A patient health monitoring system integrating behavior intervention, medication management, and examination

Abstract

The application discloses a patient health monitoring system integrating behavior intervention, medication management and examination, and belongs to the field of intelligent medical health management. The system comprises a data acquisition module, an abnormality analysis module, an intervention classification module, a path optimization module, a scheme adjustment module, a strategy matching module and a push execution module. The data acquisition module acquires patient behavior, medication and examination data. The abnormality analysis module fuses random forest and time series graph neural network to detect abnormal deviation. The intervention classification module adopts contrast learning and prototype network to obtain individualized intervention categories. The path optimization module generates an optimized path based on multi-objective deep reinforcement learning. The scheme adjustment module evaluates intervention net benefits by using double machine learning and counterfactual reasoning. The strategy matching module dynamically updates parameters through meta-learning and context multi-armed tiger machine. The push execution module generates a notification signal and acquires feedback. The application solves the problem that the existing system lacks dynamic adaptation capability, realizes precise and individualized real-time intervention, and significantly improves chronic disease management effect and patient compliance.

Description

Technical Field

[0001] This invention relates to the field of patient health monitoring technology, and in particular to a patient health monitoring system that integrates behavioral intervention, medication management, and examination. Background Technology

[0002] In the field of modern healthcare management, the importance of patient health monitoring systems is self-evident. They directly impact the daily management and disease control outcomes of patients with chronic diseases, serving as a key pillar for improving the quality of medical services and patients' quality of life. With the increasing aging population and rising incidence of chronic diseases, the construction of a comprehensive and intelligent monitoring system has become an urgent need. Such a system needs to integrate behavioral guidance, medication use monitoring, and regular checkups to provide patients with all-round health support.

[0003] However, current health monitoring tools often share a common problem: they lack deep interaction and dynamic adaptability when integrating multifaceted information. While many systems can collect various patient data, they cannot effectively adjust to individual differences and real-time changes, leading to interventions that are often out of touch with the patient's actual needs. This deficiency not only reduces patient trust in the system but also affects the actual effectiveness of health management, especially when facing complex conditions or multiple health problems, where the system often proves inadequate.

[0004] A deeper technological challenge lies in achieving precise matching and dynamic optimization of interventions. First, precise matching requires the system to comprehensively analyze a patient's behavioral habits, medication history, and test results to find the most suitable timing and method of intervention. However, due to significant differences in patients' physical conditions and lifestyles, the system often struggles to personalize interventions when processing this complex information. Second, achieving this precise matching further complicates dynamic optimization, as a patient's health status and response to intervention change over time. If the system cannot adjust its strategies promptly, the intervention's effectiveness may be significantly reduced. For example, when managing hypertensive patients, the system might need to remind them to exercise more on a particular day based on their morning high blood pressure. However, if the patient is unable to comply the next day due to fatigue, and the system fails to adjust the reminder content or method in time, this rigid intervention logic will frustrate the patient.

[0005] Therefore, achieving precise matching and dynamic optimization of intervention measures within a patient health monitoring system has become a key issue in improving system usability and user experience. Solving this problem requires not only technological breakthroughs but also a deep understanding of individual patient needs and a comprehensive optimization of the health management process. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a patient health monitoring system that integrates behavioral intervention, medication management and examination, in order to overcome the shortcomings of the prior art.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A patient health monitoring system integrating behavioral intervention, medication management, and examination, specifically including: The data acquisition module is used to collect patient behavior data, medication records, and regular check-up results through sensors and wearable devices to obtain comprehensive patient data; The anomaly analysis module is used to extract behavioral and health indicator features from comprehensive patient data, and calculates them using a random forest classifier. The divergence value is used as an anomaly deviation value, and a time-series graph neural network is further used to perform dynamic anomaly detection on multi-source heterogeneous data; The intervention classification module is used to obtain personalized intervention categories by performing deep clustering of behavioral patterns and health indicators through contrastive learning and prototype networks if the abnormal deviation value exceeds the preset threshold. The path optimization module is used to extract real-time change features from personalized intervention categories, simulate the intervention effect using a multi-objective deep reinforcement learning model, and generate an optimized path. The intervention adjustment module is used to integrate drug monitoring data and periodic inspection results through optimized pathways, and to evaluate the net benefit of the intervention based on causal inference and counterfactual reasoning, so as to obtain the adjusted intervention plan. The strategy matching module is used to determine the final matching strategy by updating the dynamic adaptation parameters through meta-learning and contextual multi-armed slot machine if the adjusted intervention plan matches the historical response data. The push execution module is used to generate notification signals based on the final matching strategy, push behavioral guidance instructions to the patient's device, and obtain execution feedback data.

[0008] Preferably, the data acquisition module includes: The real-time acquisition unit is used to collect patient behavior data, medication records, and examination results in real time. The data integration unit is used for unified formatting and integration into structured comprehensive data; The data cleaning unit is used to remove outliers and missing values; Data imputation unit, used to fill in missing values ​​whose missing proportion exceeds a threshold by historical interpolation; Anomaly classification unit is used to classify abnormal behavior or medication patterns using support vector machines; The correlation analysis unit is used to analyze the correlation between behavior and examination results, and to obtain the correspondence between behavior and health status. The strategy generation unit is used to generate personalized monitoring strategies and adjustment suggestions based on the corresponding relationships.

[0009] Preferably, the anomaly analysis module includes: The classification processing unit is used to classify behavioral patterns and health indicators based on comprehensive patient data using the random forest algorithm, and to determine abnormal deviation values. The grouping unit is used to group patient groups based on abnormal deviation values ​​using cluster analysis to obtain behavioral pattern groups; The rule extraction unit is used to extract key behavioral rules for each group based on behavioral patterns using a decision tree algorithm, thereby obtaining a rule set. The risk determination unit is used to perform correlation matching between the rule set and health indicator deviations to determine whether there are significant deviation correspondences and identify high-risk behavior patterns. The threshold adjustment unit is used to generate personalized early warning threshold data based on high-risk behavior patterns to obtain the adjusted monitoring threshold. The risk marking unit is used to compare and judge the subsequent collected data in real time according to the adjusted monitoring threshold. If the threshold is exceeded, the risk status is marked and a risk warning signal is determined. The level determination unit is used to conduct trend comparison analysis on risk warning signals in combination with historical behavior patterns to determine whether there is a continuous deviation and obtain the enhanced warning level. The anomaly analysis module is also configured as follows: Constructing a temporal heterogeneous graph ; in, For time indexing, A set of nodes, containing behavior nodes. Medication timing and physiological nodes , For a moment The edge set represents the dynamic dependencies between nodes; Update node features using graph attention mechanism: ; in: For nodes At any moment The updated feature vector, For activation function, For nodes The set of neighboring nodes, For nodes To the neighbors Attention weights The weight matrix is ​​a learnable matrix. For nodes At any moment The original feature vector, For attention parameter vectors, This represents vector concatenation. It is a non-linear activation function; Modeling temporal dependencies using temporal convolutional networks: ; in: For a moment Embedment matrix of all nodes, The time window length, Indicates temporal convolutional network operations; Calculate the outlier score: ; The first term represents the graph reconstruction error. For decoder, For the encoder's implicit variable output, For the Euclidean norm; in the second term For balance coefficient, express divergence, This represents the temporal distribution at the current moment. For historical time series distribution; if If the value exceeds the preset threshold, it is considered abnormal, and the abnormal score is used as the abnormal deviation value.

[0010] Preferably, the intervention classification module is also configured as follows: The embedding representation is optimized through contrastive learning, and the contrastive loss function is: ; in: The total number of samples, For the first The embedding vector of each sample, For the first The positive sample embedding vector of each sample. For the first Each negative sample embedding vector The cosine similarity function is used. Temperature coefficient; Build a prototype network for each intervention category. The prototype vector is: ; in: To belong to category The sample set, The number of samples in this set. For the sample Category tags; The classification probability of the new sample is: ; in: Represents the sample embedding vector Category The probability, For Euclidean distance, The total number of intervention categories; individualized intervention categories are determined based on the highest probability.

[0011] Preferably, the path optimization module is also configured as follows: Define a multi-objective reward vector: ; in: For a moment The reward vector, For the rate of improvement of health indicators, To improve medication adherence, For intervention costs or patient burden; Multi-objective Q-learning update: For each objective , ; in: For a moment In state Take action below Time corresponding target Action value function, For learning rate, For the goal Instant rewards As a discount factor, For the next state, For the next action; Define the advantage vector: ; in: As the dominant vector, , This is a state value vector; Strategy selection criteria: ; in: In the state The strategy of selection For the patient preference weight vector, Corresponding target The weights are determined based on the selected strategy to determine the optimization path.

[0012] Preferably, the scheme adjustment module is also configured as follows: Using dual machine learning to estimate the causal effects of the intervention: in: For the total sample size, For sample index, For the first Intervention indicator variables for each sample ( Indicates acceptance of intervention. (This indicates that the offer was not accepted). For covariate vectors, For the propensity score estimation function (i.e., given covariates) (Probability of accepting intervention) The estimated function for the resulting model, For observed health outcomes, and These represent the predicted outcomes under intervention and non-intervention conditions, respectively. Calculate the net benefit of individual interventions: ; in: For the first Net intervention benefit for each individual To accept the potential outcomes of the intervention, For potential outcomes without intervention, Given covariates and intervention status The conditional expectation function is used; if the net benefit of an individual intervention is positive, the intervention is deemed effective, and an adjusted intervention plan is generated based on the characteristics of an effective intervention.

[0013] Preferably, the strategy matching module is also configured as follows: Rapid adaptation using a meta-learning framework: inner layer update Outer layer optimization ; in: These are the initial parameters for the meta-learner. In the first The task refers to the updated parameters of the patient's inner layer. The inner learning rate, The outer learning rate, For gradient operators, For the first The loss function for each task Total number of tasks; Contextual multi-armed slot machine strategy selection: ; in: For a moment The chosen action is the intervention strategy. For a set of actions, For action The regression parameter vector, For a moment Context feature vectors, To explore coefficients, For action The characteristic covariance matrix, superscript Indicates vector transpose, superscript This represents finding the inverse of a matrix. Posterior probability iteration is achieved through Bayesian online updates: ; in: Given historical data The posterior distribution of the parameters after that, For the prior distribution of parameters, For observations given context and parameters The likelihood function is used to update the dynamic adaptation parameters based on the posterior probability, and the final matching strategy is determined.

[0014] Preferably, the push execution module includes: A signal generation unit is used to generate a notification signal according to the final matching strategy; The instruction conversion unit is used to convert notification signals into behavior guidance instructions according to preset encoding rules; The target determination unit is used to bind instruction content through patient device identification; The instruction sending unit is used to push instructions to the patient device. The feedback acquisition unit is used to acquire execution feedback data; The feedback parsing unit is used to parse the feedback data in a structured manner and extract the confirmation flags; The status determination unit is used to determine the policy trigger status based on the confirmation flag associated with the final matching policy.

[0015] Preferably, it also includes a comprehensive evaluation module for intervention effects, configured as follows: Calculate the overall score of intervention effect: in: To assess the overall effectiveness of the intervention, For the rate of improvement of health indicators, To improve medication adherence, For intervention costs or patient burden, The model interpretability score, ranging from 0 to 1, is obtained by weighting indicators such as feature importance consistency and counterfactual plausibility. For the corresponding weight coefficients, satisfying And all coefficients are non-negative; Calculate the dynamic adaptability index: ; in: As a dynamic adaptability indicator, The total number of samples, This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. and These are the model's response to samples before and after the policy update. The predicted probability, The standard deviation of the predicted probabilities before the policy update. The preset sensitivity threshold; if If the value exceeds the preset standard, a model update signal is generated and fed back to the policy matching module.

[0016] A patient health monitoring method integrating behavioral intervention, medication management, and examination, applied to the aforementioned patient health monitoring system integrating behavioral intervention, medication management, and examination, includes the following steps: Step S101: Collect patient behavior data, medication use records, and regular check-up results to obtain comprehensive patient data; Step S102: Extract features from comprehensive patient data, calculate KL divergence as anomaly bias value using a random forest classifier, and use a temporal graph neural network for dynamic anomaly detection; Step S103: If the abnormal deviation value exceeds the threshold, deep clustering is performed through contrastive learning and prototype network to obtain personalized intervention categories; Step S104: Extract real-time change features from personalized intervention categories, use multi-objective deep reinforcement learning to simulate intervention effects, and generate an optimized path; Step S105: Integrate drug surveillance data and inspection results through optimized pathways, assess the net benefit of the intervention based on causal inference, and obtain the adjusted intervention plan; Step S106: If the adjusted intervention plan matches the historical response data, then update the dynamic adaptation parameters through meta-learning and contextual multi-armed slot machine to determine the final matching strategy; Step S107: Generate a notification signal based on the final matching strategy, push behavioral guidance instructions to the patient device, and obtain execution feedback data; Step S108: Calculate the comprehensive score and dynamic adaptability index using the intervention effect comprehensive evaluation module, and provide feedback on optimization strategies based on the evaluation results.

[0017] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects: This invention integrates patient behavior data, medication records, and regular check-up results into comprehensive data through a data acquisition module; an anomaly analysis module uses a random forest algorithm to classify behavioral patterns and health indicators, identifying abnormal deviations; an intervention classification module generates personalized intervention categories through cluster analysis when thresholds are exceeded; a path optimization module combines a deep reinforcement learning model to simulate intervention effects and optimize the adjustment plan; and a strategy matching module updates parameters and pushes guidance instructions based on historical response data. Addressing the core issue of achieving precise intervention in dynamically changing health management, this invention innovatively integrates behavior monitoring, data analysis, and real-time feedback mechanisms. Through the collaborative work of multiple algorithms, it ensures the personalization and adaptability of intervention plans, significantly improving the efficiency and effectiveness of health management, particularly demonstrating unique value in chronic disease management and behavior guidance scenarios, providing patients with scientific and timely health support. (See attached figures.)

[0018] Figure 1 The flowchart of the patient health monitoring system integrating behavioral intervention, medication management, and examination according to the present invention is shown below; Figure 2 This is a schematic diagram of the data acquisition module of the present invention; Figure 3 This is a schematic diagram of the anomaly analysis module of the present invention; Figure 4 This is a schematic diagram of the intervention classification module of the present invention; Figure 5 This is a schematic diagram of the path optimization module of the present invention; Figure 6 This is a schematic diagram of the adjustment module of the present invention; Figure 7 This is a schematic diagram of the strategy matching module of the present invention; Figure 8 This is a schematic diagram of the push execution module of the present invention; Figure 9 This is a flowchart of the patient health monitoring method integrating behavioral intervention, medication management, and examination, as described in this invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this specification, and not all embodiments. Based on the embodiments in this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this specification.

[0020] like Figures 1 to 9 As shown, a patient health monitoring system integrating behavioral intervention, medication management, and examination, such as... Figure 1 As shown, it specifically includes: The data acquisition module is used to collect patient behavior data, medication records, and regular check-up results through sensors and wearable devices to obtain comprehensive patient data; The anomaly analysis module is used to extract behavioral and health indicator features from comprehensive patient data, and calculates them using a random forest classifier. The divergence value is used as an anomaly deviation value, and a time-series graph neural network is further used to perform dynamic anomaly detection on multi-source heterogeneous data; The intervention classification module is used to obtain personalized intervention categories by performing deep clustering of behavioral patterns and health indicators through contrastive learning and prototype networks if the abnormal deviation value exceeds the preset threshold. The path optimization module is used to extract real-time change features from personalized intervention categories, simulate the intervention effect using a multi-objective deep reinforcement learning model, and generate an optimized path. The intervention adjustment module is used to integrate drug monitoring data and periodic inspection results through optimized pathways, and to evaluate the net benefit of the intervention based on causal inference and counterfactual reasoning, so as to obtain the adjusted intervention plan. The strategy matching module is used to determine the final matching strategy by updating the dynamic adaptation parameters through meta-learning and contextual multi-armed slot machine if the adjusted intervention plan matches the historical response data. The push execution module is used to generate notification signals based on the final matching strategy, push behavioral guidance instructions to the patient's device, and obtain execution feedback data.

[0021] like Figure 2 As shown, the data acquisition module is used to collect patient behavior data, medication records, and periodic examination results through sensors and wearable devices to obtain comprehensive patient data.

[0022] The data acquisition module includes: The real-time acquisition unit is used to collect patient behavior data, medication records, and examination results in real time. The data integration unit is used for unified formatting and integration into structured comprehensive data; The data cleaning unit is used to remove outliers and missing values; Data imputation unit, used to fill in missing values ​​whose missing proportion exceeds a threshold by historical interpolation; Anomaly classification unit is used to classify abnormal behavior or medication patterns using support vector machines; The correlation analysis unit is used to analyze the correlation between behavior and examination results, and to obtain the correspondence between behavior and health status. The strategy generation unit is used to generate personalized monitoring strategies and adjustment suggestions based on the corresponding relationships.

[0023] Real-time data collection of patient behavior, medication use, and routine check-ups is achieved using sensor devices and wearable devices, yielding preliminary patient behavior data, usage records, and examination results. Based on the collected patient behavior data, usage records, and examination results, a standardized data formatting method is used to integrate them into structured comprehensive data. For the integrated comprehensive data, pre-established data cleaning rules are used to remove outliers and missing values, resulting in cleaned comprehensive data.

[0024] If the proportion of missing data in the cleaned aggregate data exceeds a preset threshold, historical data interpolation is used to impute the missing data, resulting in the imputed aggregate data. Based on the imputed aggregate data, a support vector machine algorithm is used to classify patient behavior data and medication use records to determine if any abnormal behaviors or medication patterns exist. The classification results, combined with periodic check-up results, are used to analyze the correlation between patient behavior and check-up results, obtaining a correspondence between behavior and health status. Personalized monitoring strategy data is generated based on this correspondence between behavior and health status, and targeted adjustment suggestions are output.

[0025] For example, in remote monitoring scenarios for patients with chronic diseases, wearable devices and sensors, such as smart bracelets and smart medicine bottle caps, collect real-time data on patients' daily activities, medication times, and regular blood pressure and blood sugar self-monitoring, forming preliminary behavioral data, usage records, and examination results. This data may initially be fragmented and requires standardized formatting. Specifically, the steps and heart rate recorded by the bracelets are standardized in JSON format, the timestamps of medicine bottle cap openings are standardized to UTC time, and the examination results are converted into numerical fields, thus integrating them into a structured, comprehensive dataset. This standardized processing helps subsequent algorithms read the data efficiently and avoids errors caused by format differences.

[0026] In one possible implementation, pre-established cleaning rules are applied to the integrated data. For example, outliers such as heart rates exceeding 200 beats per minute or obvious errors like daily step counts exceeding 50,000 are removed, along with records missing more than 30% of entries, resulting in cleaned integrated data. If a patient's medication records are 40% missing within a week due to equipment malfunction, exceeding a preset threshold of 20%, historical data interpolation methods are used, such as linear imputation based on the average opening time of the same period over the previous four weeks, to determine the imputed integrated data. This imputation method significantly improves data integrity and ensures the reliability of subsequent analyses.

[0027] For example, based on the imputed comprehensive data, a support vector machine (SVM) algorithm is used to classify patients' step count, sleep duration, and medication time. Specifically, normal behavioral patterns are classified as positive, while abnormalities such as staying up late or missing medication are classified as negative. Hyperplane partitioning is used to determine whether abnormal behavior or medication patterns exist. If a hypertensive patient has fewer than 2000 steps for three consecutive days and misses two medications, the algorithm can classify it as an abnormal pattern. This classification accuracy is usually high, effectively and promptly identifying risky behaviors and leading to the beneficial effect of early intervention.

[0028] In one embodiment, the correlation between low activity levels and elevated blood pressure is analyzed by combining classification results with regular check-up results, such as blood pressure values ​​exceeding 140 / 90 mmHg. For example, it was found that blood pressure increased by an average of 15 mmHg on days when medication was missed, thus establishing a correlation between behavior and health status: when medication was taken regularly and daily steps exceeded 5000, blood pressure was controlled within 130 / 80 mmHg. This correlation analysis reveals the specific impact of behavior on health, providing a basis for precision management.

[0029] For example, personalized monitoring strategy data can be generated based on the above correspondence. This could include increasing the frequency of step reminders for patients with low activity levels, sending double medication alerts to patients at high risk of missed doses, and ultimately outputting targeted adjustment suggestions, such as "It is recommended to increase daily activity by 3,000 steps and strictly adhere to medication schedules." This personalized strategy can significantly improve patient adherence, reduce the risk of relapse, and lead to better health management and optimization of medical resources.

[0030] Understandably, the entire process, from data collection to strategy generation, forms a closed loop, with each step ensuring data quality and analytical accuracy, ultimately achieving precise real-time monitoring and intervention for individual patients with chronic diseases.

[0031] like Figure 3 As shown, the anomaly analysis module is used to classify behavioral patterns and health indicators based on the comprehensive patient data using the random forest algorithm, and to determine abnormal deviation values.

[0032] The anomaly analysis module includes: The classification processing unit is used to classify behavioral patterns and health indicators based on comprehensive patient data using the random forest algorithm, and to determine abnormal deviation values. The grouping unit is used to group patient groups based on abnormal deviation values ​​using cluster analysis to obtain behavioral pattern groups; The rule extraction unit is used to extract key behavioral rules for each group based on behavioral patterns using a decision tree algorithm, thereby obtaining a rule set. The risk determination unit is used to perform correlation matching between the rule set and health indicator deviations to determine whether there are significant deviation correspondences and identify high-risk behavior patterns. The threshold adjustment unit is used to generate personalized early warning threshold data based on high-risk behavior patterns to obtain the adjusted monitoring threshold. The risk marking unit is used to compare and judge the subsequent collected data in real time according to the adjusted monitoring threshold. If the threshold is exceeded, the risk status is marked and a risk warning signal is determined. The level determination unit is used to conduct trend comparison analysis on risk warning signals in combination with historical behavior patterns to determine whether there is a continuous deviation and obtain the enhanced warning level. The anomaly analysis module is also configured as follows: Constructing a temporal heterogeneous graph ; in, For time indexing, A set of nodes, containing behavior nodes. Medication timing and physiological nodes , For a moment The edge set represents the dynamic dependencies between nodes; Update node features using graph attention mechanism: ; in: For nodes At any moment The updated feature vector, For activation function, For nodes The set of neighboring nodes, For nodes To the neighbors Attention weights The weight matrix is ​​a learnable matrix. For nodes At any moment The original feature vector, For attention parameter vectors, This represents vector concatenation. It is a non-linear activation function; Modeling temporal dependencies using temporal convolutional networks: ; in: For a moment Embedment matrix of all nodes, The time window length, Indicates temporal convolutional network operations; Calculate the outlier score: ; The first term represents the graph reconstruction error. For decoder, For the encoder's implicit variable output, For the Euclidean norm; in the second term For balance coefficient, express divergence, This represents the temporal distribution at the current moment. For historical time series distribution; if If the value exceeds the preset threshold, it is considered abnormal, and the abnormal score is used as the abnormal deviation value.

[0033] Based on comprehensive patient data, a random forest algorithm is applied to classify behavioral patterns and health indicators, identifying abnormal deviation values. Cluster analysis is then used to group patient groups based on these abnormal deviation values, resulting in behavioral pattern groups. A decision tree algorithm is then used to extract key behavioral rules for each group, creating a rule set. These rule sets are then correlated with health indicator deviations to determine if significant deviations exist, identifying high-risk behavioral patterns. Personalized warning threshold data is generated from these high-risk behavioral patterns, resulting in adjusted monitoring thresholds. Subsequent data is then compared in real-time against these adjusted thresholds; if the threshold is exceeded, a risk status is marked, establishing a risk warning signal. Finally, trend analysis is performed on these risk warning signals in conjunction with historical behavioral pattern groups to determine if persistent deviations exist, resulting in a strengthened warning level.

[0034] For example, in remote monitoring scenarios for patients with chronic diseases, the comprehensive analysis and processing of patient data is particularly important. For the classification and processing of behavioral patterns and health indicators, the random forest algorithm can be used to identify potential outliers. Specifically, the random forest constructs multiple decision trees to comprehensively evaluate a patient's daily activity data, such as steps, sleep duration, and health indicators, such as blood pressure and blood sugar, to determine if any abnormalities exist. For instance, if a patient's step count is below 1000 steps for several consecutive days, while their blood pressure is above the normal range, the algorithm will mark this as a potential outlier. This approach can quickly filter out data points that require attention.

[0035] For example, when performing cluster analysis based on outlier values, the K-means method can be used to divide the patient population into groups with different behavioral patterns. One possible implementation is to group patients with similar low activity levels and hypertension into one group, and those with higher activity levels and stable health indicators into another. For instance, if a group of patients averages only 1500 steps per day and generally has high blood pressure, they can be defined as a high-risk behavioral pattern group. This grouping facilitates subsequent targeted analysis.

[0036] For example, when using decision tree algorithms to extract key behavioral rules for each group, certain specific patterns can be identified. For instance, in the low-activity group, if patients take fewer than 2000 steps per day and sleep less than 6 hours, the probability of elevated blood pressure increases significantly. This set of rules provides a clear basis for subsequent interventions. When the set of rules is correlated with deviations in health indicators, if a significant correlation is found between steps below a certain threshold and elevated blood pressure, this can be identified as a high-risk behavioral pattern.

[0037] For example, when generating personalized warning threshold data, dynamic thresholds can be set based on patients' historical data. Suppose a patient's normal step count ranges from 4000 to 6000 steps. If the step count falls below 3000 steps for three consecutive days, the monitoring threshold is adjusted to 2500 steps, and this is used as the benchmark for subsequent real-time comparisons. If newly collected data falls below this threshold, it is marked as a risk state, and a warning signal is issued. This personalized threshold design better suits individual circumstances.

[0038] For example, when conducting trend comparison analysis on risk warning signals, historical behavioral patterns can be grouped to determine whether the deviation persists. If a patient's step count is below the threshold for a consecutive week and highly matches the characteristics of the low-activity group, the warning level can be raised, indicating the need for enhanced intervention. This trend analysis helps to promptly detect potential signs of deterioration. Through the above multi-level analysis, from classification to grouping to warning, a complete data processing chain is formed. This approach not only accurately identifies high-risk behavioral patterns but also enables real-time monitoring and intervention through dynamic thresholds and trend analysis, significantly improving the targeting and effectiveness of chronic disease management.

[0039] like Figure 4 As shown, the intervention classification module is used to perform cluster analysis on behavioral patterns and health indicators to obtain personalized intervention categories if the abnormal deviation value exceeds a preset threshold.

[0040] The intervention classification module is also configured as follows: The embedding representation is optimized through contrastive learning, and the contrastive loss function is: ; in: The total number of samples, For the first The embedding vector of each sample, For the first The positive sample embedding vector of each sample. For the first Each negative sample embedding vector The cosine similarity function is used. Temperature coefficient; Build a prototype network for each intervention category. The prototype vector is: ; in: To belong to category The sample set, The number of samples in this set. For the sample Category tags; The classification probability of the new sample is: ; in: Represents the sample embedding vector Category The probability, For Euclidean distance, The total number of intervention categories; individualized intervention categories are determined based on the highest probability.

[0041] This module is configured to perform the following steps: If abnormal deviations exceeding the preset threshold are detected through data collection, the detection process is initiated to perform preliminary screening of the collected behavioral patterns and health indicators, and to obtain preliminary classification results.

[0042] Based on the preliminary classification results, cluster analysis was used to deeply group behavioral patterns and health indicators to determine the basis for individualized category division.

[0043] Based on the criteria for classifying personalized categories, and combined with the assessment data of health indicators, corresponding intervention categories are obtained, resulting in a specific set of intervention categories.

[0044] By using intervention classification sets, matching the pattern classification features in behavioral patterns, refining the labeling of each personalized category, and determining the priority ranking of interventions.

[0045] If the evaluation data of a certain category of indicators in the intervention priority ranking continues to deviate from the preset threshold, the data collection frequency of that category will be dynamically adjusted to obtain an updated monitoring dataset.

[0046] Based on the updated monitoring dataset, the deviation assessment is re-executed, the correlation between abnormal deviations and intervention categories is analyzed, and the final intervention adjustment plan is obtained.

[0047] For the final intervention and adjustment plan, the latest data on behavioral patterns and health indicators will be incorporated into the testing process, and personalized categories will be continuously updated to determine the direction of subsequent interventions.

[0048] For example, in the daily monitoring of patients with chronic diseases, the detection and intervention of abnormal deviations in behavioral patterns and health indicators can be achieved through a series of systematic analyses and processes. When data collection reveals abnormal deviations exceeding preset thresholds, preliminary screening is a crucial first step. For instance, if a patient's daily step count is typically around 5000 steps, but has recently fallen below 2000 steps for three consecutive days, and their heart rate exceeds the upper limit of the normal range by 10 percentage points, the system will mark this as abnormal and initiate the detection process. This process will perform preliminary classification of behavioral data and physiological indicators, distinguishing potential risk signals.

[0049] For example, deep grouping is particularly important based on preliminary classification results. Cluster analysis based on behavioral patterns and health indicators can divide patients into different personalized categories. For instance, by analyzing step count, sleep quality, and blood glucose fluctuation data, patients can be divided into a low-activity group and a normal-activity group. Patients in the low-activity group may average less than 3,000 steps per day and sleep less than 6 hours, while the normal-activity group is the opposite. This classification provides a basis for subsequent interventions.

[0050] For example, combining health indicator assessment data to determine the intervention category is a crucial next step. Assuming patients in the low-activity group generally have high blood sugar, the system would categorize them into an intervention category requiring dietary adjustments and exercise promotion, while those with sleep deprivation might be categorized into an intervention category requiring psychological support. These category sets lay the foundation for subsequent refined interventions.

[0051] For example, when matching behavioral pattern features, refined labeling and prioritization are indispensable. For the low-activity group, if a patient not only has a low step count but also recent dietary records show high sugar intake, the system will label them as a high-priority intervention recipient and prioritize health guidance. This prioritization method ensures the rational allocation of resources.

[0052] For example, when a metric for a particular category consistently deviates from a threshold, dynamically adjusting the data collection frequency becomes necessary. Suppose a patient takes fewer than 2000 steps for a consecutive week; the system would adjust the step monitoring frequency from once daily to once every 6 hours to obtain a more detailed dataset. This adjustment helps capture subtle changes.

[0053] For example, based on the updated dataset, it is crucial to reassess the correlation between the deviation and the intervention category. If a significant correlation is found between persistently low step counts and elevated blood sugar, the system will adjust the intervention plan, increasing the frequency and intensity of exercise guidance. This correlation analysis provides a basis for precise intervention.

[0054] For example, the continuous updating of the final intervention plan forms a closed loop in the entire process. The system incorporates the latest data into the monitoring and dynamically adjusts the personalized categories. For instance, if a patient's step count gradually increases to 4,000 steps after intervention, the system will reassess their risk level and adjust the direction of subsequent interventions. This continuous updating approach ensures the timeliness and targeted nature of the intervention.

[0055] like Figure 5 As shown, the path optimization module is used to extract real-time change features from the personalized intervention categories, simulate the intervention effect using a deep reinforcement learning model, and determine the optimal path.

[0056] The path optimization module is also configured as follows: Define a multi-objective reward vector: ; in: For a moment The reward vector, For the rate of improvement of health indicators, To improve medication adherence, For intervention costs or patient burden; Multi-objective Q-learning update: For each objective , ; in: For a moment In state Take action below Time corresponding target Action value function, For learning rate, For the goal Instant rewards As a discount factor, For the next state, For the next action; Define the advantage vector: ; in: As the dominant vector, , This is a state value vector; Strategy selection criteria: ; in: In the state The strategy of selection For the patient preference weight vector, Corresponding target The weights are determined based on the selected strategy to determine the optimization path.

[0057] Real-time change features are extracted from personalized intervention categories to generate state representation vectors. Based on these state representation vectors, a deep reinforcement learning model is input to obtain an action selection policy. For the action selection policy, a reward function is defined and designed, and the intervention effect is simulated. Through intervention effect simulation, simulated trajectory data is generated, and a path evaluation score is determined. If the path evaluation score is higher than the current record, the action selection policy is updated to obtain an optimized intervention plan. Based on the optimized intervention plan, the state representation vector is adjusted, and the updated real-time change features are obtained. From the updated real-time change features, a new state representation vector is generated, and the optimized path is determined.

[0058] For example, in the daily monitoring of patients with chronic diseases, real-time feature extraction for personalized intervention categories can be achieved through multi-dimensional data analysis. To analyze changes in behavioral patterns and health indicators, the system records data such as the patient's steps, heart rate, and sleep duration in real time, generating a multi-dimensional state representation vector. This vector reflects the patient's current physical state and behavioral trends, providing a foundation for subsequent analysis.

[0059] For example, during the input of the state representation vector into a deep reinforcement learning model, the system predicts the optimal intervention strategy based on historical data and the current state. Suppose a patient's step count continuously declines to below 2000 steps per day, while their heart rate is high; the system might suggest increasing light exercise and adjusting their diet as initial intervention actions. The generation of this strategy relies on the model's comprehensive assessment of the patient's condition.

[0060] For example, the reward function design for action selection strategies can be implemented through simulation of intervention effects. The system sets a reward mechanism based on the execution of the intervention actions; for instance, if the patient's step count increases to over 3000 steps per day after the intervention, the reward score increases; if there is no significant improvement, the score decreases. By simulating the intervention effects, the system generates simulated trajectory data and evaluates the scores of the intervention paths. If the score of a certain intervention path is higher than the current record, the system updates the action selection strategy to form a better intervention optimization plan.

[0061] For example, when adjusting the state representation vector, the system re-extracts real-time change features based on the optimized scheme. Suppose that after intervention, the patient's step count increases to 3500 steps and their heart rate returns to the normal range, the system updates the state vector to reflect the latest health status. This process ensures the dynamic adaptability of the intervention.

[0062] For example, after regenerating the state representation vector from the updated real-time change features, the system determines whether the optimized path is reasonable. If the score of the new path is consistently higher than the historical record, such as when the patient's sleep duration increases from 5 hours to 7 hours after the intervention, the system confirms that the path is the optimal solution. This closed-loop adjustment mechanism can continuously optimize the intervention effect.

[0063] For example, in determining the path assessment score, the system also incorporates multi-faceted data for verification. Suppose a patient's dietary records show a decrease in sugar intake while increasing physical activity; the system will comprehensively assess the impact of these changes on overall health, ensuring the comprehensiveness and scientific validity of the intervention plan. This multi-dimensional verification approach helps improve the targeted nature of the intervention.

[0064] For example, regarding the continuous adjustment of the intervention optimization plan, the system dynamically updates the strategy based on patient feedback and data changes. For instance, if a patient experiences a significant increase in steps at the beginning of the intervention but later feels fatigued, the system will appropriately reduce the exercise intensity and increase rest recommendations. This flexibility ensures the sustainability of the intervention plan, providing patients with support that better meets their actual needs.

[0065] like Figure 6 As shown, the intervention plan adjustment module is used to integrate drug monitoring data and periodic inspection results through the optimization path to obtain the adjusted intervention plan.

[0066] The scheme adjustment module is also configured as follows: Using dual machine learning to estimate the causal effects of the intervention: in: For the total sample size, For sample index, For the first Intervention indicator variables for each sample ( Indicates acceptance of intervention. (This indicates that the offer was not accepted). For covariate vectors, For the propensity score estimation function (i.e., given covariates) (Probability of accepting intervention) The estimated function for the resulting model, For observed health outcomes, and These represent the predicted outcomes under intervention and non-intervention conditions, respectively. Calculate the net benefit of individual interventions: ; in: For the first Net intervention benefit for each individual To accept the potential outcomes of the intervention, For potential outcomes without intervention, Given covariates and intervention status The conditional expectation function is used; if the net benefit of an individual intervention is positive, the intervention is deemed effective, and an adjusted intervention plan is generated based on the characteristics of an effective intervention.

[0067] This module is configured to perform the following steps: Daily medication data is obtained from drug monitoring records and combined with physiological indicator data from regular check-up results to create a comprehensive health dataset.

[0068] For a comprehensive health dataset, a pre-established classification model is used to analyze the correspondence between drug monitoring data and examination results to determine the health status classification.

[0069] Based on the health status classification, the corresponding intervention plan template is obtained. If the classification result shows an abnormality during the information processing stage, the enhanced intervention template is selected to obtain the preliminary intervention plan.

[0070] Based on the initial intervention plan, and combined with medication adherence data from the monitoring records, an adjustment strategy is adopted. If adherence is lower than a preset threshold, the frequency of reminders is increased, and the adjusted intervention plan is determined.

[0071] The adjusted intervention plan generates a personalized implementation plan. By combining the latest physiological indicators from the examination results, if the indicators deviate from the normal range, the intensity of the intervention in the implementation plan is adjusted to obtain the final implementation plan.

[0072] Based on the final execution plan, a daily task list is generated. The task execution status is judged by real-time feedback data from the data source, and the execution completion rate data is obtained.

[0073] Based on the execution completion data, the effectiveness of the intervention plan is analyzed. Through information processing, it is determined whether the execution completion rate has reached the preset standard, and the effect evaluation results are obtained.

[0074] In the daily management of patients with chronic diseases, obtaining daily medication data from medication monitoring records is a fundamental step. For example, the system records whether patients take their antihypertensive and hypoglycemic drugs on time each day, and integrates this data with blood pressure and blood sugar levels from regular check-ups to form a comprehensive health dataset. This dataset can comprehensively reflect the correlation between patients' medication use and physiological indicators, providing a reliable basis for subsequent analysis.

[0075] Specifically, for comprehensive health datasets, a pre-established classification model is used to analyze the correspondence between medication monitoring data and examination results. For example, if medication records show that a patient has taken medication on time for 7 consecutive days and their blood pressure has remained stable below 130 / 80 mmHg, the model will classify the health status as normal; conversely, if medication is intermittent and blood glucose levels exceed 8 mmol / L, it will be classified as abnormal. This classification helps to quickly identify potential risks.

[0076] For example, based on the health status classification results, the system obtains the corresponding intervention plan template. If the classification shows abnormalities, the enhanced intervention template is selected first to form a preliminary intervention plan. In one embodiment, for patients with abnormal blood pressure, the preliminary plan may include increased medication dosage monitoring and lifestyle guidance to control the progression of the disease in a timely manner.

[0077] It should be noted that the initial intervention plan is further adjusted based on medication adherence data from the monitoring records. If adherence falls below the 80% threshold, the system will employ an adjustment strategy to increase the frequency of reminders. For example, the original plan of once a day might be adjusted to once in the morning and once in the evening, with the addition of voice prompts, ultimately determining the adjusted intervention plan. This adjustment effectively improves patients' initiative in adhering to medication.

[0078] In one possible implementation approach, a personalized execution plan is generated from the adjusted intervention protocol and further optimized by incorporating physiological indicators from the latest examination results. If blood glucose levels deviate from the normal range, such as fasting blood glucose reaching 9 mmol / L, the intensity of dietary control is appropriately increased, resulting in the final execution plan. This plan is more tailored to the patient's current condition, ensuring the precision of the intervention.

[0079] For example, a daily task list is generated based on the final implementation plan, such as taking medication in the morning and taking a 30-minute walk after meals. Through real-time feedback data, such as app check-in records, the system assesses task execution and obtains completion rate data. If the patient's completion rate reaches over 90% for five consecutive days, it reflects a high level of adherence.

[0080] Specifically, the effectiveness of the intervention program is assessed based on data analysis of completion rates. If the completion rate reaches the preset 85% standard, the effectiveness evaluation result is positive; otherwise, it indicates the need for further optimization of the reminder methods. This evaluation mechanism forms a closed-loop feedback loop, which helps to continuously improve the actual implementation of the intervention program and the level of patient health management. Through multi-faceted data support, such as the cross-validation of medication records and physiological indicators, the entire process ensures the scientific nature and adaptability of the program, thereby significantly improving patient compliance and health outcomes.

[0081] like Figure 7 As shown, the strategy matching module is used to update the dynamic adaptation parameters and determine the final matching strategy if the adjusted intervention plan matches the historical response data.

[0082] The strategy matching module is also configured as follows: Rapid adaptation using a meta-learning framework: inner layer update Outer layer optimization ; in: These are the initial parameters for the meta-learner. In the first The task refers to the updated parameters of the patient's inner layer. The inner learning rate, The outer learning rate, For gradient operators, For the first The loss function for each task. Total number of tasks; Contextual multi-armed slot machine strategy selection: ; in: For a moment The chosen action is the intervention strategy. For a set of actions, For action The regression parameter vector, For a moment Context feature vectors, To explore coefficients, For action The characteristic covariance matrix, superscript Indicates vector transpose, superscript This represents finding the inverse of a matrix. Posterior probability iteration is achieved through Bayesian online updates: ; in: Given historical data The posterior distribution of the parameters after that, For the prior distribution of parameters, For observations given context and parameters The likelihood function is used to update the dynamic adaptation parameters based on the posterior probability, and the final matching strategy is determined.

[0083] This module is configured to perform the following steps: Records related to the intervention plan are obtained from historical response data. Records that match the relationship are filtered out through information processing to obtain a preliminary matching dataset.

[0084] For the initial matching dataset, information processing methods are used to analyze the conditions required for dynamic adaptation. If the matching relationship meets the preset threshold, the adaptation parameters are updated to obtain the updated parameter set.

[0085] Based on the updated parameter set, a configuration scheme related to the final strategy is generated through pre-established logical rules, thus determining the direction of strategy generation.

[0086] Based on the direction of strategy generation, obtain the response data corresponding to the adjustment plan, and determine whether the response data meets the dynamic adaptation conditions through information processing. If it does, generate an adapted strategy framework.

[0087] Based on the appropriate strategy framework and the matching criteria, information processing methods are used to analyze the fit between historical responses and intervention plans to determine the basis for the final strategy implementation.

[0088] By combining the execution criteria of the final strategy with the results of parameter updates, a specific intervention execution path is generated, resulting in an operable strategy plan.

[0089] By extracting records related to the intervention plan from historical response data and filtering them through information processing, a preliminary matching dataset can be formed. This dataset focuses on patients' past feedback on intervention measures, such as adherence records after medication reminders and changes in physiological indicators. For example, in the daily management of patients with chronic diseases, the system first collects historical response data, such as the patient's daily medication reminder confirmation rate and subsequent blood glucose monitoring values. Specifically, when the screening criteria are set as a response time within 24 hours and an indicator improvement exceeding 5%, records that meet the matching criteria are retained to form the preliminary matching dataset. This process ensures the relevance and reliability of the dataset.

[0090] For the initial matching dataset, information processing methods are used to analyze the conditions required for dynamic adaptation. If the matching relationship meets a preset threshold, the adaptation parameters are updated to obtain the updated parameter set. In one embodiment, the analysis focuses on the correlation between response frequency and compliance improvement. For example, if the dataset shows that in more than 80% of the records, the proportion of patients taking medication on time after being reminded increases, then the threshold is determined to be met, and the reminder intensity parameter is updated, such as adjusting from a single push to multiple pushes at different times.

[0091] Based on the updated parameter set, a configuration scheme related to the final strategy is generated through pre-established logical rules, determining the direction of strategy generation. This direction emphasizes personalized adjustments, such as prioritizing strategies that enhance communication rather than simply increasing dose monitoring.

[0092] Based on the direction of strategy generation, response data corresponding to the adjustment plan is acquired. The information processing stage determines whether the response data meets the conditions for dynamic adaptation. If it does, an adapted strategy framework is generated. Specifically, if the response data shows that patients respond more positively to text-based guidance than to voice prompts, multimedia elements are prioritized in the framework to create a more acceptable intervention structure.

[0093] Based on the appropriate strategy framework and the matching criteria, information processing methods are used to analyze the fit between historical responses and intervention plans to determine the basis for the final strategy implementation. For example, when the fit assessment shows that the success rate of responses to dietary guidance combined with exercise reminders is over 85% in the past, this basis is established to guide subsequent implementation priorities.

[0094] By establishing the execution basis of the final strategy and combining it with the results of parameter updates, a specific intervention execution path is generated, resulting in an actionable strategy plan. One possible implementation approach could include sending personalized dietary suggestions in the morning, following up with exercise check-ins at noon, and summarizing physiological indicator feedback in the evening, forming a closed-loop execution. This approach closely aligns with patients' behavioral habits, effectively improving adherence, and achieving dynamic optimization through continuous response matching, thereby significantly improving health management outcomes.

[0095] It's important to note that the entire process, from historical response screening to path generation, forms a tightly linked logical chain, with data from various aspects mutually supporting each other to ensure the adaptability and practicality of the strategy. For example, parameter updates and fit analysis together reinforce the direction determination, while the execution criteria provide a solid foundation for the path, ultimately leading to more precise interventions and higher patient participation.

[0096] like Figure 8 As shown, the push execution module is used to generate a notification signal according to the final matching strategy, push behavioral guidance instructions to the patient device, and obtain execution feedback data.

[0097] The push execution module includes: A signal generation unit is used to generate a notification signal according to the final matching strategy; The instruction conversion unit is used to convert notification signals into behavior guidance instructions according to preset encoding rules; The target determination unit is used to bind instruction content through patient device identification; The instruction sending unit is used to push instructions to the patient device. The feedback acquisition unit is used to acquire execution feedback data; The feedback parsing unit is used to parse the feedback data in a structured manner and extract the confirmation flags; The status determination unit is used to determine the policy trigger status based on the confirmation flag associated with the final matching policy.

[0098] A notification signal is generated based on the final matching strategy. The notification signal is then converted into a behavioral guidance instruction using a preset encoding rule to obtain the instruction content. This instruction content is bound to the patient device identifier to determine the push operation target. The push operation is executed on the patient device to complete the behavioral guidance instruction transmission. Execution feedback data returned by the patient device is obtained. The execution feedback data is then structured and parsed to determine if the feedback status is complete; if complete, a confirmation flag is extracted. The confirmation flag is associated with the final matching strategy to determine the strategy trigger status.

[0099] For example, in the field of chronic disease management, the process of generating notification signals for a final matching strategy can be understood as transforming the strategy into a specific, executable signal. The generation of these notification signals is based on the core content of the strategy. For instance, for a medication reminder strategy for diabetic patients, the system would generate a timed signal, set to be pushed out every morning at 8:00 AM. This signal not only includes the time trigger but also the priority and format of the reminder; for example, a vibration combined with a ringtone might be used for higher priority reminders to ensure the patient doesn't miss it.

[0100] Specifically, notification signals are converted into behavioral guidance instructions using preset encoding rules. This can be achieved by mapping the signal into specific text or voice content. For example, if the signal contains the core information of "medication reminder," the encoding rules will convert it into the specific instruction "Please take two tablets of your blood sugar medication at 8:00 AM." This process ensures the accurate transmission of the signal.

[0101] For example, by binding instruction content to the patient's device identifier and determining the target of the push operation, precise push can be achieved through the device's unique identifier. Assuming the patient's smartwatch identifier is ID12345, the system will bind the instruction content to this identifier, ensuring that the instruction is only sent to the designated device and avoiding mis-sending. The push operation then sends a message to the device via a cloud service, such as executing a push precisely at 8:00 AM via a wireless network.

[0102] Specifically, obtaining the execution feedback data returned by the patient's device can be achieved through the device's built-in response mechanism. For example, if a patient clicks the "Medication Taken" button after receiving a notification, the device will generate feedback data, including the click time (e.g., 8:05) and the confirmation status. This data is then transmitted back to the system server via the network for subsequent analysis.

[0103] For example, in the structured parsing of execution feedback data, the completeness of the feedback status can be determined using predefined field formats. Assuming the feedback data includes a timestamp and a confirmation field, if both have values ​​(e.g., the timestamp is 8:05 and the confirmation field is "completed"), the feedback is considered complete, and the confirmation marker "completed" is extracted as key information.

[0104] Specifically, the strategy triggering status is determined by comparing the confirmation marker with the strategy objective, based on the final matching strategy. For example, if the strategy objective is "ensuring patients take their medication on time," and the confirmation marker shows "completed," then the strategy triggering status is determined to be "successful," and the system records this status for subsequent optimization and adjustments. This process ensures closed-loop management of strategy execution, improving patient adherence.

[0105] In one possible implementation, further analysis of the strategy trigger status can be combined with historical data to determine the strategy's effectiveness. For example, if the system records that the strategy trigger status was "successful" 6 times in the past 7 days, accounting for over 85%, then the current strategy design can be considered reasonable and can be continued or fine-tuned. This analysis provides data support for continuous strategy improvement, contributing to enhanced overall health management effectiveness.

[0106] It also includes a comprehensive evaluation module for intervention effectiveness, configured as follows: Calculate the overall score of intervention effect: in: To assess the overall effectiveness of the intervention, For the rate of improvement of health indicators, To improve medication adherence, For intervention costs or patient burden, The model interpretability score, ranging from 0 to 1, is obtained by weighting indicators such as feature importance consistency and counterfactual plausibility. For the corresponding weight coefficients, satisfying And all coefficients are non-negative; Calculate the dynamic adaptability index: ; in: As a dynamic adaptability indicator, The total number of samples, This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. and These are the model's response to samples before and after the policy update. The predicted probability, The standard deviation of the predicted probabilities before the policy update. The preset sensitivity threshold; if If the value exceeds the preset standard, a model update signal is generated and fed back to the policy matching module.

[0107] A patient health monitoring method integrating behavioral intervention, medication management, and examination is applied to the aforementioned patient health monitoring system integrating behavioral intervention, medication management, and examination, such as... Figure 9 As shown, it includes the following steps: Step S101: Collect patient behavior data, medication use records, and regular check-up results to obtain comprehensive patient data; Step S102: Extract features from the comprehensive patient data and calculate using a random forest classifier. Divergence is used as an anomaly deviation value, and a time-series graph neural network is used for dynamic anomaly detection; Step S103: If the abnormal deviation value exceeds the threshold, deep clustering is performed through contrastive learning and prototype network to obtain personalized intervention categories; Step S104: Extract real-time change features from personalized intervention categories, use multi-objective deep reinforcement learning to simulate intervention effects, and generate an optimized path; Step S105: Integrate drug surveillance data and inspection results through optimized pathways, assess the net benefit of the intervention based on causal inference, and obtain the adjusted intervention plan; Step S106: If the adjusted intervention plan matches the historical response data, then update the dynamic adaptation parameters through meta-learning and contextual multi-armed slot machine to determine the final matching strategy; Step S107: Generate a notification signal based on the final matching strategy, push behavioral guidance instructions to the patient device, and obtain execution feedback data; Step S108: Calculate the comprehensive score and dynamic adaptability index using the intervention effect comprehensive evaluation module, and provide feedback on optimization strategies based on the evaluation results.

[0108] This invention integrates patient behavior data, medication records, and regular check-up results into comprehensive data through a data acquisition module; an anomaly analysis module uses a random forest algorithm to classify behavioral patterns and health indicators, identifying abnormal deviations; an intervention classification module generates personalized intervention categories through cluster analysis when thresholds are exceeded; a path optimization module combines a deep reinforcement learning model to simulate intervention effects and optimize the adjustment plan; and a strategy matching module updates parameters and pushes guidance instructions based on historical response data. Addressing the core issue of achieving precise intervention in dynamically changing health management, this invention innovatively integrates behavioral monitoring, data analysis, and real-time feedback mechanisms. Through the collaborative work of multiple algorithms, it ensures the personalization and adaptability of intervention plans, significantly improving the efficiency and effectiveness of health management. It particularly demonstrates unique value in chronic disease management and behavioral guidance scenarios, providing patients with scientific and timely health support.

[0109] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A patient health monitoring system integrating behavioral intervention, medication management, and examination, characterized in that: Specifically, it includes: The data acquisition module is used to collect patient behavior data, medication records, and regular check-up results through sensors and wearable devices to obtain comprehensive patient data; The anomaly analysis module is used to extract behavioral and health indicator features from comprehensive patient data, and calculates them using a random forest classifier. The divergence value is used as an anomaly deviation value, and a time-series graph neural network is further used to perform dynamic anomaly detection on multi-source heterogeneous data; The intervention classification module is used to obtain personalized intervention categories by performing deep clustering of behavioral patterns and health indicators through contrastive learning and prototype networks if the abnormal deviation value exceeds the preset threshold. The path optimization module is used to extract real-time change features from personalized intervention categories, simulate the intervention effect using a multi-objective deep reinforcement learning model, and generate an optimized path. The intervention adjustment module is used to integrate drug monitoring data and periodic inspection results through optimized pathways, and to evaluate the net benefit of the intervention based on causal inference and counterfactual reasoning, so as to obtain the adjusted intervention plan. The strategy matching module is used to determine the final matching strategy by updating the dynamic adaptation parameters through meta-learning and contextual multi-armed slot machine if the adjusted intervention plan matches the historical response data. The push execution module is used to generate notification signals based on the final matching strategy, push behavioral guidance instructions to the patient's device, and obtain execution feedback data.

2. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, The data acquisition module includes: The real-time acquisition unit is used to collect patient behavior data, medication records, and examination results in real time. The data integration unit is used for unified formatting and integration into structured comprehensive data; The data cleaning unit is used to remove outliers and missing values; Data imputation unit, used to fill in missing values ​​whose missing proportion exceeds a threshold by historical interpolation; Anomaly classification unit is used to classify abnormal behavior or medication patterns using support vector machines; The correlation analysis unit is used to analyze the correlation between behavior and examination results, and to obtain the correspondence between behavior and health status. The strategy generation unit is used to generate personalized monitoring strategies and adjustment suggestions based on the corresponding relationships.

3. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, The anomaly analysis module includes: The classification processing unit is used to classify behavioral patterns and health indicators based on comprehensive patient data using the random forest algorithm, and to determine abnormal deviation values. The grouping unit is used to group patient groups based on abnormal deviation values ​​using cluster analysis to obtain behavioral pattern groups; The rule extraction unit is used to extract key behavioral rules for each group based on behavioral patterns using a decision tree algorithm, thereby obtaining a rule set. The risk determination unit is used to perform correlation matching between the rule set and health indicator deviations to determine whether there are significant deviation correspondences and identify high-risk behavior patterns. The threshold adjustment unit is used to generate personalized early warning threshold data based on high-risk behavior patterns to obtain the adjusted monitoring threshold. The risk marking unit is used to compare and judge the subsequent collected data in real time according to the adjusted monitoring threshold. If the threshold is exceeded, the risk status is marked and a risk warning signal is determined. The level determination unit is used to conduct trend comparison analysis on risk warning signals in combination with historical behavior patterns to determine whether there is a continuous deviation and obtain the enhanced warning level. The anomaly analysis module is also configured as follows: Constructing a temporal heterogeneous graph ; in, For time indexing, A set of nodes, containing behavior nodes. Medication timing and physiological nodes , For a moment The edge set represents the dynamic dependencies between nodes; Update node features using graph attention mechanism: ； in: For nodes At any moment The updated feature vector, For activation function, For nodes The set of neighboring nodes, For nodes To the neighbors Attention weights The weight matrix is ​​a learnable matrix. For nodes At any moment The original feature vector, For attention parameter vectors, This represents vector concatenation. It is a non-linear activation function; Modeling temporal dependencies using temporal convolutional networks: ; in: For a moment Embedment matrix of all nodes, The time window length, Indicates temporal convolutional network operations; Calculate the outlier score: ; The first term represents the graph reconstruction error. For decoder, For the encoder's implicit variable output, For the Euclidean norm; in the second term For balance coefficient, express divergence, This represents the temporal distribution at the current moment. For historical time series distribution; if If the value exceeds the preset threshold, it is considered abnormal, and the abnormal score is used as the abnormal deviation value.

4. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, The intervention classification module is also configured as follows: The embedding representation is optimized through contrastive learning, and the contrastive loss function is: ; in: The total number of samples, For the first The embedding vector of each sample, For the first The positive sample embedding vector of each sample. For the first Each negative sample embedding vector The cosine similarity function is used. Temperature coefficient; Build a prototype network for each intervention category. The prototype vector is: ; in: To belong to category The sample set, The number of samples in this set. For the sample Category tags; The classification probability of the new sample is: ; in: Represents the sample embedding vector Category The probability, For Euclidean distance, The total number of intervention categories; individualized intervention categories are determined based on the highest probability.

5. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, The path optimization module is also configured as follows: Define a multi-objective reward vector: ; in: For a moment The reward vector, For the rate of improvement of health indicators, To improve medication adherence, For intervention costs or patient burden; Multi-objective Q-learning update: For each objective , ; in: For a moment In state Take action below Time corresponding target Action value function, For learning rate, For the goal Instant rewards As a discount factor, For the next state, For the next action; Define the advantage vector: ; in: As the dominant vector, , This is a state value vector; Strategy selection criteria: ; in: In the state The strategy of selection For the patient preference weight vector, Corresponding target The weights are determined based on the selected strategy to determine the optimization path.

6. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, The scheme adjustment module is also configured as follows: Using dual machine learning to estimate the causal effects of the intervention: in: For the total sample size, For sample index, For the first Intervention indicator variables for each sample ( Indicates acceptance of intervention. (This indicates that the offer was not accepted). For covariate vectors, For the propensity score estimation function (i.e., given covariates) (Probability of accepting intervention) The estimated function for the resulting model, For observed health outcomes, and These represent the predicted outcomes under intervention and non-intervention conditions, respectively. Calculate the net benefit of individual interventions: ; in: For the first Net intervention benefit for each individual To accept the potential outcomes of the intervention, For potential outcomes without intervention, Given covariates and intervention status The conditional expectation function is used; if the net benefit of an individual intervention is positive, the intervention is deemed effective, and an adjusted intervention plan is generated based on the characteristics of an effective intervention.

7. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, The strategy matching module is also configured as follows: Rapid adaptation using a meta-learning framework: inner layer update Outer layer optimization ; in: These are the initial parameters for the meta-learner. In the first The task refers to the updated parameters of the patient's inner layer. The inner learning rate, The outer learning rate, For gradient operators, For the first The loss function for each task Total number of tasks; Contextual multi-armed slot machine strategy selection: ; in: For a moment The chosen action is the intervention strategy. For a set of actions, For action The regression parameter vector, For a moment Context feature vectors, To explore the coefficient, For action The characteristic covariance matrix, superscript Indicates vector transpose, superscript This represents finding the inverse of a matrix. Posterior probability iteration is achieved through Bayesian online updates: ; in: Given historical data The posterior distribution of the parameters after that, For the prior distribution of parameters, For observations given context and parameters The likelihood function is used to update the dynamic adaptation parameters based on the posterior probability, and the final matching strategy is determined.

8. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, The push execution module includes: A signal generation unit is used to generate a notification signal according to the final matching strategy; The instruction conversion unit is used to convert notification signals into behavior guidance instructions according to preset encoding rules; The target determination unit is used to bind instruction content through patient device identification; The instruction sending unit is used to push instructions to the patient device. The feedback acquisition unit is used to acquire execution feedback data; The feedback parsing unit is used to parse the feedback data in a structured manner and extract the confirmation flags; The status determination unit is used to determine the policy trigger status based on the confirmation flag associated with the final matching policy.

9. The patient health monitoring system integrating behavioral intervention, medication management, and examination as described in claim 1, characterized in that, It also includes a comprehensive evaluation module for intervention effectiveness, configured as follows: Calculate the overall score of intervention effect: in: To assess the overall effectiveness of the intervention, For the rate of improvement of health indicators, To improve medication adherence, For intervention costs or patient burden, The model interpretability score, ranging from 0 to 1, is obtained by weighting indicators such as feature importance consistency and counterfactual plausibility. For the corresponding weight coefficients, satisfying And all coefficients are non-negative; Calculate the dynamic adaptability index: ; in: As a dynamic adaptability indicator, The total number of samples, This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. and These are the model's response to samples before and after the policy update. The predicted probability, The standard deviation of the predicted probabilities before the policy update. The preset sensitivity threshold; if If the value exceeds the preset standard, a model update signal is generated and fed back to the policy matching module.

10. A patient health monitoring method integrating behavioral intervention, medication management, and examination, applied to the patient health monitoring system integrating behavioral intervention, medication management, and examination as described in any one of claims 1 to 9, characterized in that, Includes the following steps: Step S101: Collect patient behavior data, medication use records, and regular check-up results to obtain comprehensive patient data; Step S102: Extract features from comprehensive patient data, calculate KL divergence as anomaly bias value using a random forest classifier, and use a temporal graph neural network for dynamic anomaly detection; Step S103: If the abnormal deviation value exceeds the threshold, deep clustering is performed through contrastive learning and prototype network to obtain personalized intervention categories; Step S104: Extract real-time change features from personalized intervention categories, use multi-objective deep reinforcement learning to simulate intervention effects, and generate an optimized path; Step S105: Integrate drug surveillance data and inspection results through optimized pathways, assess the net benefit of the intervention based on causal inference, and obtain the adjusted intervention plan; Step S106: If the adjusted intervention plan matches the historical response data, then update the dynamic adaptation parameters through meta-learning and contextual multi-armed slot machine to determine the final matching strategy; Step S107: Generate a notification signal based on the final matching strategy, push behavioral guidance instructions to the patient device, and obtain execution feedback data; Step S108: Calculate the comprehensive score and dynamic adaptability index using the intervention effect comprehensive evaluation module, and provide feedback on optimization strategies based on the evaluation results.

Images