Intervention information pushing method and device

Abstract

The invention relates to the field of data processing, and provides an intervention information pushing method and device. The method comprises the steps of performing standardization processing on multi-source health data and storing the multi-source health data in a unified data base to generate health feature data; constructing a personal health portrait and a behavior mode based on the health feature data; inputting the personal health portrait, the behavior mode and the health management target into a health decision model to generate a personalized behavior intervention plan; tasks in the behavior intervention plan are pushed to the user through the interaction terminal, and task execution conditions and user feedback are collected; and based on the execution condition and the user feedback evaluation intervention effect, updating the personal health portrait and the behavior mode to adjust the behavior intervention plan. Through a unified data base, a decision model and a behavior intervention mechanism, user behavior changes are guided in multiple scenes, and the health management efficiency is improved.

Description

Technical Field

[0001] This application relates to the field of data processing, and more specifically, to a method and apparatus for disseminating intervention information. Background Technology

[0002] In recent years, with the popularization of wearable devices, mobile internet, and cloud computing, digital health management tools for the general public have developed rapidly. These systems typically focus on a few quantifiable indicators such as steps, heart rate, sleep duration, and weight, and provide users with basic functions such as exercise check-in reminders, diet records, and sleep check-ins through fixed threshold rules or experience templates.

[0003] However, existing technologies mostly remain at the stage of data collection, recording and display, and simple reminders, lacking in-depth modeling of individual user differences, behavioral patterns, and long-term health goals, making it difficult to form a continuous and effective closed loop of behavioral guidance. Most applications rely on fixed thresholds or experience templates to provide general suggestions, lacking deep personalization and dynamic adjustment mechanisms based on individual differences, behavioral habits, and long-term goals. Analysis dimensions are mostly limited to superficial rule matching, making it difficult to achieve true intelligent coaching. Furthermore, health services lack integrated coordination across multiple dimensions such as exercise, diet, sleep, and psychology, requiring users to switch between multiple applications and manually integrate scattered information. Therefore, there is an urgent need to propose a technical solution to address at least one of the above-mentioned technical problems. Summary of the Invention

[0004] In this context, the embodiments of this application aim to provide an intervention information push method and apparatus, which is a technical solution that can continuously guide changes in user behavior in multiple scenarios, and improve health management efficiency through a unified data foundation, intelligent decision-making model and behavior intervention mechanism.

[0005] In a first aspect of the embodiments of this application, an intervention information push method is provided. The method includes: acquiring multi-source health data, standardizing the multi-source health data and storing it in a unified data foundation to generate health feature data; constructing a personal health profile and behavior pattern based on the health feature data; the health feature data is related to the current scenario of the target user; receiving health management goals for the target user, inputting the personal health profile, behavior pattern and health management goals into a health decision model to generate a personalized behavior intervention plan including phased goals and specific task suggestions; pushing tasks in the behavior intervention plan to the user through an interactive terminal bound to the user, and collecting task execution status and user feedback; evaluating the intervention effect based on the execution status and user feedback, updating the personal health profile and behavior pattern, and adjusting the behavior intervention plan according to the updated personal health profile and behavior pattern.

[0006] In a second aspect of the embodiments of this application, an intervention information push device is provided. The device includes the following units: a collection unit, used to acquire multi-source health data, standardize the multi-source health data and store it in a unified data base to generate health feature data; collect task execution status and user feedback; a construction unit, used to construct a personal health profile and behavior pattern based on the health feature data; a push unit, used to receive health management goals for target users, input the personal health profile, behavior pattern and health management goals into a health decision model, generate a personalized behavior intervention plan including phased goals and specific task suggestions; push the tasks in the behavior intervention plan to the user through an interactive terminal bound to the user; and an adjustment unit, used to evaluate the intervention effect based on the execution status and user feedback, update the personal health profile and behavior pattern, and adjust the behavior intervention plan according to the updated personal health profile and behavior pattern.

[0007] In a third aspect of the embodiments of this application, a terminal device is provided, the terminal device including: at least one processor, a memory and an input / output unit; wherein, the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the intervention information push method of any of the first aspects.

[0008] This application discloses an intervention information push method and apparatus. The technical solution first acquires multi-source health data, standardizes the data, and stores it in a unified data platform to generate health characteristic data. Then, based on the health characteristic data, a personal health profile and behavioral pattern are constructed; the health characteristic data is related to the target user's current scenario. Next, health management goals for the target user are received, and the personal health profile, behavioral pattern, and health management goals are input into a health decision-making model to generate a personalized behavioral intervention plan containing phased goals and specific task suggestions. Tasks in the behavioral intervention plan are pushed to the user through an interactive terminal bound to the user, and task execution status and user feedback are collected. Finally, the intervention effect is evaluated based on the execution status and user feedback, the personal health profile and behavioral pattern are updated, and the behavioral intervention plan is adjusted according to the updated personal health profile and behavioral pattern. This application's technical solution, through a unified data platform, decision-making model, and behavioral intervention mechanism, continuously guides user behavior changes in multiple scenarios, improving health management efficiency. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating an intervention information push method as shown in this application; Figure 2 This is a schematic diagram of the structure of an intervention information push device shown in this application. Detailed Implementation

[0010] The following is for reference. Figure 1 , Figure 1 This is a flowchart illustrating an intervention information push method according to an embodiment of this application. It should be noted that the implementation methods of this application can be applied to scenarios such as routine physical examinations, daily monitoring, and chronic disease maintenance.

[0011] To address the technical problem that existing technologies in the aforementioned scenarios mostly remain at the stage of data collection, recording and display, and simple reminders, lacking in-depth modeling of individual user differences, behavioral patterns, and long-term health goals, and making it difficult to form a continuous and effective closed loop for behavioral guidance, this application provides an intervention information push method and device.

[0012] Specifically, in this application embodiment, on the one hand, by uniformly standardizing and centrally storing multi-source health data, the problems of scattered data sources, inconsistent formats, and difficulty in fusion modeling in the prior art are solved. This enables multi-dimensional information such as exercise, diet, sleep, and psychology to be associated under a unified time and user index, forming a health feature data layer that can be directly used for intelligent analysis, providing a high-quality data foundation for subsequent accurate modeling. On the other hand, this application embodiment simultaneously constructs personal health profiles and behavioral patterns based on health feature data. It not only quantifies the health level of target users in various dimensions from the perspective of health risk and status, but also models behavioral feature dimensions such as compliance, habit stability, task preference, and situational sensitivity. This allows the model to formulate intervention strategies based on a full understanding of user needs and user habits. On this basis, the personal health profile, behavioral patterns, and health management goals are jointly input into the health decision-making model to generate personalized behavioral intervention plans for different stages of goals. This achieves fine-grained matching of intervention tasks in terms of type, intensity, frequency, and execution timing, overcoming the shortcomings of the prior art that rely on fixed templates and are difficult to reflect individual differences. Furthermore, this application embodiment pushes tasks to users in multiple scenarios via interactive terminals and collects task execution status and subjective feedback in real time. The execution results are fed back to evaluate intervention effectiveness and dynamically update health profiles and behavioral patterns, forming a closed-loop optimization mechanism of guidance, intervention effect feedback, and readjustment of guidance strategies. This allows intervention plans to adaptively adjust as users' health status and behavior change. In summary, this application embodiment not only improves the efficiency of health data utilization and modeling accuracy but also further enhances the individualization and continuous optimization capabilities of intervention strategies. It can continuously and effectively guide user behavior changes in multiple scenarios, thereby improving health management effectiveness and overall operational efficiency.

[0013] Figure 1 The flowchart of an intervention information push method provided in one embodiment of this application includes: Step S101: Acquire multi-source health data, standardize the multi-source health data, and store it in a unified data platform to generate health feature data. Multi-source health data can come from wearable devices, mobile terminals, smart medical devices, and user-entered data.

[0014] In this embodiment, the unified data foundation is the basic data platform supporting subsequent health profile construction and intelligent decision-making model training. As an optional embodiment, in step S101, multi-source health data is standardized and stored in the unified data foundation to generate health feature data. This includes: unifying the timestamps and normalizing the units of the original health data from different devices in the multi-source health data; converting the original indicators of steps, heart rate, sleep stage, food intake, and mood scores into basic features under a unified scale; resampling and time-aligning the basic features according to a preset time granularity to construct a time-series feature matrix indexed by user identifiers and time slices; performing data quality processing on missing and abnormal data in the time-series feature matrix; and based on the processing... The processed time series feature matrix is ​​used to calculate derived health feature indicators, which include at least one of the following: sleep quality score, heart rate variability index, activity intensity level, dietary nutritional balance, and mood fluctuation index. Analytical dimension labels are established for these derived health feature indicators, including at least physiological, psychological, behavioral, and contextual dimensions. The labeled derived health feature indicators are then bound to the target user's index information and stored in a unified database. Based on a profile request for the target user, health feature data matching the target analytical dimensions in the profile request is extracted from the unified database.

[0015] Specifically, the raw data from wearable devices, mobile applications, smart medical devices, and user-entered data are first processed to unify timestamps and normalize units. The local time format used by each data source is mapped to a unified standard timeline, and raw indicators such as steps, heart rate, sleep stage labels, food intake, and mood scores are converted into basic features under a unified scale.

[0016] For example, the energy consumption of certain devices, expressed in kilocalories or kilojoules, can be uniformly converted into standard energy units; the sleep stage division results of different manufacturers can be mapped into unified light sleep, deep sleep, and REM sleep stage identifiers; and the emotion scores obtained from different questionnaires or scales can be normalized into the same scoring range.

[0017] Subsequently, the aforementioned basic features are resampled and time-aligned on a unified time axis according to a preset time granularity, for example, using minutes or hours as time slices. The mean or representative value of high-frequency data such as heart rate is calculated within each time slice, while low-frequency data such as step count is accumulated within each time slice, thereby constructing a time-series feature matrix indexed by user identifier and time slice. Inevitably, missing or anomalous data will exist in this time-series matrix. This embodiment corrects these through data quality processing algorithms. Short-term missing data can be filled using forward imputation or linear interpolation. For long-term missing or significantly anomalous data, estimation or replacement can be performed by combining user historical patterns and population statistical distribution, thereby ensuring the continuity and rationality of the feature sequence.

[0018] After completing the basic feature cleaning and alignment, derived health feature indicators are further calculated based on the time series feature matrix. For example, a sleep quality score is generated by combining the time to fall asleep, total sleep duration, number of awakenings and the proportion of each sleep stage within a night window; a heart rate variability index is calculated based on continuous heart rate changes; an activity intensity level is generated based on the number of steps taken during the day, activity duration and estimated metabolic equivalent; a dietary nutrition balance is generated based on the deviation between the nutrient structure corresponding to the intake of various foods over a period of time and the recommended intake structure; and a mood fluctuation index is generated based on the mean, fluctuation amplitude and state switching frequency of continuous mood scores.

[0019] Furthermore, to support subsequent profiling and decision-making from different analytical perspectives, this embodiment also establishes associated analytical dimension labels for each derived health feature indicator. Indicators reflecting physiological state are labeled as physiological dimensions, indicators reflecting emotional and psychological state are labeled as psychological dimensions, indicators reflecting daily behavioral habits are labeled as behavioral dimensions, and contextual features dependent on time, location, or device status are labeled as situational dimensions. The derived health features with analytical dimension labels are then bound to the target user's index information and stored uniformly in a unified database. Based on this, when an upper-layer service initiates a profiling request for a target user, it can directly retrieve the corresponding health feature data from the unified database by user identifier and target analytical dimension. For example, it can extract only physiological dimension features related to sleep for sleep profiling analysis, or extract behavioral and situational dimension features for behavioral pattern modeling.

[0020] Through the above steps, the unified data foundation in this embodiment not only achieves spatiotemporal alignment and semantic normalization of multi-source heterogeneous health data, but also improves the expressive power and usability of the data through derived feature calculation and dimension label management. This provides high-quality data support for subsequent personalized health profile construction, behavior pattern recognition, and AI-based intervention effect prediction, effectively improving the analysis accuracy and decision-making efficiency of the health management system in complex multi-scenario scenarios.

[0021] Optionally, in step S101, time alignment of multi-source health data from different data sources with different sampling frequencies includes: mapping the timestamps of wearable device data, mobile application record data, and smart medical device uploaded data to a unified time axis; using forward padding, linear interpolation, or interpolation algorithms based on adjacent time slices to fill in intermediate time values ​​for indicators with sampling frequencies lower than the target time granularity; and using summation, average, extreme value, or percentile statistics to aggregate indicators with sampling frequencies higher than the target time granularity within each time slice, thereby obtaining a basic feature sequence aligned at a unified time granularity.

[0022] For example, data from multiple sources—whether wristbands, mobile apps, or home blood pressure monitors—is aligned using an hourly time granularity. Wristbands report heart rate and steps every minute, mobile apps record photos of food and subjective mood scores each time, and blood pressure monitors occasionally upload a blood pressure value during user measurements. First, the timestamps of all three data types are uniformly converted to the same time zone and format. Then, using each hour as a time slice, the average heart rate data from the wristbands within that hour is calculated, and the step count is accumulated. One or more blood pressure measurements taken during this period are then filled into the corresponding time slice based on the most recent measurement. Hours without measurements are left blank. For low-frequency metrics like steps, if an hour has no records but adjacent hours have valid data, interpolation from nearby time slices or referencing the user's historical average for the same period can be used to estimate a reasonable step count to fill the gap, thus forming a basic feature sequence with continuous hourly records.

[0023] Further optionally, the imputation of missing data and the detection and correction of abnormal data include: setting anomaly detection rules based on the value range, statistical distribution and rate of change of each type of basic feature; marking values ​​that exceed reasonable thresholds or have abnormal change magnitudes as abnormal data; correcting abnormal data by truncating, smoothing or replacing them with normal values ​​from adjacent time slices; imputing missing data for short time periods by interpolation of adjacent time slices or historical averages of the same time period; and estimating long-term missing data that crosses a preset duration threshold by using a prediction model based on user profiles and similar population characteristics, or explicitly marking it as a missing feature in the feature matrix.

[0024] Further, optionally, the generation of derived health characteristic indicators based on the processed raw data includes: aggregating sleep-related basic characteristics (such as sleep duration, sleep latency, number of nighttime awakenings, and percentage of deep sleep) according to a preset time window, and calculating a sleep quality score using a weighted summation formula or a trained scoring model; calculating a heart rate variability index based on the instantaneous heart rate changes and heart rate interval changes of a continuous heart rate sequence within the time window; and calculating an activity intensity level based on steps, cadence, activity duration, and estimated energy consumption.

[0025] Furthermore, it can also generate dietary nutritional balance and mood fluctuation indices by: performing nutritional analysis on dietary record data, converting the food consumed at each meal into the intake of nutrients such as energy, protein, fat, and carbohydrates, and comparing it with the recommended intake within a preset statistical period to calculate dietary nutritional balance; and statistically analyzing the time series of results from mood scoring, mood tags, or voice emotion recognition to calculate the positive and negative ratio of emotions, the frequency of mood state switching, and the variance of mood scores within the statistical period, thereby generating a mood fluctuation index.

[0026] Step S102: Construct a personal health profile and behavior pattern based on health characteristic data.

[0027] In this embodiment, health feature data is related to the target user's current scenario. Specifically, based on health feature data in a unified data foundation, a personal health profile and behavioral pattern related to the target user's current scenario are constructed. This health feature data includes not only multi-dimensional health indicators such as exercise, diet, sleep, and mental health, but also scenario information related to time period, geographical location, weekday or rest day, and device online status, ensuring that each feature record corresponds to a specific life situation. For example, if a user is identified as being in a low-activity state for an extended period between 9:00 AM and 12:00 PM on a weekday and is within the office geographical area, then the number of steps, sedentary time, and heart rate during that period are associated with a work scenario tag. Similarly, after 10:00 PM, sleep-related features are associated with a bedtime / nighttime scenario tag. Based on this, health feature data from different scenarios are aggregated and analyzed within a preset statistical period to extract health status and behavioral features related to specific scenarios. This makes the personal health profile no longer a static overall average, but rather able to distinguish health performance in different scenarios such as "working hours," "commuting time," and "bedtime." By using this scenario-based modeling approach, this embodiment can identify typical problems such as "severe sitting on weekday mornings", "excessive energy intake on weekends", and "prolonged sleep latency due to using mobile devices before bed". This provides a basis for developing intervention tasks based on the scenario, thereby improving the pertinence and feasibility of intervention recommendations and further enhancing the effectiveness of health management.

[0028] In this embodiment, the personal health profile is used to characterize the health status and risk distribution of target users across multiple dimensions, including exercise, diet, sleep, and psychology. The personal health profile can be represented in the model structure as a multi-dimensional feature vector plus several discrete labels. Its input includes derived health feature indicators summarized by dimension from a unified data foundation, as well as basic user information. The output is a comprehensive score and risk level for each health dimension. For example, inputting the activity intensity level, resting heart rate, and weight change over the past month into the scoring model yields a "moderate to low exercise dimension score," inputting sleep quality and sleep stability into the model yields a "high sleep dimension risk," and inputting dietary nutritional balance and body fat percentage trends into the model yields a "moderate diet dimension risk." Combined with age, gender, and chronic disease history, this ultimately forms a profile label of "middle-aged sedentary individuals with a tendency towards sleep problems."

[0029] Behavioral patterns are used to characterize the behavioral habits, execution preferences, and compliance of target users in different scenarios and task types. They represent a more dynamic form of behavioral feature representation. Behavioral patterns take task execution records, interaction logs, sleep patterns, and mood fluctuation characteristics as inputs, and learn user behavior patterns in different time periods and task types through clustering or classification models.

[0030] For example, during training, using historical data from multiple users as samples, supervised or semi-supervised learning methods are used to simultaneously optimize the profile scoring model and the behavior pattern classification or clustering model. This ensures that the output health profile and behavior pattern features not only meet the rules of medical and behavioral science but also reflect real-world usage behavior. In practical applications, the personal health profile provides the health decision-making model with information on priority areas for improvement, while the behavior pattern provides the model with analysis results on intervention methods that users are more likely to accept. Combining the two can generate intervention plans that both meet health risk priorities and match user behavioral habits. This further improves the execution rate and long-term adherence of intervention plans while ensuring safety and scientific rigor, thereby enhancing the capabilities of intelligent health coaching and behavior guidance.

[0031] As an optional embodiment, in step S102, constructing a personal health profile and behavioral pattern based on health feature data includes: calculating the comprehensive score and risk level of each target analysis dimension within a preset statistical period based on the extracted health feature data, obtaining a multidimensional health score vector to characterize the health status of multiple dimensions including exercise, diet, sleep, and psychology; classifying the target user into preset population stratification and risk stratification categories based on the multidimensional health score vector and the target user's basic information, including age, gender, and / or chronic disease history, to obtain population stratification labels to characterize health risk levels and management priorities; identifying behavioral stages and clustering behavioral types for the target user based on the historical intervention task completion status, application usage behavior, work-rest patterns, and emotional fluctuation characteristics recorded in the health feature data, to obtain behavioral pattern features to characterize compliance, habit stability, and intervention sensitivity; generating a personal health profile corresponding to the target user based on the multidimensional health score vector and population stratification labels; and generating a behavioral pattern corresponding to the target user based on the behavioral pattern features.

[0032] Specifically, firstly, within a pre-defined statistical period, health characteristic data associated with the target user in a unified data foundation are aggregated and calculated according to analytical dimensions. For target analytical dimensions such as exercise, diet, sleep, and psychology, basic characteristics such as steps, activity intensity, weight change, sleep quality score, dietary nutritional balance, and mood fluctuation index are input into a pre-defined scoring function or a trained scoring model to obtain a comprehensive score for each dimension and a risk level indicating whether it exceeds a safety threshold, thus forming a health score vector reflecting multidimensional health status. For example, within a one-month statistical period, if a user's average daily steps are low and sedentary time is long, the exercise dimension score is low and the sedentary risk is marked as high. If sleep duration is insufficient and there are multiple awakenings at night, the sleep dimension score is low and the sleep risk is marked as high. Based on this, combined with the user's basic information, including age, gender, and whether there is a history of chronic diseases such as hypertension and diabetes, the multidimensional health score vector is mapped to pre-defined healthy population stratification and risk stratification categories, such as young low-risk groups, middle-aged obese high-risk groups, and elderly people with multiple chronic diseases, resulting in population stratification labels used to characterize the overall health risk level and health management priorities.

[0033] Simultaneously, by utilizing historical intervention task completion data, application usage behavior, daily routines, and mood fluctuations recorded in health characteristic data, behavioral characteristics of target users are modeled. Specifically, indicators such as task completion rate, on-time completion rate, consecutive check-in days, and common reasons for interruption are combined with user activity levels, login frequency, report viewing frequency, and sleep-wake time distribution across different time periods to construct behavioral feature vectors. Based on behavioral science theories regarding behavioral change stages, rules are designed or classification models are trained to identify whether a user is currently in the preparation, action, or maintenance stage. Furthermore, the behavioral feature vectors of a large number of users are input into a clustering algorithm to categorize users into behavioral types such as highly self-disciplined, interest-driven, and easily abandoned. Different types exhibit different characteristics in task difficulty tolerance, push notification frequency sensitivity, and preferred time periods. Through behavioral stage identification and behavioral type clustering, a set of behavioral pattern features reflecting user compliance, habit stability, and sensitivity to intervention is obtained. For example, a user can be identified as having "high compliance in the evening, sensitivity to high-frequency reminders, and a tendency to abandon excessively intense exercise tasks."

[0034] Based on the above calculations, the multidimensional health score vector is combined with population stratification labels to generate a personal health profile corresponding to the target user. This profile clearly identifies the current status, main risk points, and priority intervention directions for each health dimension, such as marking information like "low sleep quality, high risk of sedentary lifestyle, and a generally balanced diet but high fat content." Simultaneously, behavioral pattern features are input into the behavioral pattern, solidifying them in a structured form to characterize the user's task acceptance preferences, execution pace preferences, and reminder tolerance. This is used to assess the feasibility and expected compliance of different task combinations at the execution level during subsequent intervention decisions. Through this modeling process, this embodiment not only obtains a user's health status profile at a static level but also constructs a computable description of the user's execution habits at a dynamic behavioral level. This improves the targeting and executability of intervention strategies, reduces plan churn and user fatigue caused by mismatch, and enhances the overall health management effect and platform intelligence level.

[0035] Step S103: Receive the health management goals for the target user, input the personal health profile, behavior patterns and health management goals into the health decision-making model, and generate a personalized behavior intervention plan that includes phased goals and specific task suggestions.

[0036] In this embodiment, the health decision-making model can be an intelligent decision-making engine that comprehensively calculates the target user's personal health profile, behavioral patterns, and health management goals to generate a personalized behavioral intervention plan. The model adopts a structure combining rule constraints and data-driven prediction. On one hand, it introduces expert rules based on medical guidelines and behavioral science theories to impose hard constraints on the type of intervention task, safety boundaries, and maximum execution intensity, ensuring that the generated plan is medically reasonable, safe, and feasible. On the other hand, it introduces a machine learning model that takes health feature data, health management goals, and candidate intervention task combinations as input to predict the changing trends of health indicators and the probability of users completing tasks under different intervention combinations. Thus, under the premise of meeting safety constraints, it optimizes the intervention combination with the goal of maximizing health benefits and increasing compliance.

[0037] During training, a training sample set is constructed using a large amount of historical user data. The actual changes in multidimensional health indicators are used as supervision signals for the intervention effect prediction sub-model, while the completion status and degree of historical tasks are used as supervision signals for the compliance prediction sub-model. The system learns the mapping relationship between health features, task features, and outcomes. The compliance prediction sub-model can be implemented using an algorithm such as gradient boosting trees, deep neural networks, or recurrent neural networks, or a combination of multiple algorithms.

[0038] When making decisions using the health decision-making model, the first step is to break down the dimensions requiring priority improvement and the target range based on the risk levels and health management goals of each dimension in the individual's health profile. Then, a set of tasks matching the target dimensions and satisfying behavioral pattern constraints is selected from the candidate intervention task library. The intervention effect prediction sub-model and the compliance prediction sub-model are then used to score different task combinations. Finally, the solution with the best overall score across the dimensions of expected health improvement and task completion probability is selected as the decision result. Through these steps, the health decision-making model can integrate medical safety, individual health risk, and behavioral feasibility into a unified decision-making framework. Compared to traditional solutions based on fixed templates and empirical rules, this further improves the scientific rigor, individualization, and long-term effectiveness of intervention strategies.

[0039] For example, for an office worker with mild obesity and a significant problem with staying up late, the health decision-making model might prioritize low-threshold tasks in the first stage, such as "light walking after dinner" or "going to bed 30 minutes earlier and reducing screen time before bed," rather than directly recommending high-intensity interval training and a strict low-carb diet, to reduce the risk of abandonment. Through this decision-making mechanism that combines rule constraints with machine learning, the health decision-making model in this embodiment can continuously optimize its recommendation logic under training with massive amounts of historical data, ensuring that the generated intervention plans are both consistent with scientific evidence and highly tailored to individual differences, thereby improving the intelligence and effectiveness of health management.

[0040] Personalized behavioral intervention plans are actionable health behavior arrangements output by a health decision-making model for a single target user. Their core includes time-based phased goals and a set of specific task suggestions configured around each phased goal.

[0041] Phased goals are used to break down a user's long-term health management goals into several relatively short-term, measurable sub-goals. For example, the goal of "losing 10 kg in weight and improving sleep within six months" can be broken down into "establishing a regular sleep schedule and reducing high-calorie snacks in the first month," "stabilizing weight and slightly decreasing it in the next two months and increasing daily steps," and "following up to further optimize body fat percentage and consolidate habits." This allows users to focus on a limited area of ​​improvement at each stage. Specific task suggestions are the smallest behavioral units that support the implementation of phased goals. They are usually presented in the form of "in what scenario, at what time, what to do, and to what extent." For example, tasks for sleep goals could be "turn off the lights and screen before 11 pm on weekdays and not use your phone for one hour before bed" or "do 15 minutes of relaxation stretching or breathing exercises at least three days a week in the evening." Tasks for activity goals could be "get up and move around for 5 minutes every 60 minutes on weekdays" or "do 30 minutes of moderate-intensity walking after dinner at least twice a week." The content, intensity, frequency, and recommended time windows of these tasks are automatically determined by the health decision-making model, combining the risk dimensions displayed in the individual's health profile and the execution preferences reflected in behavioral patterns. For example, for users with low adherence and heavy overtime work on weekdays, it is preferred to break them down into multiple short, dispersed micro-tasks, and schedule their execution time more often in the morning, during lunch breaks, or on weekends. By organizing the intervention plan with phased goals as the main thread and specific task suggestions as the key, this embodiment can transform abstract health goals into daily, perceptible, and actionable action paths for users. Subsequently, the phased goals and task configurations are dynamically adjusted based on task completion and user feedback, thereby forming a continuously iterative and optimized closed loop of behavioral change, effectively improving long-term adherence rates and overall health management effectiveness.

[0042] As an optional embodiment, in step S103, the personal health profile, behavioral patterns, and health management goals are input into the health decision-making model to generate a personalized behavioral intervention plan that includes phased goals and specific task suggestions. This includes: determining the priority phased goal dimensions and corresponding phased goal intervals based on the health management goals and the risk levels of each health dimension in the personal health profile; selecting candidate intervention tasks that match the phased goal dimensions from the candidate intervention task set; matching the task type, execution content, intensity and frequency parameters, and recommended execution time window of the selected candidate intervention tasks with the behavioral patterns to obtain a subset of tasks to be evaluated that meets the target user's safety constraints and execution preferences; calling the intervention effect prediction sub-model and the compliance prediction sub-model to evaluate the health indicator change trends and task completion probabilities of each task combination in the subset of tasks to be evaluated to obtain a target task combination that matches the phased goals and has the best comprehensive score of expected effect and compliance; and generating a personalized behavioral intervention plan divided into phases based on the target task combination. The personalized behavior intervention plan includes at least a specific task list for different stages of goals, a task execution time window, a recommended execution frequency, and corresponding incentive rules, so that the task difficulty and push frequency can be dynamically adjusted according to the behavior pattern when the behavior stage changes.

[0043] Specifically, in the above embodiments, when generating a personalized behavioral intervention plan, the health decision-making model first combines the user's input health management goals with the risk levels of each health dimension in the personal health profile to automatically determine the dimensions that need priority intervention and reasonable phased target intervals. For example, if both overweight and poor sleep quality exist simultaneously, the model can prioritize "improving sleep quality" as the first-stage goal based on risk weights, and break down "weight loss of 5%" into subsequent phased target intervals. Subsequently, candidate tasks matching the target dimensions are selected from a pre-built set of candidate intervention tasks. The task types (such as walking, diet replacement, pre-sleep stress reduction exercises), execution content (such as walking for 20 minutes after dinner, 10 minutes of breathing relaxation before bed), intensity and frequency parameters, and recommended execution time windows are matched with the user's compliance, preferred time periods, and reach sensitivity recorded in the behavioral pattern. Tasks that are obviously difficult to execute in the current behavioral stage or conflict with safety constraints are filtered out, resulting in a subset of tasks to be evaluated that matches the user's characteristics. Next, the intervention effect prediction sub-model and the compliance prediction sub-model are invoked. Different combinations of current health characteristics, phased goal information, and the subset of tasks to be evaluated are used as input to predict the changing trends of health indicators such as weight, sleep score, and mood fluctuations, as well as the expected probability and degree of task completion for each combination. Through weighted scoring or multi-objective optimization algorithms, health benefits and feasibility are integrated into a single decision index, selecting the target task combination that best matches the phased goal and has the optimal overall score. Based on this target task combination, a phased personalized behavioral intervention plan is generated in chronological order. For each phase, a specific task list, task execution time window, and recommended execution frequency are provided. Simultaneously, corresponding incentive rules are configured based on the estimated user incentive needs in the behavioral pattern. For example, for users with average compliance, a virtual badge or points reward can be obtained by completing a certain number of consecutive days. When the behavioral phase transitions from the preparation phase to the maintenance phase, the intervention plan is dynamically upgraded or its workload reduced by adjusting task intensity, increasing or decreasing push frequency, and the number of micro-tasks. Through the above decision-making process, this embodiment can not only ensure that the intervention plan is highly consistent with the health goals and risk status as a whole, but also take into account the user's actual behavior patterns and ability to implement them in detail. This makes the final personalized behavior intervention plan both scientific and reasonable and highly executable, which is conducive to improving the long-term adherence rate and intervention effect, thereby further enhancing the capabilities of intelligent health coaching and behavior guidance.

[0044] In another optional embodiment, step S103, the step of generating a personalized behavioral intervention plan that includes phased goals and specific task suggestions, further includes: calling an expert rule engine to impose rule constraints on the applicable population, execution intensity, maximum frequency, and safety boundaries of candidate intervention tasks based on preset medical guidelines, exercise prescriptions, nutritional advice, and behavioral science theories, and eliminating task combinations that do not meet medical safety and long-term feasibility requirements; calling a data-driven learning model to learn the comprehensive impact relationship of different task types, execution order, and push rhythm on health improvement and compliance based on a set of historical intervention effect indicators, and outputting candidate task combination scores for different user profiles and behavioral patterns; weightedly fusing the compliant task set output by the expert rule engine and the task combination score results output by the data-driven learning model to obtain a comprehensive intervention plan that meets medical safety constraints and has a high comprehensive score in expected health benefits and compliance indicators; and generating the final personalized behavioral intervention plan based on the comprehensive intervention plan, and structurally configuring each phased goal with corresponding task combinations, execution windows, frequency suggestions, and incentive rules.

[0045] In the above embodiments, when generating a personalized behavioral intervention plan that includes phased goals and specific task suggestions, the health decision-making process is achieved through the collaborative work of an expert rule engine and a data-driven learning model, which ensures medical safety while making full use of empirical patterns in historical data.

[0046] Specifically, the expert rule engine has a pre-built set of rules based on authoritative medical guidelines, exercise prescriptions, nutritional advice and behavioral science theories. It imposes hard constraints on the applicable population, execution intensity, maximum frequency and combination relationship between candidate intervention tasks. For example, it prohibits recommending high-impact and ultra-high-intensity interval training for users with underlying cardiovascular diseases, restricts the early addition of multiple high-load task combinations for users with long-term sleep deprivation, and sets safety boundaries for total daily exercise duration, number of high-intensity exercise sessions per week and lower limit of energy intake.

[0047] Optionally, the establishment and configuration of the expert rule engine can be based on domain knowledge engineering. Specifically, experts in medicine, nutrition, and sports rehabilitation, based on existing authoritative guidelines and consensus, should first identify key constraints related to exercise safety, chronic disease management, dietary control, and sleep intervention. These constraints should then be formalized into structured rules, for example, by using population and scenario conditions to jointly determine the permitted and prohibited task ranges, as well as upper and lower limits for parameters. In implementation, a rule-based inference engine can be used to encode various conditions into machine-readable rule entries. For example, a conditional expression could describe that high-intensity interval running is prohibited for individuals over 60 years of age with cardiovascular disease, or that users with severe sleep deprivation should not exceed 30 minutes of total daily exercise time and should not engage in more than two high-intensity tasks in the two weeks prior to intervention. These rules can be hierarchically organized according to task type, execution intensity, frequency limits, time periods, and task combinations, and support the configuration of rule priorities and effective scope. During runtime, after constructing a set of candidate intervention tasks, the system inputs the target user's personal health profile, behavioral patterns, and current scenario information into the rule engine. It then matches each rule condition against the data, eliminating tasks or task combinations that do not meet safety constraints or automatically adjusting task parameters. Tasks that meet the conditions are marked as compliant. As the platform operates for longer periods, rules can be added, deleted, and their weights adjusted based on expert review opinions and statistics of serious adverse events during real-world interventions, gradually forming a maintainable and iteratively updatable rule base. In this process, the expert rule engine provides safety boundaries and hard constraints for the data-driven model, preventing solutions that might violate medical common sense from being derived solely from data learning in extreme cases, thus ensuring the reliability and safety of the intervention strategy in long-term operation.

[0048] After filtering by the rules engine, a data-driven learning model is invoked to evaluate the effectiveness and compliance of the retained set of compliant tasks. This data-driven learning model uses a set of historical intervention effectiveness indicators as training data to learn the comprehensive impact of different task types, execution order, and push rhythm combinations on health indicators such as weight, blood pressure, sleep scores, and mood fluctuation index, as well as compliance indicators such as task completion rate and continuous check-in days. In terms of implementation, gradient boosting trees, deep neural networks, or sequence models can be used to encode user profile features, behavioral pattern features, and candidate task combinations as input, and output a task combination score for a specific user profile and behavioral pattern.

[0049] Optionally, for the establishment and training of the data-driven learning model, this embodiment can adopt a multi-input multi-output model structure oriented towards task combination scoring. Specifically, training samples are first constructed from historical operational data. Each sample includes user profile features, behavioral pattern features, encoded representations of candidate task combinations, and health effect indicators and compliance indicators within the corresponding time window. For example, the multidimensional health score vector, population stratification labels, and behavioral pattern features in the personal health profile can be concatenated into user-side features, and the candidate task combinations in the current stage can be encoded into task-side features. Task-side features can include one-hot encoding of task type, execution intensity, execution frequency, duration, execution time period, and other numerical parameters, as well as combination identifiers used to indicate whether to combine certain sensitive tasks. At the same time, changes in weight, sleep quality, blood pressure, mood fluctuation index, task completion rate, and number of consecutive execution days at the end of the training cycle are used as supervision signals to form the target output. In terms of model structure, gradient boosting tree models can be used to model the aforementioned structured features, or deep learning networks can be constructed using multilayer perceptrons or sequence models. For example, task combinations can be treated as a sequence input, user features can be mapped to a latent vector space and concatenated with task sequence features, and then output as health benefit scores and compliance scores after passing through several fully connected layers or attention layers. During training, the model is supervised using historical sample sets to minimize the error between the model's predictions and the actual results. A multi-task learning strategy can also be employed, using health improvement scores and compliance scores as two output branches, sharing a front-end feature extraction layer, enabling the model to simultaneously learn the trade-off between health benefit goals and compliance goals. After training, in the online decision-making phase, forward reasoning is performed on each candidate task combination filtered by the expert rule engine to obtain a comprehensive score for each combination under the current user profile and behavioral pattern. This score is then weighted and fused with the compliance markers from the expert rule engine to generate a comprehensive intervention plan that is well-suited to different user groups and behavioral types. Through this structural design, the data-driven learning model, while adhering to the safety boundaries set by the rule engine, makes full use of the long-term accumulated intervention effect data to discover task combinations and push rhythm patterns that are more conducive to improving health in real application scenarios and are easier for users to adhere to in the long term, thereby further improving the precision of intervention decisions and overall health benefits.

[0050] Furthermore, during decision-making, on the one hand, the set of compliant tasks output by the expert rule engine is used to ensure medical safety and long-term feasibility. On the other hand, a data-driven model is used to give each task combination a comprehensive score reflecting health benefits and compliance. The results of the two are then weighted and integrated. For example, combinations that violate medical safety rules are directly assigned zero weight, while combinations that show high improvement and low churn in historical data are given higher weight. This results in a comprehensive intervention plan that not only meets the guidelines but also scores highly in terms of expected health benefits and compliance indicators.

[0051] Based on this comprehensive intervention plan, the phased goals are structured and configured with corresponding task combinations, execution time windows, recommendation frequencies, and incentive rules. For example, for the first phase goal of improving sleep, tasks such as turning off lights and screens before 11 PM on weekdays, engaging in 10 minutes of relaxation training before bed, and reducing caffeinated drinks before bed are assigned to a user who stays up late. The execution time window is limited to 10 PM to 11 PM, the frequency is set to 5 days a week, and additional points are awarded for completing 7 consecutive days as an incentive. In subsequent phases, when the behavior pattern identifies that the user has entered the maintenance phase and has high compliance, the task intensity can be automatically increased or more challenging task combinations can be added. Through the above mechanism, this embodiment achieves an organic combination of bottom-line constraints based on medical rules and effect optimization based on historical data at the decision-making level. This balances the generated personalized behavioral intervention plan between safety, scientific rigor, and feasibility, effectively improving the reliability of the intervention strategy, the long-term user adherence rate, and the overall technical effectiveness of health management.

[0052] It should be noted that any step in the embodiments of this application involving the judgment of medical conclusions, setting information, and guidance suggestions must be reviewed and verified by relevant personnel or professional institutions, which will not be elaborated here.

[0053] Step S104: Push tasks from the behavior intervention plan to the user through the interactive terminal bound to the user, and collect task execution status and user feedback.

[0054] As an optional embodiment, in step S104, pushing tasks from the behavior intervention plan to the user through an interactive terminal bound to the user includes: obtaining the target user's contextual environment information at the current moment, including at least time period, geographical location, schedule, device online status, and recent interaction activity; determining the push timing, push frequency, and priority of the corresponding tasks based on the contextual environment information, the preferred time period and reach sensitivity recorded in the behavior pattern, and the recommended execution time window and urgency of each task in the personalized behavior intervention plan, and generating an intervention push strategy for the current moment; selecting a target push channel and push format from at least one push channel among mobile applications, instant messaging tools, smart speakers, and wearable device reminders, and pushing the tasks to be executed to the target user in the form of notifications, dialogues, or contextual reminders, and accepting feedback operations from the target user on the interactive terminal, including accepting, delaying, or rejecting; recording the feedback operations and the actual task initiation status, and synchronously writing the push results and subsequent task execution status into the task execution record for updating the behavior pattern and intervention push strategy.

[0055] Here, in step S104, when pushing tasks from the behavior intervention plan through the user-bound interactive terminal, a context-aware intelligent push mechanism is introduced to improve the timing and user acceptance of intervention. Specifically, the contextual environment information of the target user at the current moment is first obtained from the terminal and business side, including whether the current time period is a weekday morning, evening, or weekend leisure time, whether the geographical location is an office, home, or outdoor location, whether there are meetings or appointments in the schedule, whether the user's frequently used devices are currently online, and the user's recent interaction activity in the application. Combining the preferred time periods and sensitivity to reminder frequency recorded in the pre-established behavior pattern, as well as the recommended execution time window and urgency of each task in the personalized behavior intervention plan, a set of rules or a lightweight scoring model is used to calculate the push timing score, suggested push frequency, and priority for the currently pushable tasks, thereby generating an intervention push strategy for the user at a given time point.

[0056] In practical applications, optionally, a set of rules or lightweight scoring models can employ a weighted scoring function based on heuristic rules. This involves pre-assigning weights to factors such as time period matching, task urgency, current scenario suitability, recent push notification frequency, and user interaction activity. The values ​​of these factors are then linearly or non-linearly combined using a fixed formula to obtain a comprehensive score that measures whether a task is suitable for push notification at the current moment. For example, a higher score is given when the user is within their preferred time period and has not been frequently disturbed recently, while a lower score is given when the user is in a meeting or resting at night. Alternatively, a machine learning model trained based on historical push notification records and completion status can be used. This could include logistic regression models, shallow neural network models, or small-scale gradient boosting tree models. These models take current contextual features, behavioral patterns, and task characteristics as input and output the probability that the task will be accepted or completed by the user at the current moment. This probability can be directly used as the scoring result or weighted and fused with rule-based scoring results, thereby achieving a refined quantitative evaluation of push timing, frequency, and priority while maintaining low computational overhead.

[0057] For example, when a user is in a meeting for several hours on a weekday morning and their phone usage is low, reminders for exercise tasks requiring a longer duration will be postponed. Instead, reminders will be scheduled after get off work or after dinner, periods marked as high-feasibility times based on their behavioral patterns. Conversely, for short relaxation exercises suitable for lunch breaks, reminders will be proactively sent when the system detects that the user is in the company break area and their device is online.

[0058] After generating an intervention push strategy for the current moment, based on the user's device binding and historical preferences, the system selects the most suitable target channel and display format from various push channels such as mobile applications, instant messaging tools, smart speakers, and wearable device reminders. Tasks to be executed are pushed to the user via notification pop-ups, conversational interactions, or contextually matched reminder cards. The user interface allows for feedback actions such as accepting, delaying, or rejecting each task. For example, for a 20-minute walk after dinner, a reminder with the expected end time and step target can be pushed via a mobile application, allowing the user to choose to start immediately, delay by 30 minutes, or skip it for the day. Furthermore, the system records user feedback and actual task initiation in real time, and monitors completion time and progress during task execution, forming a complete task execution record. These records, along with the push timing, channel, and reminder format from the corresponding push strategy, are written into the task fulfillment data for subsequent iterations to update behavioral patterns and push strategy parameters. When it's detected that a user consistently delays or rejects morning exercise tasks but has a high completion rate for evening bedtime relaxation tasks, the system automatically adjusts the user's preferred time slots and reach sensitivity parameters. This reduces disruptions during inappropriate times and increases push frequency within high-response windows. Through this closed-loop mechanism, this embodiment can deliver the same intervention plan to different users and in different scenarios in a more appropriate way. Push notifications are no longer mechanical reminders at fixed times, but rather flexible guidance tailored to the user's rhythm and context, thereby increasing the probability of task acceptance and completion, avoiding push fatigue, and improving the overall effectiveness of behavioral intervention and user experience.

[0059] In another embodiment, step S104, pushing tasks from the behavior intervention plan to the user via an interactive terminal bound to the user, further includes: acquiring text interaction content, voice signals, and physiological signals related to the target user; performing single-modal sentiment analysis on the text interaction content, voice signals, and physiological signals to obtain a first identification result related to emotion type and emotion intensity, and a second identification result related to stress level and motivation level; performing multimodal fusion processing combining feature-level fusion and decision-level fusion based on the first and second identification results, and extracting the target user's comprehensive emotional state and comprehensive motivation level from the fusion results; matching corresponding communication postures and behavior strategies in a pre-constructed communication posture and behavior strategy matrix according to the comprehensive emotional state and comprehensive motivation level, generating empathic feedback content to guide the target user to execute the personalized behavior intervention plan by combining the target user's personal health profile and behavior patterns, and pushing the empathic feedback content as part of the personalized behavior intervention plan to the target user via the interactive terminal bound to the target user.

[0060] Optionally, the communication posture and behavior strategy matrix is ​​a two-dimensional matrix, with rows representing user emotional states and columns representing user motivation levels. User emotional states include at least one of the following: frustration, fatigue, excitement, anxiety, and calmness. User motivation levels are divided into multiple grades according to a preset scoring standard. Each cell in the matrix corresponds one-to-one with a combination of an emotional state and a motivation level, used to store the communication posture and behavior strategies for that combination. The behavior strategies include at least one or more of the following: reducing the difficulty of the daily goal, breaking down phased goals into multiple sub-goals, adjusting task types and execution times, showcasing short-term intervention results, and setting motivating challenges. During operation, the communication posture and behavior strategy matrix is ​​periodically updated based on the user's acceptance rate of empathic feedback and the execution rate of the corresponding intervention tasks.

[0061] In a preferred embodiment, the personalized health intervention prediction platform also includes a multimodal emotion computing and empathic feedback engine for identifying users' emotional and motivational states and generating empathic feedback. This engine accurately identifies the target user's overall emotional state and motivational level by fusing multimodal data such as text, voice, and physiological signals. Based on a pre-built communication posture and behavior strategy matrix, it determines communication methods and behavioral strategies that match the current state, thus breaking through the traditional approach of relying solely on a single modality for emotion recognition and outputting fixed scripts. This achieves intelligent coaching interaction that is more tailored to the individual's psychological state. In its implementation, the multimodal emotion computing and empathy feedback engine includes multiple unimodal recognition sub-models and a multimodal fusion module. The text emotion recognition sub-model uses a pre-trained language model combined with a convolutional neural network to encode the dialogue text between the target user and the platform, outputting the corresponding emotion type and intensity score. The speech emotion recognition sub-model uses a convolutional neural network combined with a recurrent neural network, taking the acoustic feature sequence of the user's speech as input, analyzing speech rate, pitch, and energy changes, and outputting the corresponding emotion type and intensity. The physiological signal recognition sub-model uses time-series models such as gated recurrent units to model physiological signals such as continuous heart rate and skin conductance, outputting the user's stress level score and motivation level score, thus physiologically characterizing whether the user is in a state of high stress, low motivation, or high motivation. Each of these unimodal recognition sub-models can be trained under supervised supervision on a multimodal training set containing labeled text, speech, and physiological signal data to improve recognition accuracy and generalization ability in real-world interaction scenarios.

[0062] After completing single-modal recognition, the multimodal emotion computing and empathy feedback engine obtains the target user's comprehensive emotional state and comprehensive motivation level through a combination of feature-level fusion and decision-level fusion. Specifically, in the feature-level fusion stage, the text feature vectors output by the text sub-model, the speech feature vectors output by the speech sub-model, and the physiological feature vectors output by the physiological sub-model are concatenated or mapped to a unified feature space to obtain a fused feature representation containing multimodal information. In the decision-level fusion stage, weights are assigned to the recognition results of each modality based on their recognition reliability in historical evaluations. The emotion type and motivation level output by each modality are then integrated through an attention mechanism or weighted summation method to obtain the final comprehensive emotional state and comprehensive motivation level. For example, when text emotion recognition is highly accurate in the current scenario, a higher weight can be assigned to the text modality, while in noisy environments or when the speech signal quality is low, the weight of the speech modality is reduced to ensure the stability and robustness of the comprehensive recognition result. Through the above multimodal fusion strategy, the platform can still make more reliable judgments on emotional and motivational states by combining physiological signal information, even when the user only inputs a small amount of text or a short speech.

[0063] To transform emotion recognition results into actionable behavioral intervention plans, this embodiment further designs a communication posture behavior strategy matrix to establish a mapping relationship between emotional states and motivation levels, and between communication styles and behavioral strategies. The matrix uses user emotional states as the row dimension and user motivation levels as the column dimension, categorizing user emotions into several types such as frustration, fatigue, excitement, anxiety, and calmness. User motivation levels are divided into multiple levels according to preset scoring criteria, ensuring that each cell in the matrix corresponds one-to-one with a combination of an emotional state and a motivation level. Each combination is associated with a set of communication posture descriptions and behavioral strategy configurations. For example, when a user is identified as frustrated and with low motivation, the corresponding cell in the matrix can be set to a communication posture focused on reassurance and stress reduction, with behavioral strategies such as reducing the difficulty of the day's goals or assigning lighter alternative tasks. When a user is identified as anxious and with moderate motivation, the corresponding cell can adopt a communication posture focused on empathy and explanation, coupled with behavioral strategies such as displaying recent minor improvements or slightly adjusting task times to alleviate the user's concerns about the effectiveness. When a user is identified as being in a positive state with high motivation, the matrix can select a communication stance that combines incentives and prospects, and configure challenging tasks and phased reward mechanisms to guide the user to appropriately increase the intensity of intervention within their tolerance. As the platform's operating time increases, the matrix can be periodically optimized based on the user acceptance rate and task execution rate of the corresponding strategies under different matrix cells. For example, strategy combinations with consistently low acceptance and execution rates can be adjusted or replaced, thereby ensuring that the communication stance behavior strategy matrix maintains high effectiveness in practical applications.

[0064] In the empathic feedback generation phase, the multimodal affective computing and empathic feedback engine takes comprehensive emotional state, comprehensive motivation level, and personal health profile and behavioral patterns as input. First, it matches the communication posture and behavioral strategy configuration of the corresponding cell in the communication posture and behavior strategy matrix. Then, combining the target user's health risk dimension, current stage goals, and historical performance, it generates empathic feedback content in natural language form through a personalized dialogue generation engine. With the support of knowledge graphs and health management rule bases, the personalized dialogue generation engine can transform abstract communication postures and behavioral strategies into concrete, easily understood, and actionable text or voice prompts. For example, when it identifies that a weight-loss user is emotionally frustrated and has low motivation due to not completing their daily exercise task, the platform can generate feedback that expresses understanding and reassurance while appropriately reducing the difficulty of the daily exercise goal, guiding the user to complete an easier alternative task first, thus preventing the user from completely abandoning their long-term plan due to frustration. For example, the communication posture and behavior strategy matrix can be shown in the table below:

[0065] Through this empathic feedback mechanism based on multimodal emotion recognition and communication posture behavior strategy matrix, this embodiment can enhance affinity and acceptability at the interaction level while maintaining the rationality of health intervention goals, effectively improve user compliance with intervention plans and long-term adherence, thereby further enhancing the overall health management effect.

[0066] Furthermore, in an optional embodiment of step S104, the collection of task execution status and user feedback includes: recording the planned start time, actual start time, completion time, completion level, and interruption reason for each intervention task by the target user, and generating an execution record associated with the task identifier; receiving subjective experience feedback information submitted by the target user through at least one of the following methods: scale scoring, preset options, text input, or voice input, wherein the subjective experience feedback information includes at least an evaluation of task difficulty, execution feeling, and intervention satisfaction; associating and storing the subjective experience feedback information with the corresponding task execution record to form a task performance dataset for representing the objective execution status and subjective feelings of the intervention process; and quantitatively evaluating the comfort and acceptability of different types of intervention tasks based on the task performance dataset.

[0067] In this embodiment of the application, the process of collecting task execution status and user feedback aims to build a task performance data foundation that can simultaneously reflect the task execution status, emotional state, or behavioral state during the intervention process, for subsequent quantitative evaluation of the comfort and acceptability of different intervention tasks.

[0068] Specifically, after each intervention task is issued to the target user, the system automatically records the planned start time, actual start time, completion time, completion percentage, and reason for interruption at key points in the task lifecycle. This information is then linked to the task's unique identifier to form an objective execution record. For example, for the task "Walk for 20 minutes after dinner," the planned execution time is recorded as 8:00 to 8:30 PM. If the user clicks "Start" at 8:15 PM and finishes at 8:35 PM, the corresponding actual start and finish times are recorded accurately. If the user stops midway or only completes 10 minutes, the completion percentage and reason for interruption are noted in the execution record, such as due to overtime work, weather conditions, or physical discomfort. Building upon this foundation, users are guided to provide subjective feedback upon task completion through mobile application questionnaires, preset option ratings, text comments, or voice input. Feedback should include at least their perceived task difficulty, comfort or fatigue during execution, and overall satisfaction with the intervention. For example, users can choose options such as "task difficulty was high, I felt slightly tired after execution" or "the pace was suitable, I felt good after execution," or rate their experience on a scale of 0 to 10. This subjective feedback is then linked to the corresponding objective task execution records at the data level and synchronously stored in the task performance dataset. This ensures that each record includes both objective facts such as time and completion rate, as well as the user's emotional and physical evaluation of the intervention.

[0069] With a sufficiently large dataset of task performance, the comfort and acceptability of different types of intervention tasks can be quantitatively evaluated through statistical analysis and simple modeling. Specifically, records of multiple executions of the same or similar tasks can be aggregated according to dimensions such as task type, execution intensity, time period, and user characteristics. Indicators such as average completion rate, early termination rate, average difficulty score, and satisfaction score can be statistically analyzed to calculate a comprehensive comfort score and acceptability score for each task category. For example, it might be found that "moderate-intensity walking on weekday evenings" has a high completion rate and high satisfaction score among most office workers, while "high-intensity interval training on weekday mornings" has a significantly higher interruption rate and difficulty score among the same group. Therefore, the former can be judged to have better comfort and acceptability, while the latter requires more caution or a reduction in intensity when recommending it. At the algorithmic level, a weighted average or regression model can be used to comprehensively map completion rate, subjective difficulty, and satisfaction to a task acceptability index ranging from zero to one. This index can then be used as a constraint or important feature in health decision-making models and push strategy optimization, causing subsequent intervention plans to tend to select task types and configurations with high comfort and acceptability. Through the aforementioned closed-loop data collection and quantitative evaluation mechanism, this application embodiment can not only accurately record the user's actual performance in each intervention, but also transform subjective experience into calculable indicators. In the long-term operation, it can continuously eliminate task combinations with poor experience and strengthen intervention paths with good experience, thereby systematically improving the humanization of intervention strategies and the overall health management effect.

[0070] Step S105: Evaluate the intervention effect based on the implementation status and user feedback, update the individual health profile and behavior pattern, and adjust the behavior intervention plan according to the updated individual health profile and behavior pattern.

[0071] As an optional embodiment, in step S105, the intervention effect is evaluated based on the implementation status and user feedback, and the personal health profile and behavior pattern are updated, including: Within a pre-defined evaluation period, the changes in multidimensional health score vectors before and after intervention, task completion rate, phased goal achievement rate, user retention rate, and comfort score output by the experience evaluation model are statistically analyzed to form raw statistical data on the intervention effect for the evaluation period. Based on the raw statistical data, an intervention effect indicator set is calculated, which includes at least health improvement indicators, behavioral compliance indicators, and user experience indicators. The parameters of the health decision-making model, intervention effect prediction sub-model, and compliance prediction sub-model are iteratively updated using the intervention effect indicator set. Based on the updated model output results, the multidimensional health score and risk stratification label in the personal health profile are adjusted, and the behavioral stages, behavioral types, and execution preference characteristics in the behavioral pattern are updated. Based on the updated personal health profile and behavioral pattern, the priority of health management goals is reassessed, and the task difficulty, task type combination, and push frequency in the personalized behavioral intervention plan for subsequent periods are adaptively adjusted.

[0072] In this embodiment, step S105 achieves closed-loop optimization of the health intervention effect by introducing a periodic evaluation and model adaptive update mechanism. Specifically, within a preset evaluation period, based on the aforementioned task execution status and user feedback data, the changes in the target user's health status and behavioral performance during this period are comprehensively statistically analyzed. First, the multidimensional health score vectors before and after the intervention are compared, and the changes in various indicators such as weight, blood pressure, sleep quality score, and mood fluctuation index are calculated. At the same time, the overall completion rate of various tasks, the achievement rate of each stage goal, the retention rate of users continuing to use the platform, and the task comfort and intervention satisfaction scores output by the experience evaluation model are statistically analyzed during this period, thereby forming the original statistical data of the intervention effect for this evaluation period. On this basis, the above-mentioned original statistical data are summarized and quantified to construct a set of intervention effect indicators including health improvement indicators, behavioral compliance indicators, and user experience indicators. For example, health improvement indicators may include the percentage of weight loss and the average improvement score of sleep quality; behavioral compliance indicators may include the task completion rate and the number of consecutive check-in days; and user experience indicators may include the average comfort score and satisfaction score. Using this set of intervention effect indicators, the parameters of the health decision-making model, the intervention effect prediction sub-model, and the compliance prediction sub-model can be iteratively updated. Incremental training or periodic retraining can be used to gradually correct the prediction biases of different task combinations and compliance, making the model more closely reflect the real intervention results.

[0073] After the model is updated, the health status and behavioral characteristics of the target users are recalculated based on the latest model output. The multidimensional health scores and risk stratification labels in the personal health profile are adjusted. For example, when sleep quality scores improve further and sleep-related risks decrease continuously over multiple assessment cycles, the risk level of the sleep dimension can be lowered, gradually shifting the intervention focus to weight control or emotional management. Simultaneously, the behavioral stages, behavioral types, and execution preference characteristics in the behavioral patterns are updated based on the latest task completion rates and experience scores. For instance, when a user maintains a high completion rate and provides positive feedback on task difficulty over several consecutive cycles, their behavioral stage can be adjusted from the preparation stage to the maintenance stage, and the upper limit of acceptable task difficulty can be appropriately increased. Finally, based on the updated personal health profile and behavioral patterns, the priority of overall health management goals is reassessed, and the personalized behavioral intervention plan for the next assessment cycle is adaptively adjusted accordingly. This includes moderately increasing task difficulty while ensuring medical safety boundaries, adjusting the proportion of task type combinations, and optimizing push frequency and timing. For example, for users who have adapted to the current exercise intensity and report "too low difficulty," the walking time per session can be increased or a light strength training task can be introduced in the next cycle. For users who report "too many push notifications," the number of daily reminders will be appropriately reduced and concentrated during their preferred time periods. Through the closed-loop process of evaluation, updating, and replanning described above, the embodiments of this application enable the health decision-making model and intervention plan to evolve dynamically over time, continuously absorbing information on the actual intervention effects, and responding promptly to individual differences and behavioral changes, thereby improving the accuracy, sustainability, and overall health management effectiveness of the intervention strategy.

[0074] As an optional embodiment, after step S105, based on the health characteristic data, personal health profiles, intervention execution data, and intervention effect indicators of multiple target users, cluster analysis is performed on the multiple target users to form multiple healthy population subcategories. Each healthy population subcategory includes at least combinations of different risk levels, different behavioral patterns, and different target preferences. For each healthy population subcategory, a corresponding group-level intervention strategy template is generated by comprehensively considering historical intervention effect indicators and expert experience. The group-level intervention strategy template includes at least priority intervention dimensions, recommended task type combinations, and basic push rhythm. When generating personalized behavioral intervention plans for target users, the personal health profiles and behavioral patterns of target users are matched with the feature centers of each healthy population subcategory to determine the healthy population subcategory to which the target user belongs. A basic intervention plan is selected from the group-level intervention strategy templates corresponding to the healthy population subcategory, and the task intensity, execution frequency, and push timing are fine-tuned based on the target user's multidimensional health score, risk stratification label, and behavioral pattern characteristics to obtain the final intervention plan for the target user.

[0075] After evaluating the intervention effect and updating the profile of individual users, a health population stratification mechanism based on cluster analysis can be introduced at the group level to form a reusable group-level intervention strategy template. Specifically, within a preset time window, health characteristic data, personal health profiles, intervention execution data, and intervention effect indicators of multiple target users are aggregated. Each user is represented as a high-dimensional feature vector containing multidimensional health scores, risk stratification labels, behavioral pattern characteristics, task completion rates, experience scores, and goal preferences. Clustering algorithms (such as distance-based clustering methods or density-based clustering methods) are used to group users, resulting in several healthy population subclasses. Each healthy population subclass has similar characteristics in risk level, behavioral patterns, and goal preferences. For example, it can form combination types such as "middle-aged people who are sedentary, have poor sleep quality but a strong desire to lose weight" and "young people who experience high stress, significant mood swings, but have a good exercise foundation."

[0076] For each subcategory of healthy individuals, based on the clustering results and by integrating historical intervention effectiveness indicators and expert experience, we analyze the average health improvement, task completion rate, and satisfaction level of this subgroup under different intervention strategies. We then extract the intervention configurations with better results and construct corresponding group-level intervention strategy templates. The templates should include at least the health dimensions to be prioritized, the recommended combination of task types, and the basic push rhythm suitable for this subgroup. For example, for "sedentary individuals with sleep problems," we suggest the strategy of "first optimizing sleep and light activity, and then gradually increasing the intensity of exercise," while for "high-stress individuals with mood swings," we suggest the strategy of "increasing mood regulation and relaxation exercises and controlling caffeine intake during overtime work."

[0077] Furthermore, in conjunction with the intelligent health coach scenario of this application, the clustering process no longer simply uses numerical features such as body mass index and step count for Euclidean distance clustering. Instead, it incorporates dimensions such as health risk level, behavioral stage, and goal preference into the similarity function. For example, when comparing two middle-aged users, if user A is obese with hypertension and has a low sleep score, they are marked as high-risk, in the preparation stage, and their main goals are to control blood pressure and improve sleep. User B, on the other hand, is only slightly overweight, has no history of chronic diseases, is in the maintenance stage, and their main goals are to shape their body and improve athletic performance. Even if their average daily step counts are similar, the differences in health risk and behavioral stage will be given greater weight in the distance calculation, increasing the distance between them and thus placing them in different subcategories. This avoids mixing "high-risk, low-compliance" users with "low-risk, high-compliance" users in the same category, preventing the inability to uniformly implement group-level strategies. For example, consider a group of young internet professionals. Some users might have high mood swing indices and prefer to set their goals as anxiety relief, exhibiting behaviors such as nighttime online activity and poor early morning adherence. Others might have stable mood scores, primarily aiming to improve physical fitness and habitually running in the morning. In traditional clustering methods that only consider exercise volume and age, these users might be misclassified into the same group. The similarity function in this application imposes penalties on the differences between "emotional risk" and "goal preference," clearly separating these two types of users during clustering. This facilitates designing group strategies focused on mood regulation and sleep optimization for the former, and configuring strategies focused on morning exercise and progressive intensity for the latter. During the clustering algorithm's operation, constraints are also set on cluster size and intra-cluster variance. For example, the number of users in each healthy subclass is limited to a certain threshold, and the variance of intra-cluster health scores and behavioral indicators does not exceed a preset range. If a clustering operation produces a "special case cluster" containing only a very small number of users with highly discrete features, it will be handled by merging, re-initializing the center, or delaying its inclusion in the strategy template, avoiding modeling individual extreme individuals as separate groups that are difficult to generalize. As the platform continues to operate, new intervention cycle data will be continuously written into the unified data base. Incremental updates can be used to fine-tune the cluster centers, or clustering can be re-executed at fixed time intervals. The health improvement effects and behavioral changes of the most recent cycles can be incorporated into the stratification criteria. For example, when a subgroup of people who originally had "severe sleep problems and average compliance" improves their average sleep score and completion rate in multiple cycles, the cluster center of that subgroup will move towards "medium risk, behavior stage tending to maintain". Ultimately, it may be split or renamed into a new group type.In the results interpretation and strategy design stages, the multidimensional features of each cluster center are analyzed in a rule-based manner and solidified into a three-dimensional label of risk level, behavioral pattern, and goal preference. For example, "middle-aged people who are sedentary, have a moderate risk of hypertension, and tend to choose low-impact tasks" and "young people with high blood pressure, high mood fluctuations, and a preference for short-term relaxation and light exercise" are used. Experts can quickly develop and validate group-level intervention strategy templates for each group based on these labels, so that the clustering results can directly serve the actual strategy design and evaluation. Thus, in this application, the clustering algorithm is not only a mathematical grouping method, but also an interpretable and reusable strategy hierarchical tool for health intervention, which effectively improves the pertinence and implementation effect of group strategy design.

[0078] When generating personalized behavioral intervention plans for new or existing users, the similarity between the user's personal health profile and behavioral patterns and the feature centers of various healthy population subcategories can be calculated. The user is then assigned to the most similar population subcategory, and a basic intervention plan is selected from the corresponding group-level intervention strategy template. By combining the user's current multidimensional health score, the latest risk stratification label, and behavioral pattern characteristics, fine-tuning of parameters such as task intensity, execution frequency, and push timing can be made. For example, for users with good compliance within the same population, the exercise intensity and frequency can be appropriately increased, while for users with low compliance, easier tasks and lower push frequencies can be selected. This achieves precise adaptation at the individual level while retaining mature experience paths, improving the efficiency of plan design and ensuring that the intervention strategy is well adapted to different population types and individual characteristics, thereby enhancing the overall scalability of the platform and the effectiveness of health management.

[0079] Optionally, after step S105, an intervention effect prediction sub-model can be used to predict the trend of health indicator changes for the target user under different combinations of intervention tasks, based on health characteristic data and health management goals. A compliance prediction sub-model can be used to predict the probability of the target user completing the candidate intervention tasks. Based on the trend of health indicator changes and the probability of completion, the candidate intervention tasks are combined and screened to generate phased goals and specific task suggestions in a personalized behavioral intervention plan. The candidate intervention task set is constructed based on a pre-set task library, user health profile, health management goals, and safety constraints. Each candidate intervention task includes at least the task type, execution content, intensity and frequency parameters, recommended execution time window, applicable population, and safety constraint information.

[0080] In other words, the embodiments of this application can also utilize the intervention effect prediction sub-model and the compliance prediction sub-model to conduct a refined evaluation of the candidate task combination, assisting in the generation of better phased goals and specific task suggestions. In this process, firstly, a set of candidate intervention tasks is constructed based on a pre-set task library, user health profile, current health management goals, and expert rule constraints. The pre-set task library predefines a variety of intervention task templates for different health dimensions. Each template includes at least the task type (such as aerobic exercise, strength training, dietary adjustment, sleep habit improvement, mood regulation, etc.), execution content (such as walking for 20 minutes after dinner, 10 minutes of breathing relaxation exercises before bed, reducing sugary drinks at lunch), intensity and frequency parameters, recommended execution time window, and applicable population and safety constraint information. The rule engine eliminates unsuitable tasks based on the user's age, chronic disease history, current risk status, etc., forming a safe and compliant candidate task set.

[0081] Next, the current health characteristic data, health management goals, and different arrangements of candidate task combinations are input into the intervention effect prediction sub-model. This sub-model is trained on historical intervention samples to learn the impact patterns of different task combinations on health indicators such as weight, sleep scores, and mood indices, and outputs the predicted trend of health indicator changes for each combination within a given period. Simultaneously, the same combinations, along with user profiles and behavioral pattern characteristics, are input into the compliance prediction sub-model. This sub-model learns the actual completion probability and degree of different users under different task combinations and push notification rhythms, and outputs a compliance score for each candidate combination. Taking into account both the predicted results and completion probabilities of health indicator improvements, candidate task combinations are scored and screened. A group or several tasks with the expected improvement and high compliance are selected as the target task combination for the current stage. This process is then used to formulate corresponding stage goals and specific task suggestions. For example, if the prediction results indicate that moderately increasing daily steps and optimizing pre-sleep behavior can further improve sleep scores with a high completion probability, then "increasing daily steps to 8,000 steps" and "moving bedtime before 11 PM" will be set as stage goals. Specific task suggestions include "walking one stop to and from work every day," "walking easily for 15 to 20 minutes after dinner," and "not using a mobile phone for one hour before bedtime." Through a task combination evaluation mechanism centered on the prediction model, this embodiment can quickly converge to a solution that balances effectiveness and feasibility from a large set of candidate tasks, avoiding reliance solely on human experience or simple rules to combine tasks, thus improving the achievement rate and health benefits of personalized behavioral intervention plans in actual implementation.

[0082] Further optionally, in the above embodiments, before using the intervention effect prediction sub-model to predict the trend of health indicator changes of the target user under different combinations of intervention tasks based on health feature data and health management goals, the method further includes: generating a training sample set based on historical multi-source health data, historical health management goals, historical combinations of intervention tasks, and changes in health indicators before and after intervention. The training sample set includes at least supervised samples that take health feature data, health management goals, and combinations of candidate intervention tasks as inputs and the actual changes of the target user in multiple health indicators as outputs. The intervention effect prediction sub-model is trained based on the training sample set so that it learns the influence relationship between different health features and different combinations of intervention tasks on multidimensional health indicators. The intervention effect prediction sub-model is a machine learning model among gradient boosting tree model, deep neural network model, and recurrent neural network model.

[0083] The training objective of the intervention effect prediction sub-model is to enable it to perform intervention tasks based on the user's current health characteristics and the planned intervention tasks. To this end, a training sample set is first constructed from multi-source health data accumulated during historical operation. Specifically, this includes a user's health characteristic data at the beginning of an assessment cycle and the set health management goals for that period, such as weight loss, improved sleep, or anxiety relief. Simultaneously, the actual combination of intervention tasks performed during that cycle is collected, such as walking for 20 minutes after dinner daily, strength training three times a week, and reducing electronic device use before bedtime. At the end of the cycle, the actual changes in multidimensional health indicators before and after the intervention are calculated, and this change is used as a supervision signal. Thus, each training sample contains health characteristic data, health management goals, and task combinations as model input, and the actual changes in multiple health indicators as model output. Subsequently, a gradient boosting tree model, a deep neural network model, or a recurrent neural network model is used to supervise the learning of the training sample set, enabling the model to gradually grasp the influence relationships between different initial health states and different task combinations on multidimensional health indicators. For example, through training with a large number of samples, the model can learn that for users who are sedentary and have poor sleep, the combination of increasing moderate-intensity evening walking and going to bed earlier often significantly improves sleep quality scores, while simply increasing high-intensity daytime exercise has limited effect on sleep improvement. The trained intervention effect prediction sub-model can receive the user's current health characteristics, health management goals, and the combination of tasks to be evaluated during the online phase. It outputs predictions of short-term trends in indicators such as weight and sleep quality scores. Based on this, the decision-making module selects task combinations with better expected health benefits. Thus, when planning intervention programs, it not only relies on empirical rules but also introduces data-driven effect prediction, improving the predictability and accuracy of the intervention strategy's health improvement effect.

[0084] Optionally, in the above embodiments, before predicting the probability of a target user completing a candidate intervention task using the compliance prediction sub-model, the method further includes: generating a compliance training sample set based on historical task execution records, user context information, and user profiles. The compliance training sample set includes at least supervised samples that take health feature data, the task type, intensity, and frequency parameters of the candidate intervention task, the recommended execution time window, and user behavior pattern features as input, and the task completion status and degree as output. The compliance prediction sub-model is then trained based on the compliance training sample set, enabling it to learn the influence relationship between different user characteristics, task characteristics, and contextual environment on the task completion probability. The compliance prediction sub-model is a machine learning model selected from logistic regression, gradient boosting tree, or deep neural network models.

[0085] Regarding the compliance prediction sub-model, this application embodiment models historical task execution records and user context information to predict the probability of a user completing a task under a given task and scenario, thereby helping the decision-making module avoid task assignments that seem effective but are difficult to maintain in the long term. Before training, a compliance training sample set is constructed based on historical data. Each sample includes at least health feature data, the task type, intensity and frequency parameters of the candidate intervention task, the recommended execution time window, and user behavior pattern features and contextual environment features, such as whether the user was at work, in a frequently used location, or their activity level in the past week when the task was pushed. The supervision signal corresponding to these input features is whether the task was completed and the degree of completion, such as fully completed, partially completed, or not started. By collecting a large number of user execution records under different task configurations and scenarios, a mapping sample between task features, user features, context features, and completion results is formed. The compliance prediction sub-model can be trained using machine learning methods such as logistic regression models, gradient boosting tree models, or deep neural network models to learn the influence relationship between different combinations of user features, task features, and context on the probability of task completion. For example, the model gradually learns that for some office workers, the completion probability of a long-duration, high-intensity exercise task pushed on weekday evenings is significantly lower than that of a moderate-intensity task pushed on weekend mornings. Conversely, for users with good early rising habits, the completion rate of a light exercise task in the morning is higher. The trained compliance prediction sub-model receives the current user's health characteristic data, parameter descriptions of candidate tasks, and stage and preference features of behavioral patterns during the online decision-making phase. It predicts the completion probability and expected completion level of each candidate task or task combination. The decision-making module then considers this compliance score along with the health benefit prediction output by the intervention effect prediction sub-model, selecting task combinations that have both good health improvement potential and high feasibility. This avoids recommending overly idealistic but practically difficult-to-implement plans, thereby further improving the long-term effectiveness and user stickiness of personalized behavioral intervention programs.

[0086] In another optional embodiment, step S101, which stores multi-source health data on a unified data platform, further includes: performing desensitization processing on user identification information and sensitive fields from different data sources; encrypting the data stored on the unified data platform using a hierarchical encryption strategy according to the data sensitivity level; configuring access control policies for service modules that access health feature data, allowing access to health feature data only after authorization verification via encrypted calls; and performing calculations based on authorized encrypted data during the training and inference process of the health decision model, so as to complete the fusion analysis of multi-source health data and personalized health intervention prediction while ensuring user privacy and security.

[0087] Specifically, during the data access phase, user identification information and health-sensitive fields from different data sources are anonymized, replacing direct identification information with anonymous identifiers, and adding masquerading or generalization processing to highly sensitive fields. Based on data category and sensitivity level, data stored in a unified data platform is encrypted using a hierarchical encryption strategy. For highly sensitive data containing patient information, health diagnosis information, or medication information, stronger encryption algorithms and key management strategies are employed. Fine-grained access control policies are configured for internal service modules accessing health feature data and raw health data, limiting the scope and methods of access for different modules. Access to health feature data is only permitted through encrypted calls or controlled decryption after identity authentication and permission verification. In the above embodiments, when training and inferring the health decision-making model, intervention effect prediction sub-model, and compliance prediction sub-model, calculations are performed based on authorized encrypted or anonymized data. Access auditing and operation log recording mechanisms ensure that multi-source health data fusion analysis and personalized health intervention prediction are completed while guaranteeing user privacy, security, and compliance.

[0088] Compared to existing solutions that only perform anonymization and encryption at the storage and simple access levels, this embodiment makes multi-dimensional collaborative improvements to the privacy protection mechanism of a unified data foundation for personalized health intervention scenarios. This embodiment integrates health characteristic data, personal health profiles, behavioral patterns, and intervention execution records into a unified sensitivity grading system, performing fine-grained classification and management for different analytical dimensions and uses. For example, direct identification and diagnostic information are classified as the highest sensitivity level, profile tags and behavioral pattern features as medium sensitivity level, and aggregated group statistical features as low sensitivity level. Based on this, different encryption algorithm strengths and key management strategies are configured to ensure that the high-dimensional features required for model training remain usable while minimizing the exposure of direct identity.

[0089] On the other hand, the access control strategy in this embodiment is deeply integrated with specific algorithm modules and business processes. It not only refines the coarse-grained access control between the server and the access control table, but also specifies which feature fields can be accessed and in what form of desensitization or encryption during the training and inference phases of the health decision model, intervention effect prediction sub-model and compliance prediction sub-model. For example, it restricts plaintext features related to behavior patterns to be used only in a controlled and explicit manner within the model, and only exposes abstracted or aggregated risk labels and behavior types to the outside world, thereby reducing the diffusion path of sensitive raw data.

[0090] Furthermore, this embodiment links authorization verification, encryption / decryption operations, and access auditing during model training and online inference, achieving end-to-end recording. Abnormal access behavior is incorporated into behavior pattern monitoring and security policy updates, enabling data security policies to dynamically evolve with changes in business operations and attack surfaces. Through the aforementioned multi-level sensitivity labeling, access control coupled with algorithm modules, and end-to-end auditing of training and inference processes, this application effectively reduces the risk of individual privacy leaks while ensuring the high-value utilization of multi-source health data. This allows the unified data foundation to support complex individualized intervention and prediction models while meeting increasingly stringent privacy compliance requirements in healthcare scenarios, thus achieving a better balance between security and usability.

[0091] In an optional embodiment, after evaluating the effect of a single intervention in step S105, a behavioral pattern analysis and churn warning intervention mechanism are introduced on a long-term time scale to achieve a shift from passively responding to user churn to proactively preventing user churn. Specifically, long-term behavioral pattern features are constructed based on multi-source behavioral data of the target user over a recent period. The behavioral pattern feature sequence, reflecting changes in execution frequency, degree of behavioral fluctuation, and cross-behavioral correlations, is input into a time series model to obtain a quantitative score for characterizing behavioral stability. Then, combined with static features in the user profile, a survival analysis model is used to estimate the churn probability within several future time windows. When the prediction results show that the target user has a high risk of churn within the next 7 or 14 days, an appropriate intervention level and form will be automatically selected according to the risk level to generate a corresponding personalized churn warning intervention plan. This plan will be presented through interactive terminals in the form of dialogue, task recommendation, or care reminders, thereby proactively recalling users and reactivating their healthy behaviors before actual churn occurs.

[0092] In another specific embodiment, a hybrid churn risk prediction model is constructed based on the aforementioned behavioral pattern feature sequences. This hybrid model consists of a cascaded behavioral stability prediction sub-model and a churn risk prediction sub-model. The behavioral stability prediction sub-model employs a Long Short-Term Memory (LSTM) network structure, taking the behavioral pattern feature sequences from the past few weeks (e.g., the past 12 weeks) as input in weekly increments. By storing long-term dependencies and short-term fluctuation information, it outputs a behavioral stability score, which can range from 1 to 10. A lower score indicates more unstable user behavior and a higher likelihood of churn. The churn risk prediction sub-model uses a Cox proportional hazards model, incorporating the behavioral stability score along with the user's static characteristics as covariates to estimate the churn probability of the target user in the next 7 and 14 days. During the model training phase, training samples are constructed based on large-scale historical user data. Users who have not used the application or completed any intervention tasks for 14 consecutive days are marked as having experienced churn events. Five-fold cross-validation is used to evaluate the model performance.

[0093] Optionally, based on the churn risk prediction results, target users are divided into different churn risk levels, and corresponding personalized intervention strategies are configured for each level. When a user's churn probability within the next 7 days is lower than a preset threshold, personalized health knowledge or lightweight check-in reminders related to their health management goals and interests are pushed only at appropriate times to maintain their basic activity. When a user's churn probability is in the medium-risk range, a brief caring conversation is initiated to inquire about the reasons for recent missed check-ins or uncompleted tasks, and a small regression task is generated based on the user profile. For example, completing a diet record or a 10-minute stretching exercise can be considered as completing the daily goal, thus lowering the threshold for resuming behavior. When a user's churn probability exceeds the high-risk threshold, a more intensive and personalized deep intervention program is launched, such as daily targeted suggestions from an AI health coach, provision of phased review plans, and linkage with offline doctor consultations or human customer service support in scenarios such as chronic disease management. After the intervention is implemented, the system continuously tracks whether users resume checking in and whether they complete the phased goals again during the subsequent 7 or 14-day observation period. If a number of users are expected to succeed, the current strategy is solidified; if a number of users are expected to fail, the strategy combination is automatically adjusted. For example, the strategy may be upgraded from simple knowledge reminders to introducing social support and inviting friends to participate in the goals. The intervention mechanism is dynamically adjusted according to the risk of churn.

[0094] The embodiments of this application, through the above steps S101 to S105, not only improve the efficiency of health data utilization and modeling accuracy, but also enhance the individualization and continuous optimization capabilities of intervention strategies. They can continuously and effectively guide changes in user behavior in multiple scenarios, thereby improving the effectiveness of health management and overall operational efficiency.

[0095] Having described the method of an exemplary embodiment of this application, the following description will describe an intervention information push device according to an exemplary embodiment of this application. (Reference) Figure 2 As shown, the device includes: a data acquisition unit for acquiring multi-source health data, standardizing the multi-source health data, storing it in a unified data base, and generating health characteristic data; collecting task execution status and user feedback; a construction unit for constructing a personal health profile and behavior pattern based on the health characteristic data; a push unit for receiving health management goals for target users, inputting the personal health profile, behavior pattern, and health management goals into a health decision-making model, generating a personalized behavior intervention plan containing phased goals and specific task suggestions; pushing tasks in the behavior intervention plan to users through an interactive terminal bound to the user; and an adjustment unit for evaluating the intervention effect based on execution status and user feedback, updating the personal health profile and behavior pattern, and adjusting the behavior intervention plan according to the updated personal health profile and behavior pattern. The above device can implement the steps described in the above method implementation, and the specific implementation methods of each step will not be repeated here.

[0096] After introducing the methods and apparatus of the exemplary embodiments of this application, a terminal device of the exemplary embodiments of this application will be described next. The terminal device can implement the steps described in the above method embodiments, and the specific implementation of each step will not be repeated here.

[0097] After introducing the methods, apparatus, and terminal devices of exemplary embodiments of this application, the computer-readable storage medium of exemplary embodiments of this application will now be described. The computer-readable storage medium is an optical disc, on which a computer program (i.e., a program product) is stored. When the computer program is run by a processor, it implements the steps described in the above method embodiments. The specific implementation methods of each step will not be repeated here. It should be noted that the above embodiments are merely specific embodiments of this application, used to illustrate the technical solutions of this application, and not to limit it. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the scope of the technology disclosed in this application, or make equivalent substitutions for some of the technical features. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for intervening in information push, characterized in that, The method includes: Acquire multi-source health data, standardize the multi-source health data and store it in a unified data base to generate health feature data; A personal health profile and behavioral pattern are constructed based on the aforementioned health characteristic data; the health characteristic data is related to the target user's current situation. Receive health management goals for target users, input the personal health profile, the behavior pattern and the health management goals into the health decision model, and generate a personalized behavior intervention plan that includes phased goals and specific task suggestions; Tasks from the behavioral intervention plan are pushed to users through an interactive terminal bound to the user, and task execution status and user feedback are collected. The effectiveness of the intervention is evaluated based on the implementation status and user feedback. The personal health profile and the behavioral pattern are then updated, and the behavioral intervention plan is adjusted according to the updated personal health profile and the behavioral pattern.

2. The intervention information push method according to claim 1, characterized in that, After evaluating the intervention effect based on the implementation status and user feedback, updating the personal health profile and the behavioral pattern, and adjusting the behavioral intervention plan according to the updated personal health profile and behavioral pattern, the method further includes: Using the intervention effect prediction sub-model, based on the health feature data and the health management goals, the model predicts the trend of health indicator changes of target users under different combinations of intervention tasks for the candidate intervention task set. Using a compliance prediction sub-model, the probability of the target user completing the candidate intervention task is predicted; Based on the trends in the health indicators and the completion probabilities, candidate intervention tasks are combined and screened to generate phased goals and specific task suggestions for the personalized behavioral intervention plan.

3. The intervention information push method according to claim 1, characterized in that, The standardization process of the multi-source health data and its storage in a unified data foundation to generate health feature data includes: The original health data from different devices in the multi-source health data are unified in terms of timestamps and normalized in terms of units, and the original indicators such as steps, heart rate, sleep stage, food intake and mood score are converted into basic features under a unified scale. The basic features are resampled and time-aligned according to a preset time granularity to construct a time series feature matrix indexed by user identifier and time slice. Data quality processing is performed on the missing and outlier data in the time series feature matrix; Derived health characteristic indicators are calculated based on the processed time series feature matrix. Establish associated analytical dimension labels for the derived health feature indicators. The analytical dimension labels include at least physiological, psychological, behavioral, and contextual dimensions. Then, bind the derived health feature indicators with associated labels to the index information of the target user and store them in a unified database. Based on the profile request for the target user, health feature data that matches the target analysis dimension in the profile request is extracted from the unified database.

4. The intervention information push method according to claim 3, characterized in that, The construction of a personal health profile and behavioral pattern based on the health characteristic data includes: Based on the extracted health feature data, the comprehensive score and risk level of each target analysis dimension are calculated within a preset statistical period to obtain a multidimensional health score vector. Based on the multidimensional health score vector and the basic information of the target users, the target users are divided into preset population stratification and risk stratification categories to obtain population stratification labels used to characterize health risk level and management priority. Based on the historical intervention task completion status, application usage behavior, work and rest patterns and emotional fluctuation characteristics recorded in the health feature data, the target users are identified in terms of behavioral stages and clustered in terms of behavioral types to obtain behavioral pattern features that represent compliance, habit stability and intervention sensitivity. Based on the multidimensional health score vector and the population stratification labels, a personal health profile corresponding to the target user is generated; based on the behavioral pattern features, a behavioral pattern corresponding to the target user is generated.

5. The intervention information push method according to claim 1, characterized in that, The process of inputting the individual health profile, behavioral patterns, and health management goals into a health decision-making model to generate a personalized behavioral intervention plan that includes phased goals and specific task suggestions includes: Based on the aforementioned health management goals and the risk levels of each health dimension in the individual health profile, the priority intervention phased target dimensions and corresponding phased target intervals are determined. Candidate intervention tasks that match the phased target dimension are selected from the candidate intervention task set. The task type, execution content, intensity and frequency parameters, and recommended execution time window of the selected candidate intervention tasks are matched with the behavior pattern to obtain a subset of tasks to be evaluated that meet the target user's safety constraints and execution preferences. The intervention effect prediction sub-model and the compliance prediction sub-model are invoked to evaluate the health indicator change trends and task completion probabilities of each task combination in the task subset to be evaluated, so as to obtain the target task combination that matches the phased goal and has the best comprehensive score of expected effect and compliance. Based on the aforementioned combination of target tasks, a personalized behavioral intervention plan is generated, divided into stages.

6. The intervention information push method according to claim 1, characterized in that, The step of pushing tasks from the behavioral intervention plan to the user via an interactive terminal bound to the user includes: Obtain the target user's context information at the current moment; Based on the contextual information, the preference time periods and reach sensitivity recorded in the behavior pattern model, and the recommended execution time window and urgency of each task in the personalized behavior intervention plan, the push timing, push frequency and priority of the corresponding task are determined, and an intervention push strategy for the current moment is generated. According to the intervention push strategy, a target push channel and push format are selected from at least one push channel among mobile applications, instant messaging tools, smart speakers and wearable device reminders, and the task to be performed is pushed to the target user in the form of notification, dialogue or contextual reminder.

7. The intervention information push method according to claim 1, characterized in that, The step of pushing tasks from the behavioral intervention plan to the user via an interactive terminal bound to the user also includes: Acquire text interaction content, voice signals, and physiological signals related to the target user; Single-modal sentiment analysis was performed on text interaction content, voice signals and physiological signals to obtain first identification results related to emotion type and emotion intensity, and second identification results related to stress level and motivation level. Based on the first and second identification results, a multimodal fusion process combining feature-level fusion and decision-level fusion is performed, and the comprehensive emotional state and comprehensive motivation level of the target user are extracted from the fusion results. Based on the overall emotional state and overall motivation level, the corresponding communication postures and behavioral strategies are matched in the pre-constructed communication posture and behavior strategy matrix. The system generates empathic feedback content based on the target user's personal health profile and behavioral patterns to guide the target user in implementing the personalized behavioral intervention plan, and pushes the empathic feedback content to the target user through an interactive terminal bound to the target user.

8. The intervention information push method according to claim 1, characterized in that, The step of generating a personalized behavior intervention plan that includes phased goals and specific task suggestions also includes: The expert rule engine is invoked to constrain the applicable population, execution intensity, maximum frequency, and safety boundaries of candidate intervention tasks based on preset medical guidelines, exercise prescriptions, nutritional advice, and behavioral science theories. The data-driven learning model is invoked to learn the comprehensive impact of different task types, execution order and push rhythm on health improvement and compliance based on the set of historical intervention effect indicators, and outputs candidate task combination scores for different user profiles and behavior patterns. The set of compliance tasks output by the expert rule engine and the task combination scoring results output by the data-driven learning model are weighted and fused to obtain a comprehensive intervention plan; Based on the comprehensive intervention plan, a final personalized behavioral intervention plan is generated.

9. The intervention information push method according to claim 1, characterized in that, After evaluating the intervention effect based on the implementation status and user feedback, updating the personal health profile and the behavioral pattern, and adjusting the behavioral intervention plan according to the updated personal health profile and behavioral pattern, the method further includes: Based on the health characteristic data, personal health profiles, intervention execution data and intervention effect indicators of multiple target users, cluster analysis is performed on multiple target users to form multiple healthy population subclasses. The healthy population subclasses include at least combinations of different risk levels, different behavioral patterns and different target preferences. For each subcategory of healthy individuals, a corresponding group-level intervention strategy template is generated by combining historical intervention effect indicators and expert experience. The group-level intervention strategy template includes at least priority intervention dimensions, recommended task type combinations, and basic push rhythm. When generating the personalized behavior intervention plan for the target user, the personal health profile and behavior pattern of the target user are matched with the feature centers of each healthy population subclass to determine the healthy population subclass to which the target user belongs. A basic intervention plan is selected from the group-level intervention strategy templates corresponding to the subclass of healthy people. The task intensity, execution frequency and push timing are then fine-tuned based on the target user's multidimensional health score, risk stratification label and behavioral pattern characteristics to obtain the final intervention plan for the target user.

10. An intervention information push device, characterized in that, The device includes the following units: The acquisition unit is used to acquire multi-source health data, standardize the multi-source health data and store it in a unified data base to generate health feature data. Collect task execution status and user feedback; The construction unit is used to construct a personal health profile and behavior pattern based on the health feature data; The push unit is used to receive health management goals for target users, input the personal health profile, the behavior pattern and the health management goals into the health decision model, generate a personalized behavior intervention plan that includes phased goals and specific task suggestions, and push the tasks in the behavior intervention plan to the user through an interactive terminal bound to the user. The adjustment unit is used to evaluate the intervention effect based on the execution status and user feedback, update the personal health profile and the behavior pattern, and adjust the behavior intervention plan according to the updated personal health profile and the behavior pattern.

Images