=0.25, T2=0.55, and T3=0.85. Since 0.5643 ≥ 0.55 and < 0.85, a secondary intervention instruction is triggered.

[0095] 7. Implementation of Tiered Intervention: Specific implementation of secondary intervention

[0096] The Level 2 intervention instruction requires cognitive guidance in a moderately invasive manner. The system first performs user status detection: the accelerometer shows that the arm movement amplitude was less than 0.1g in the past 30 seconds (at rest), and the ambient light sensor did not detect a sudden drop (not sleep), therefore it is determined that voice interaction is suitable.

[0097] The watch initiated a voice conversation via a bone conduction speaker: "XX, I noticed your heart rate is about 15% faster than usual, and you just expressed 'annoyance' and 'pain.' Would you like me to accompany you for a 2-minute deep breathing exercise? This can help relieve anxiety caused by pain." The patient responded, "Okay."

[0098] The system then played a pre-set breathing guidance audio: "Please follow my rhythm: Inhale... 4, 3, 2, 1, hold... 2, 1, exhale... 4, 3, 2, 1. Repeat four times." Simultaneously, a dynamic waveform was displayed on the watch screen to assist with the breathing rhythm. Two minutes later, the system sampled the patient's emotions again: the patient selected "Feeling better" on the screen and entered "Sour" (indicating a decrease in intensity). The system recorded this intervention as successful, and the subsequent reinforcement learning module will appropriately increase the weight β of the emotion dimension to encourage timely intervention.

[0099] Meanwhile, the system backend generated a log of this intervention, including the trigger time, scores for each dimension, and user response results, which served as training samples for updating the weight coefficients. Since the score did not reach T3, no emergency message push was triggered.

[0100] 8. Supplementary Context: Triggering of the Three-Level Intervention

[0101] At 2:00 AM on the 10th day post-surgery, the patient experienced a sudden drop in heart rate to 42 beats / min (lasting 20 seconds) during sleep. Simultaneously, the RMSSD of HRV decreased sharply from 35 ms to 8 ms, and skin conductance jumped from 1.1 μS to 3.8 μS. Sensor data generated a calibrated physiological characteristic sequence (no emotional sampling due to sleep state). The cardiovascular load index in the healthy hidden state was 0.28 at the previous time (before sleep), but recursive updates detected an extreme deviation at the current time, with a Mahalanobis distance of 4.7 and a logistic regression output h=0.91. Contextual dimension: Activity context = Sleep (w C =0.1), spatial context = family (unchanged). Emotional dimension: no recent emotional anchor, the instantaneous emotion recognition model infers from physiological signals that it is most likely panic or pain, estimating w E =0.85 (directly output from the model). Weighting coefficients are maintained at α=0.2, β=0.5, γ=0.3. Score is calculated as: Score = 0.2×0.1 + 0.5×0.85 + 0.3×0.91 =0.02 + 0.425 + 0.273 = 0.718. Although the absolute value did not reach T3=0.85, considering the extremely low weight of the nighttime sleep scenario (0.1) which led to the score being suppressed, the system simultaneously detected the prior probability of an acute health event h=0.91 and the severity of the sudden drop in heart rate, triggering the additional condition for Level 3 intervention: when the original h of the health risk dimension score exceeds 0.85 and the absolute heart rate is below 45 beats / min for 10 seconds, it directly enters Level 3 intervention. Therefore, the system outputs the Level 3 intervention instruction: the watch vibrates three times continuously at maximum intensity, while simultaneously emitting a rapid alarm sound through the bone conduction speaker and shouting "Abnormal heart rate, please wake up immediately". The linked mobile phone automatically generated an emergency message: "Patient XX, heart rate dropped to 42 beats / min, lasting 20 seconds, location [latitude and longitude] sent. Immediate medical attention recommended." The message was pushed to the patient's family and community family doctor. After being awakened by the vibration, the patient pressed the emergency button on the side of the watch to confirm consciousness, and the system then guided the patient through the supine awakening procedure. The doctor later determined it to be a postoperative vagal nerve reflex, which improved after adjusting the anticoagulant medication.

[0102] 9. Overall Effectiveness Evaluation

[0103] Within 8 weeks post-surgery, the system processed 124 patient-initiated queries, achieving a 96.2% accuracy rate in correctly answering numerical statistical questions and an 88.5% accuracy rate in answering multi-step reasoning questions. 47 proactive interventions were triggered, including 28 Level 1 interventions (user ignore rate 32%), 17 Level 2 interventions (user acceptance rate 82%, feedback improvement rate 76%), and 2 Level 3 interventions (both successfully roused the patient or elicited a response from family members). Compared to traditional wearable devices that only monitor heart rate and steps, this solution captured the evolution of the patient's post-operative psychological distress through emotional anchors (from "fear + pain" on days 1-3 to "irritability + fatigue" on days 4-7, and then to "boredom + anticipation" on days 8-14), and personalized rehabilitation recommendations and intervention timing accordingly. During a follow-up visit 14 days after discharge, the patient stated, "It's not like a cold, impersonal machine; it asks me how I'm feeling and knows whether I'm genuinely uncomfortable or just venting." This example fully validates the effectiveness of this technical solution in the surgical rehabilitation scenario.

[0104] This invention has the following advantages: First, through an emotion-physiology dual-stream alignment mechanism, it utilizes the deep fusion of ultra-short-term active emotion sampling and passive physiological perception from wearable devices to acquire the psychological contextual semantics driving physiological changes with a low user burden (no more than twice a day, requiring only 1-5 words each time). This effectively bridges the semantic gap in traditional health monitoring where "they know what but not why," allowing physiological signals such as heart rate fluctuations and HRV changes to be accurately attributed to specific psychological states such as pain, anxiety, or fatigue, thereby improving the accuracy of health assessment. Second, by employing health event primitive compression and dual-granularity rolling memory technology, continuous physiological data is automatically matched into standardized primitives such as "sleep onset period," "deep sleep period," and "stress recovery curve" for storage. This compresses 90 days of health data to less than 3% of the original amount, while supporting health pattern comparisons and trend analysis across weeks to months, resolving the contradiction between storage and inference in long-term health monitoring. Furthermore, the separate dual-channel intelligent agent architecture decouples numerical computation from semantic generation. A graph neural network-based computational channel is dedicated to precise retrieval and statistics (including mean, trend, t-test, correlation, etc.), while a medically fine-tuned large language model is dedicated to empathic expression and interpretation. This overcomes the error-proneness and unreliability of general-purpose large language models in health time-series calculations, with an expected numerical query accuracy exceeding 90%. Finally, the three-factor adaptive proactive reminder model comprehensively considers the user's current activity context (sleep, rest, exercise, etc.), real-time emotional load, and the probability of acute health risks, dynamically adapting to a three-tiered intervention strategy ranging from gentle vibration to emergency alarms. This reduces ineffective disturbances by over 60% while ensuring health and safety. In summary, this invention provides a novel data processing solution for wearable health devices that combines psychological intelligence, long-term memory, precise reasoning, and humanized interaction.

[0105] The preferred embodiments of the present invention have been described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the protection scope of the present invention.

Claims

1. A data processing method for a health-oriented conversational intelligent agent based on emotion perception, characterized in that, include: Based on the multimodal physiological signals collected in parallel by wearable devices and the ultra-short emotional text actively input by users, timestamp alignment and cross-attention fusion processing are performed to obtain a calibrated physiological feature sequence with emotional labels. The calibration physiological feature sequence is subjected to dual-granularity rolling memory processing to obtain short-term cache data and long-term compressed memory data, and the current health hidden state is recursively updated based on the calibration physiological feature sequence and long-term compressed memory data; Receive user natural language health queries, perform routing decision processing on user natural language health queries, and obtain at least one numerical computation subtask and at least one semantic generation subtask; The computation channel is invoked to retrieve and compute from short-term cached data and / or long-term compressed memory data according to the numerical computation subtask, so as to obtain accurate numerical results; Invoke the semantic channel, generate natural language response fragments based on semantic subtasks and combine them with precise numerical results, and output the complete dialogue response; Based on the sensor data collected in real time by the wearable device, the current emotional state extracted from the calibrated physiological feature sequence, and the current hidden health status obtained by recursion, a three-factor weighted scoring process is performed to obtain the trigger score. Based on the comparison between the trigger score and the preset threshold, the system outputs a first-level intervention instruction, a second-level intervention instruction, or a third-level intervention instruction.

2. The method according to claim 1, characterized in that, Based on multimodal physiological signals acquired in parallel by wearable devices and ultra-short emotional text actively input by users, timestamp alignment and cross-attention fusion processing are performed to obtain a calibrated physiological feature sequence with emotional labels, including: Signal conditioning and feature extraction are performed on the multimodal physiological signals collected by wearable devices to obtain physiological embedding vectors; When the time-triggered, event-triggered, or context-triggered conditions are met, the system receives a very short emotional text input by the user, inputs the very short emotional text into a lightweight emotional encoder for encoding processing, and obtains an emotional embedding vector. The sentiment embedding vectors generated within the same time window are used as the query matrix, and the physiological embedding vectors are used as the key matrix and value matrix. The sentiment-physiological association weights are calculated through a multi-head cross-attention layer to generate a calibrated physiological feature sequence with sentiment labels.

3. The method according to claim 2, characterized in that, Signal conditioning and feature extraction are performed on multimodal physiological signals collected by wearable devices to obtain physiological embedding vectors, including: Photoplethysmography (PPG) and electrocardiogram (ECG) signals are acquired using PPG / ECG sensors, triaxial acceleration signals are acquired using an IMU accelerometer, EDA signals are acquired using a skin conductance sensor, and body surface temperature signals are acquired using a temperature sensor to obtain multi-channel raw physiological signals. By applying a bandpass filter and motion artifact suppression algorithm to the multi-channel raw physiological signals, filtered physiological time series are obtained. The HRV time-domain features, HRV frequency-domain features, skin conductance response amplitude and attenuation constant, differential rate of change of body surface temperature, and acceleration signal sample entropy value are calculated on the filtered physiological time series to obtain a multidimensional feature vector. The multidimensional feature vectors are mapped to 128-dimensional physiological embedding vectors through a fully connected layer.

4. The method according to claim 1, characterized in that, The calibrated physiological characteristic sequences are processed using a two-granularity rolling memory method to obtain short-term cached data and long-term compressed memory data. The current health hidden state is then recursively updated based on the calibrated physiological characteristic sequences and the long-term compressed memory data, including: A circular buffer is created at the edge gateway layer, and the calibrated physiological characteristic sequences for each minute of the most recent 72 hours are written into the circular buffer as short-term cache data. For historical calibrated physiological feature sequences exceeding 72 hours, dynamic time warping matching is performed with a predefined health event primitive type library to obtain the matching health event primitive identifiers and matching confidence. If the matching confidence is greater than the confidence threshold, only the health event primitive identifier, duration stamp, matching confidence, and corresponding emotional anchor association identifier are stored to form long-term compressed memory data; if the matching confidence is less than or equal to the confidence threshold, the corresponding historical calibrated physiological feature sequence is downsampled and stored as an abnormal event fragment, and included in the long-term compressed memory data. Read the health hidden state vector of historical moments, obtain the calibration physiological feature sequence of the current moment, the emotion embedding vector of the current moment, and the time interval between the current moment and the historical moment. Input the health hidden state vector, calibration physiological feature sequence, emotion embedding vector, and time interval into the time-aware gated recursive unit for recursive calculation to obtain the updated current health hidden state vector.

5. The method according to claim 1, characterized in that, The system receives user natural language health queries, performs routing decision processing on these queries, and obtains at least one numerical computation subtask and at least one semantic generation subtask, including: The user's natural language health query is input into the semantic channel, which is a medical fine-tuning large language model, and the query type label is output. The query type label is one of numerical statistical type, semantic interpretation type or multi-step reasoning type. When the query type label is numerical statistics, the route controller generates a numerical calculation subtask and a semantic generation subtask. When the query type label is semantic interpretation type, the route controller generates only at least one semantic generation subtask; When the query type label is multi-step inference, the routing controller decomposes the user's natural language health query into multiple subtasks, which include at least two numerical calculation subtasks and at least one semantic generation subtask, and are arranged in the order of execution.

6. The method according to claim 5, characterized in that, The computation channel is invoked to retrieve and compute precise numerical results from short-term cached data and / or long-term compressed memory data according to the numerical computation subtask, including: The computation channel is a health data query engine built on graph neural networks; Based on the query time range and query metrics specified in the numerical computation subtask, retrieve the corresponding data fragments from short-term cached data or long-term compressed memory data. Perform numerical statistics, trend calculation, pattern matching, or hypothesis testing on a data segment, and output precise numerical results and confidence intervals in floating-point or integer form.

7. The method according to claim 1, characterized in that, Based on real-time sensor data collected by the wearable device, the current emotional state extracted from the calibrated physiological characteristic sequence, and the current hidden health status obtained through recursive updates, a three-factor weighted scoring process is performed to obtain the trigger score, including: Based on real-time sensor data, the current activity context and spatial context are identified, and the identification results are mapped to context weights w. C ; Based on the emotional anchor embedding vector within the most recent time period and the instantaneous emotional state inferred from physiological signals in real time, the emotional load value e is calculated, and the emotional load value is then weighted by the function w. E The emotional dimension score is obtained by transforming (e) = sigmoid(α·e+β). Based on the expected baseline distribution of the calibrated physiological characteristic sequence and the current health hidden state vector at the current moment, the Mahalanobis distance is calculated as the degree of deviation. The degree of deviation is then converted into the prior probability h of the acute health event through a logistic regression model, and the prior probability h is then weighted by a weighting function w. H (h)=h γ The transformation yields a health risk dimension score, where γ > 1; Calculate the trigger score: Score = α·w C + β·w E + γ·w H , where α+β+γ=1, and α, β, and γ are dynamically adjusted based on the user's historical feedback through reinforcement learning.

8. The method according to claim 7, characterized in that, Based on real-time sensor data collected by wearable devices, the current activity context and spatial context are identified, and the identification results are mapped to context weights w. C ,include: Real-time sensor data is acquired from the wearable device's accelerometer, gyroscope, ambient light sensor, and optional GPS. The current activity context is identified using a lightweight random forest classifier, and the activity context includes at least one of rest, sleep, exercise, commuting, and social interaction. The current spatial context is identified using a lightweight random forest classifier, wherein the spatial context includes at least one of home, office, outdoors, and vehicle. The combination of activity context and spatial context is mapped to context weight w. C During sleep, w C The value is 0.1, and w is in the resting state. C The value is 0.9; in other scenarios, w... C The value ranges between 0.1 and 0.

9.

9. The method according to claim 1, characterized in that, Based on the comparison between the trigger score and the preset threshold, the system outputs a first-level intervention instruction, a second-level intervention instruction, or a third-level intervention instruction, including: The trigger score is compared with the first low threshold, the second medium threshold, and the third high threshold; If the trigger score is less than the first low threshold, no intervention instruction will be output; If the trigger score is greater than or equal to the first low threshold and less than the second medium threshold, a first-level intervention instruction is output. The first-level intervention instruction is used to control the wearable device to push empathic text and emotion sampling input box in a gentle vibration manner. If the trigger score is greater than or equal to the second threshold and less than the third high threshold, a second-level intervention instruction is output. The second-level intervention instruction is used to play breathing guidance audio or initiate a voice dialogue through a bone conduction speaker after detecting that the user is in a resting state and has no large movements. If the trigger score is greater than or equal to the third highest threshold, and the context weight w C If the user is deemed capable of responding, a third-level intervention command is output. This third-level intervention command is used to simultaneously trigger a strong vibration and voice alarm on the wearable device, and automatically generate an emergency message containing the user's location and health summary to be pushed to a preset contact.

10. The method according to claim 1, characterized in that, The method also includes edge-cloud collaborative computing and privacy protection processing: After extracting features from the multimodal physiological signals and the ultra-short emotional text on the wearable device to obtain physiological embedding vectors and emotional embedding vectors, the original data of the multimodal physiological signals and the original data of the ultra-short emotional text are immediately deleted. Differential privacy scrambling is applied to the physiological and emotional embedding vectors uploaded to the cloud to obtain scrambled embedding vectors; Short-term cached data is stored at the edge gateway layer, while long-term compressed memory data is stored in the cloud. The cloud only stores the embedded vectors and health event primitive sequences after differential privacy scrambling, and does not store the original signals.