Health management method and system based on multi-dimensional data analysis

Abstract

The invention relates to the technical field of artificial intelligence and health management, in particular to a health management method and system based on multi-dimensional data analysis, and the method comprises the steps: collecting the physiological, behavior, environment and psychological data of a user, and carrying out the modeling through a recurrent neural network, and generating health state data; extracting a key causal relationship of the health risk based on a causal analysis model of a Bayesian network; comprehensively analyzing health changes through the short-term time sequence model and the long-term trend prediction model, and generating health risk data; a personalized health intervention scheme is designed by using a conditional generative adversarial network, and safe sharing and dynamic management of the scheme are realized through a block chain technology. According to the method, the health risk can be accurately identified, scientific intervention measures can be designed, the health management process is dynamically optimized, powerful technical support is provided for an intelligent old-age nursing scene, and the health management efficiency and effect are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and health management technology, and in particular to a health management method and system based on multidimensional data analysis. Background Art

[0002] With the intensification of the problem of population aging, the importance of health management for the elderly has become increasingly prominent. Health management methods based on multidimensional data analysis can provide comprehensive assessment and precise intervention of health status by integrating multimodal data such as physiological, behavioral, psychological and environmental data of users. Especially in the elderly care scenario, through real-time monitoring and analysis of the health data of the elderly, potential health risks can be identified and personalized intervention plans can be designed, thereby extending healthy life expectancy and improving quality of life. However, existing technologies have obvious deficiencies in health status analysis, and more scientific and comprehensive analysis methods are urgently needed.

[0003] The prior art (Chinese invention patent, publication number: CN114496264B, name: Health index analysis method, device, equipment and medium based on multidimensional data) usually relies on analysis methods based on a single or small amount of health data, such as simple alarm prompts based on the user's current heart rate or blood pressure data. The method proposed in the prior art uses the medical history data and monitoring data of the user group to perform time series analysis and trend fitting to calculate the comprehensive health index. Although this method introduces multidimensional data analysis, it is mainly limited to the analysis of disease causes and potential risks, and does not provide sufficient support for individualized real-time health management. Specific defects include: Existing technologies rely more on historical data and trend analysis, lack the ability to dynamically adjust the user's current status, and find it difficult to respond to sudden health risks in a timely manner; they fail to deeply explore the complex relationships between multimodal health data, for example, the indirect impact of the environment on the psychological state is not fully reflected; the core of existing technologies is the calculation of health indexes, but lack an effective health intervention generation and optimization mechanism, making it difficult to dynamically manage individual health status. Summary of the invention

[0004] In view of the many problems existing in the above-mentioned prior art, the present invention provides a health management method and system based on multidimensional data analysis. The present invention realizes the accurate expression of health status through multimodal data fusion and recursive neural network modeling; combines the causal analysis model of Bayesian network to dynamically mine the source of health risks; generates personalized health intervention plans through conditional generative adversarial networks, and uses blockchain technology to ensure the safe sharing and real-time update of the plans. The present invention effectively improves the accuracy of health risk assessment and the scientific nature of health intervention, and significantly improves the health management effect of the elderly in smart elderly care scenarios.

[0005] A health management method based on multidimensional data analysis comprises the following steps: Collecting the user's physiological data, behavioral data, environmental data and psychological data, preprocessing the collected data to generate preprocessed health data; performing anomaly analysis on the health data using an anomaly detection model based on time series analysis to generate anomaly labeled data; Based on the pre-processed health data and abnormal marked data, extract the user's health characteristics and generate health characteristic data; combine the health characteristic data to generate health status data; analyze the health status data through the causal analysis model based on the Bayesian network, extract the causal characteristics of health changes, and generate causal characteristic data; Based on health status data and causal characteristic data, health risk analysis is carried out, using short-term time series models to capture the rapid fluctuation trend of data, and combined with long-term trend prediction models to analyze the long-term trend of health changes to generate health risk data; based on health risk data, health intervention plans are generated; Collect users' intervention execution feedback data, optimize the health intervention plan based on the feedback data, and generate an optimized health intervention plan; share the optimized health intervention plan to relevant terminal devices to achieve closed-loop optimization of health management.

[0006] Preferably, the anomaly detection model is a sliding window model based on time series analysis, which performs local anomaly detection on the preprocessed health data and generates abnormal marker data by setting the size of the sliding window and dynamically adjusting the threshold.

[0007] Preferably, the health status data is generated by combining dynamic health feature coding data and abnormal marker data, and health status modeling is performed using a recursive neural network, which takes time series data as input and outputs health status data.

[0008] Preferably, the extraction of the dynamic health feature encoding data is achieved through a dynamic attention mechanism, which assigns weights based on the importance of time series features, where the weights The calculation of includes the following steps: Calculate time points The feature importance score of , the feature importance score is calculated by It is obtained by weighted summing of the eigenvalue of and the eigenvalues ​​of adjacent time points; Calculate time points The feature weight Specifically: , in, For time point The corresponding feature weights; For time point The feature importance score is calculated by the trend of the feature value change of the time series; is the index of the feature in the time series; is the total length of the time series; represents the natural exponential function; According to the feature weight at each time point And the corresponding eigenvalues, generate dynamic health feature encoding data.

[0009] Preferably, the extraction of the causal features is achieved through a causal analysis model based on a Bayesian network, where the causal strength The calculation of includes the following steps: Constructing a causal graph of health status data, wherein the causal graph includes variable nodes in the health status data and causal relationships between the nodes; For Node and nodes Causal strength The calculation formula for causal strength is: , in, For Node and nodes the causal strength between them; For the node After the intervention node The conditional probability distribution of ; For Node Probability distribution in the non-intervention state; Based on the causal strength calculation results, variable relationships with high causal strength are extracted to generate causal feature chain data, where the causal feature chain is used to describe the key causal relationships between variables in the health status data.

[0010] Preferably, the health risk analysis comprises the following steps: Using a short-term time series model to capture the rapid fluctuation trend of health feature data, the short-term time series model is a convolutional neural network; Analyzing the long-term trend of health changes based on a long-term trend prediction model, wherein the long-term trend prediction model is a long short-term memory network; Health risk data are generated by integrating the rapid fluctuation trend of the short-term time series model and the trend analysis results of the long-term trend prediction model.

[0011] Preferably, the generation of the health intervention plan is achieved by a conditional generative adversarial network, which includes a generator and a discriminator, wherein: The generator generates multiple health intervention candidates based on the health risk data; The discriminator evaluates the effectiveness and pertinence of candidate health intervention plans and selects the optimal health intervention plan as the final output.

[0012] Preferably, the intervention execution feedback data includes the user's physiological feedback data, behavior execution data, environmental response data and psychological feedback data, wherein: Physiological feedback data includes real-time physiological indicators of heart rate, blood pressure and blood oxygen saturation; Behavioral execution data include users’ completion of intervention tasks; Environmental response data include dynamic changes in indoor humidity, air quality, and noise levels; Psychological feedback data includes emotional change characteristics obtained through voice sentiment analysis.

[0013] Preferably, the optimized health intervention plan is shared to relevant terminal devices through blockchain technology, and the blockchain technology includes building a smart contract to dynamically manage access rights to the health intervention plan.

[0014] A system for implementing the health management method based on multidimensional data analysis, comprising: A data processing module, used to collect physiological data, behavioral data, environmental data and psychological data of the user, pre-process the collected data, and perform anomaly analysis on the pre-processed health data using an anomaly detection model based on time series analysis to generate anomaly labeled data; The health status modeling module is used to extract the user's health characteristics based on the preprocessed health data and abnormal marker data, generate health characteristic data, combine the health characteristic data to generate health status data, and analyze the health status data through a causal analysis model based on a Bayesian network to extract the causal characteristics of health changes and generate causal characteristic data; The health risk analysis module is used to capture the rapid fluctuation trend of data based on health status data and causal characteristic data using a short-term time series model, and analyze the long-term trend of health changes in combination with a long-term trend prediction model, generate health risk data, and generate health intervention plans based on the health risk data; The health intervention optimization module is used to collect users' intervention execution feedback data, optimize the health intervention plan based on the feedback data, generate an optimized health intervention plan, and share the optimized health intervention plan to relevant terminal devices to achieve closed-loop optimization of health management.

[0015] Compared with the prior art, the advantages and beneficial effects of the present invention are: The present invention uses multimodal health data fusion technology to achieve comprehensive modeling of the user's health status, dynamically pay attention to the short-term fluctuations and long-term trends of the user's health status, and significantly improve the accuracy of risk identification; The present invention reveals the causal relationship between variables in health status data through causal analysis technology based on Bayesian networks, provides a scientific explanation of the source of health risks, and provides a more targeted basis for intervention design; The present invention uses conditional generative adversarial network technology to dynamically generate personalized health intervention plans, which can provide users with accurate intervention measures and improve intervention effects based on real-time assessment of health risks; The present invention uses blockchain technology to achieve secure sharing and dynamic management of health intervention plans, protect user privacy, promote multi-party collaboration, and improve the transparency and efficiency of health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the health status modeling process in the present invention; Figure 3 It is a schematic diagram of the cause-effect analysis and health risk assessment process in the present invention; Figure 4 A schematic diagram of the process of generating and sharing the health intervention plan in the present invention; Figure 5 It is a structural block diagram of the system of the present invention. DETAILED DESCRIPTION

[0017] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure.

[0018] like Figure 1 As shown, a health management method based on multidimensional data analysis includes the following steps: like Figure 2 As shown, the physiological data, behavioral data, environmental data and psychological data of the user are collected, and the collected data are preprocessed to generate preprocessed health data; the health data is analyzed for abnormalities using an abnormality detection model based on time series analysis to generate abnormality marked data; Preferably, the anomaly detection model is a sliding window model based on time series analysis, which performs local anomaly detection on the preprocessed health data and generates abnormal marker data by setting the size of the sliding window and dynamically adjusting the threshold.

[0019] In the present invention, an anomaly detection model based on time series analysis is used for anomaly detection of the user's physiological data, behavioral data, environmental data, and psychological data. The model analyzes the time series characteristics of multimodal data to identify abnormal data points that may reflect health risks. The anomaly detection model in the present invention is particularly suitable for dynamic data environments, such as user's heart rate fluctuations, behavioral trajectory changes, instantaneous changes in indoor environmental noise, and mood fluctuations. On this basis, abnormal labeling data is generated to provide accurate input for subsequent health feature extraction and risk analysis.

[0020] The anomaly detection model based on time series analysis uses sliding window technology, which limits the local range of time series data by setting the window size, and combines it with the dynamic threshold adjustment strategy to achieve efficient anomaly detection. The principle of sliding window is to divide the time series data into multiple continuous window intervals, calculate the distribution characteristics of specific indicators in each window, compare these characteristics with the dynamic threshold, and determine whether the data point is an outlier. The dynamic threshold is dynamically updated based on the historical distribution of health data, which can effectively adapt to the changes in the health status of individual users.

[0021] In the present invention, the implementation of the anomaly detection model includes the following specific steps: The collected multimodal health data (such as heart rate, behavior trajectory, humidity data, and mood swing data) is cleaned and standardized to remove invalid data points and ensure the continuity of the time series. For example, for heart rate data, single point anomalies caused by device noise are eliminated.

[0022] Set the size of the sliding window , continuous time series data Split into multiple windows, each containing For example, assuming the sampling frequency is 1 second and the window size is If it is set to 30, each window contains 30 seconds of data.

[0023] In each sliding window, the statistical characteristics of the health data (such as mean, standard deviation, maximum, minimum, etc.) are calculated. These characteristics are used to evaluate the overall distribution of the data in the window. For example, for heart rate data, the mean value in the window is calculated. and standard deviation .

[0024] The detection threshold is dynamically adjusted based on the distribution characteristics of historical data. The threshold setting formula is: , in, is the dynamic threshold; is the mean of historical data; is the standard deviation of historical data; It is an adjustment factor, set according to the sensitivity of the detection.

[0025] For each current data point in the window , to determine the degree of deviation from the threshold: Judgment conditions: ,If the degree of deviation exceeds the dynamic threshold, the data point is marked as an outlier.,The detected outliers are marked as outliers in the original time series,,generating anomaly labeled data to reflect the abnormal features in the,user's health data.

[0026] Dynamic thresholds can be adjusted according to the distribution characteristics of user historical data to adapt to individual differences and dynamic changes in health status. For example, under the normal condition of lower heart rate at night for the elderly, the model will dynamically reduce the sensitivity of anomaly detection to avoid false positives. The sliding window technology reduces the global calculation requirements for the entire time series, controls the complexity within the local range, and achieves rapid detection. By combining local statistical features and dynamic thresholds, it can effectively identify small abnormal fluctuations, such as sudden high heart rate, abnormal interruption of behavioral trajectory, and rapid changes in ambient humidity. The model can process multimodal health data simultaneously and generate unified abnormal labeling data, providing a reliable basis for subsequent health feature extraction and health risk analysis.

[0027] In the elderly care management, the anomaly detection model is used to monitor the daily health status of the elderly and identify abnormal conditions that may indicate health risks. For example, an elderly person lives alone in a smart elderly care community. The wearable device he wears monitors his heart rate in real time. Environmental sensors in the community collect humidity and noise data. At the same time, the smart home system records the daily behavior of the elderly. The data is uploaded to the health management platform through the system.

[0028] Collect multimodal data, including heart rate data (physiological data), movement trajectory data (behavioral data), indoor humidity data (environmental data), and voice emotion data (psychological data). Preprocess the collected data, such as removing device data noise and interpolating time series to generate continuous health data. Apply sliding window technology (window size is 30 seconds) to perform local analysis on heart rate data and calculate the mean and standard deviation within the window. Dynamically set thresholds based on historical data of the elderly, such as dynamically adjusting The system generates abnormal marking data when it detects that the heart rate of an elderly person at night increases from 55 beats per minute to 120 beats per minute in a short period of time and the behavior data records a sudden stop in the movement trajectory.

[0029] The system identifies in real time the risk of heart problems or falls among the elderly, and triggers an emergency notification mechanism for community caregivers through abnormally marked data, allowing for rapid intervention to avoid further deterioration of health risks.

[0030] Based on the pre-processed health data and abnormal marked data, extract the user's health characteristics and generate health characteristic data; combine the health characteristic data to generate health status data; analyze the health status data through the causal analysis model based on the Bayesian network, extract the causal characteristics of health changes, and generate causal characteristic data; The extraction of health features is centered on time series analysis and multimodal data fusion, and the key features of users are extracted from health data through dynamic modeling. Health status data is the aggregated result of dynamic health features, reflecting the overall health status of users in multiple dimensions. In the causal analysis stage, the causal analysis model based on Bayesian network is used to construct the causal graph of variables in health status data, and determine the causal relationship and its strength between variables through causal reasoning.

[0031] Preprocessed health data and abnormal labeled data are used as input, including the user's physiological data (such as heart rate, blood pressure), behavioral data (such as movement trajectory, sleep duration), environmental data (such as humidity, air quality) and psychological data (such as mood swings). Abnormal labeled data is used to highlight possible key data points.

[0032] Based on the dynamic change law of time series, each type of health data is segmented and analyzed to extract time-dependent features (such as change rate, periodic features, etc.). Through the feature fusion algorithm, different types of data are uniformly mapped into a multidimensional feature space to construct a multimodal feature expression. For example, the relationship between heart rate fluctuations and sleep time is mapped to a common feature vector space to generate comprehensive features that reflect the user's physiological-behavioral correlation. Using the dynamic attention mechanism, higher weights are automatically assigned to key features in health data. For example, sudden high heart rate data points in abnormally labeled data will be given higher weights to ensure the accuracy of feature expression.

[0033] Combined with dynamic health features, health status data is constructed through recursive neural networks. Recursive neural networks capture long-term trends and short-term fluctuations in user health data by learning the dynamic changes of time series, and aggregate health features into health status data. Health status data is represented in the form of multidimensional vectors, including the user's physiological state (such as normal, abnormal, critical), behavioral state (such as active, static), and environmental adaptability (such as suitable, unsuitable).

[0034] The variables in the health status data are modeled through the Bayesian network to construct a causal graph. The causal graph uses nodes to represent variables (such as heart rate, sleep duration, and emotional state) and directed edges to represent the causal relationship between variables. For example, low humidity may cause sleep disruption, which in turn causes heart rate fluctuations. Using the causal reasoning algorithm, the key causal relationships and causal strengths between variables are extracted from the causal graph. By simulating intervention operations (such as assuming an increase in humidity), the changing trend of the health status is analyzed to further confirm the directionality and credibility of the causal relationship. Variable relationships with high causal strength are extracted, and causal feature data are generated. Causal feature data is used to describe the most critical variable associations in the health status and their influencing mechanisms, providing a basis for subsequent health risk analysis and intervention optimization.

[0035] Through the dynamic attention mechanism and multimodal fusion algorithm, it is possible to automatically focus on key data features, reduce the interference of irrelevant data, and generate more representative health feature data. As a comprehensive expression of dynamic features, health status data not only reflects the user's current health status, but also reveals the interactive relationship between data of various dimensions, laying the foundation for the analysis of health change trends. Causal graphs and causal reasoning algorithms can reveal the hidden causal relationships in health data, and provide a scientific decision-making basis for the prediction and intervention of health risks. The present invention can update health characteristics and causal relationships in real time, adapt to changes in the user's health status, and provide dynamic health management support.

[0036] Embodiment: In a retirement community, the present invention is used to analyze the health characteristics and causal relationships of elderly users and provide support for personalized health management. In a retirement community, the health status of elderly user A needs to be monitored in real time. Smart devices in the community collect multimodal health data of users, including heart rate data (physiological data), daily activity (behavioral data), indoor humidity (environmental data), and mood swings (psychological data). After preprocessing, continuous time series health data is generated and possible abnormal points are marked.

[0037] Based on time series analysis, the daytime fluctuation rate and nighttime stability of heart rate data are extracted, combined with the regularity of daily activity, to generate dynamic health characteristics of users. Abnormally marked humidity change data is given a higher weight to emphasize the potential impact of the environment on health.

[0038] Through recursive neural network modeling, the above features are aggregated into health status data, showing that the user’s current status is “large heart rate fluctuations, low activity level, and low ambient humidity.”

[0039] A causal graph was constructed to show the causal relationship of "low humidity → interrupted sleep → heart rate fluctuations". The intervention operation of "increasing humidity" was simulated through causal reasoning, and the results showed that the risk of heart rate fluctuations could be reduced by 40%.

[0040] Extract key causal features and record the strong causal relationship between the variables "humidity", "sleep" and "heart rate" to provide a scientific basis for subsequent health interventions.

[0041] Through this analysis process, the retirement community determined that the main risk to the user's current health status came from insufficient environmental humidity, and gave priority to adjusting humidity in subsequent health interventions. Community staff designed a set of personalized intervention plans based on causal characteristic data, which improved management efficiency and reduced health risks.

[0042] Preferably, the health status data is generated by combining dynamic health feature coding data and abnormal marker data, and health status modeling is performed using a recursive neural network, which takes time series data as input and outputs health status data.

[0043] In this invention, in order to accurately express the user's health status, dynamic health feature encoding data is combined with abnormal marking data, and health status modeling is performed through a recurrent neural network (RNN) to generate health status data that can reflect health change trends and risk points. Due to its advantages in processing time series data, recurrent neural networks can capture long-term dependencies and short-term changes in dynamic feature data, thereby modeling the user's health status with high precision.

[0044] The dynamic health feature encoding data is extracted through time series analysis and feature fusion algorithms, reflecting the main features of users in multimodal health data, while the abnormal marking data is used to mark important points in the feature data (such as abnormal fluctuations or possible risk points), providing key references for the modeling process. During the modeling process, the recurrent neural network gradually processes the features of the time series through its recurrent units (such as LSTM or GRU), taking into account the impact of the previous features on the current state, thereby generating health status data containing health trends and real-time status.

[0045] Dynamic health feature encoding data and abnormality labeling data are the inputs of this modeling process: Dynamic health feature encoding data, which contains a multidimensional vector of user health features, is generated based on time series analysis and feature fusion. For example, the heart rate feature encoding data of the elderly may include the daytime fluctuation amplitude of the heart rate, the nighttime average value, etc.

[0046] Abnormally labeled data, marking key points in health feature data that may have risks. For example, the time point when the ambient humidity is too low is marked as an abnormal point.

[0047] The two types of data are arranged in chronological order and then input into the modeling module of the recurrent neural network.

[0048] The specific process of health status modeling is as follows: (1) Structure of recurrent neural network: Long short-term memory network (LSTM) is selected as the core unit of recurrent neural network. LSTM can effectively alleviate the gradient vanishing problem of traditional RNN in long time series. Input time series data , where each Indicates at a point in time The recurrent unit gradually processes the time series data through the gating mechanism (input gate, forget gate, output gate) and updates the hidden state , to learn the dynamic changes of feature data.

[0049] (2) Core calculation of state modeling: at each time point The hidden state Determined by the unit update function of the recurrent neural network: , in, For time point The hidden state of is the hidden state at the previous time point; For time point Input vector (dynamic health feature encoding data and abnormal label data); are the parameters of the network, including the weights of the gating mechanism; Represents the update rule of a recurrent unit, such as the gating operation of an LSTM.

[0050] (3) Generation of health status data: hidden state The output is mapped to a multidimensional state vector , indicating that the user is at a point in time The mapping process is completed through the fully connected layer: , in, Health status data; and are the weights and biases of the fully connected layer.

[0051] (4) Output complete health status data: Convert the time series The combination forms health status data that describes the dynamic changes in the user's health status. This data can reflect the interactive relationship between multi-dimensional health characteristics and the trend over time.

[0052] By learning the dynamic changes of time series, recurrent neural networks can capture the long-term trends and short-term fluctuations of users' health characteristics. For example, it can detect the long-term deterioration trend and short-term drastic changes in the heart rate fluctuations of the elderly. The combination of dynamic health feature encoding data and abnormal marking data enables health status data to reflect the complex interactive relationship between multimodal data. For example, low indoor humidity causes poor sleep quality, which further leads to heart rate fluctuations. By giving higher weights to key points through abnormal marking data, the modeling process is more sensitive to abnormal changes in key health characteristics, thereby improving the accuracy of predicting health risks. The health status modeling module can update health status data in real time according to different data sources of users (such as wearable devices, environmental sensors), adapting to the needs of different scenarios.

[0053] In the embodiment, elderly user B in a retirement community is monitored for multimodal health data, including heart rate data (physiological characteristics), movement trajectory (behavioral characteristics), indoor air humidity (environmental characteristics), and mood swings (psychological characteristics). Among them, abnormal marker data of the elderly user's heart rate and indoor humidity indicate an increase in recent health risks.

[0054] Extract key health features of elderly users from time series data. For example, the heart rate fluctuation is 3 times / minute, and the number of steps per day in the action trajectory is less than 3,000. Mark that the indoor humidity has been below 40% 5 times in the past two weeks, and the heart rate has suddenly increased to 100 times / minute 3 times. Input the dynamic health feature encoding data and the abnormal marking data into the recurrent neural network, and learn the dynamic interaction between the data through the LSTM unit. The health status data generated by the model shows that the user's current state is "fatigue state, high risk of insufficient sleep". The output health status data shows the dynamic relationship of "humidity decreases → sleep quality decreases → heart rate fluctuations intensify", prompting nursing staff in the elderly community to adjust the indoor environment and pay attention to the user's nighttime heart rate fluctuations.

[0055] Through recursive neural network modeling, retirement communities can accurately understand the current health status and potential risks of elderly users, and thus design more scientific and personalized health intervention measures. For example, they can improve indoor humidity to 40%-60% and monitor users' nighttime heart rate and breathing rate through sleep tracking devices.

[0056] Preferably, the extraction of the dynamic health feature encoding data is achieved through a dynamic attention mechanism, which assigns weights based on the importance of time series features, where the weights The calculation of includes the following steps: Calculate time points The feature importance score of , the feature importance score is calculated by It is obtained by weighted summing of the eigenvalue of and the eigenvalues ​​of adjacent time points; Calculate time points The feature weight Specifically: , in, For time point The corresponding feature weights; For time point The feature importance score is calculated by the trend of the feature value change of the time series; is the index of the feature in the time series; is the total length of the time series; represents the natural exponential function; According to the feature weight at each time point And the corresponding eigenvalues, generate dynamic health feature encoding data.

[0057] In the present invention, in order to extract more accurate and dynamic health features from multimodal health data, a dynamic attention mechanism is used to assign weights to the importance of time series features and generate dynamic health feature encoding data. The dynamic attention mechanism is a feature weight assignment method based on time series, which can automatically focus on the most critical time point features in health data, thereby more accurately reflecting the changes in the user's health status.

[0058] The dynamic attention mechanism assigns importance scores (i.e., feature weights) to the features at each time point in the time series, and generates the final dynamic health feature encoding data based on these weights. The weight calculation of the attention mechanism is based on the trend of feature value changes in the time series, and the feature importance scores are calculated by weighted summation. The soft attention mechanism is used to normalize the weight distribution, making the total weight distribution in the time series more reasonable.

[0059] The extraction of dynamic health feature coding data takes time series data as input, and the data sources include the user's physiological data (such as heart rate, blood pressure), behavioral data (such as movement trajectory), environmental data (such as humidity, temperature) and psychological data (such as voice emotion characteristics). These data are preprocessed to form a continuous time series.

[0060] The implementation steps of the dynamic attention mechanism include the following core links: (1) Feature importance score calculation: feature importance score at each time point Based on the time point feature value And the weighted sum of the eigenvalues ​​of its adjacent time points is calculated as: , in, Indicates time point The characteristic value of and Respectively represent the characteristic values ​​of the adjacent time points before and after it.

[0061] (2) Feature weight allocation: Calculate at each time point The feature weight , the specific formula is: , This formula constrains the weights of all time points to be between 0 and 1 through normalization, ensuring that the weight distribution conforms to the characteristics of the time series.

[0062] (3) Feature encoding data generation: based on the calculated feature weights , combined with the corresponding feature value, to generate dynamic health feature coding data. The generation method is: , This calculation performs a weighted summation of all feature values ​​within the time series to generate a feature-encoded data that comprehensively reflects the dynamic changes in the time series features.

[0063] The generation of dynamic health feature coding data is widely used in subsequent health status modeling and risk analysis. For example, feature points with higher weights will be given higher influence in the modeling process, thereby enhancing the sensitivity of modeling results to key health factors.

[0064] The dynamic attention mechanism can effectively focus on the most critical time points in the time series, ignore irrelevant or minor feature points, and make the generated feature encoding data more focused on the core area of ​​health risks. Because the weight allocation process is dynamically adjusted, the dynamic attention mechanism can adapt to the time series characteristics of different users and different health states. For example, the heart rate data of elderly users may be more stable at night, and the model automatically reduces the feature weights at night time points. Dynamic health feature encoding data is suitable for multimodal data scenarios and can process physiological, behavioral, environmental and psychological data at the same time, providing unified feature input for subsequent modeling and analysis. Since the dynamic attention mechanism performs local calculations based on time series, it can generate feature encoding data in real time to meet the real-time requirements in health management.

[0065] In this example, the health status of an elderly user C in a retirement community needs to be monitored in real time. The system collects the user's heart rate data, movement trajectory, indoor humidity, and voice emotion characteristics. After the time series data is preprocessed, dynamic health feature coding data needs to be extracted as the core input for subsequent health status modeling.

[0066] Heart rate data, movement trajectory, indoor humidity, and voice emotion features are arranged in chronological order to form a continuous time series. Based on the heart rate data, the system detects that the heart rate fluctuations during the day are large, and the abnormal points are marked with high importance scores (such as ); Time points with long periods of no movement in the action trajectory data are also given higher importance scores (such as ). Through normalization calculation, the feature weight at each time point is adjusted to between 0 and 1. For example, when the indoor humidity is detected to be below 40% for a long time, the feature weight is increased to , reflecting its important impact on health status. Combining the feature weights and the corresponding time series feature values, all data points are weighted and summed to generate comprehensive dynamic health feature coding data to describe the dynamic health characteristics of elderly users.

[0067] By generating dynamic health feature coding data, the system can clearly identify the current health risk areas of elderly users (such as heart rate fluctuations and humidity effects), providing accurate input for subsequent health status modeling. At the same time, the elderly care community can take targeted intervention measures for high-weight feature points (such as abnormal humidity and lack of movement) to improve the health management effect of users.

[0068] Preferably, the extraction of the causal features is achieved through a causal analysis model based on a Bayesian network, where the causal strength The calculation of includes the following steps: Constructing a causal graph of health status data, wherein the causal graph includes variable nodes in the health status data and causal relationships between the nodes; Constructing a causal graph of health status data, wherein the causal graph includes variable nodes in the health status data and causal relationships between the nodes; For Node and nodes Causal strength The calculation formula for causal strength is: , in, For Node and nodes the causal strength between them; For the node After the intervention node The conditional probability distribution of ; For Node Probability distribution in the non-intervention state; Based on the causal strength calculation results, variable relationships with high causal strength are extracted to generate causal feature chain data, where the causal feature chain is used to describe the key causal relationships between variables in the health status data.

[0069] In the present invention, in order to reveal the causal relationship between variables in health status data, a causal analysis model based on Bayesian network is used to extract causal features. Bayesian network is a directed acyclic graph structure used to describe the causal relationship between variables, with nodes representing variables and edges representing causal relationships. In the present invention, Bayesian network is used to construct a causal graph of health status data, and causal strength is calculated through intervention simulation to extract key causal relationships between variables.

[0070] The calculation of causal strength is based on intervention theory, which holds that by applying an intervention operation to a certain variable (i.e., fixing the value of the variable), the degree of its influence on other variables can be observed. In the present invention, causal strength is quantified by calculating the difference in probability distribution of the variable under intervention and without intervention, thereby intuitively reflecting the strength of the causal relationship between the variables.

[0071] The input of causal feature extraction is health status data. Health status data is generated through time series modeling and multimodal feature fusion, covering the user's physiological, behavioral, environmental and psychological dimensions. For example, the health status data of elderly users may include heart rate, blood pressure, activity level, indoor humidity, and voice emotion characteristics.

[0072] Constructing a causal graph of health status data through a Bayesian network includes the following specific steps: Each indicator in the health status data is defined as a variable node. For example, "heart rate", "humidity" and "activity level" are considered as nodes in the graph. Through a data-driven approach, the time series characteristics of the health status data are used to generate causal relationships between nodes and establish directed edges between variables. For example, if humidity changes are identified as the cause of activity changes, a causal edge from "humidity" to "activity level" is established.

[0073] The core of the causal graph is to reflect the causal relationship between variables, rather than simple correlation. For example, although "lower humidity" and "increased heart rate" may have a high correlation in the data, the Bayesian network reveals the causal chain of "humidity → sleep quality → heart rate" through causal reasoning.

[0074] For any two nodes in the graph and , causal strength The calculation of includes the following steps: (1) By simulating the variables intervention operation (such as fixing the humidity at 40%), calculate the variable Conditional probability distribution under intervention state .

[0075] (2) Calculation variables Natural probability distribution under unintervention conditions .

[0076] (3) Causal strength calculation formula: .

[0077] This formula quantifies the strength of the causal relationship. >0, indicating intervention right There is a positive causal effect; if <0, it indicates a negative causal impact.

[0078] Based on the calculation results of causal strength, variable relationships with high causal strength are extracted to generate causal feature chains. Causal feature chains are the most important set of causal relationships in health status data, which are used to describe the key impact paths of health risks. For example, the causal chain of "humidity drop → sleep interruption → heart rate fluctuation" can intuitively reveal the complex causal relationship between environment, behavior and physiology, providing a basis for health intervention.

[0079] Through the causal analysis of the Bayesian network, the causal relationship between variables can be accurately identified to avoid misjudging simple correlation as causal relationship. For example, "humidity change → sleep quality" instead of "humidity change → heart rate fluctuation". Through the calculation of causal strength, the strength of the causal relationship between variables can be clearly quantified. For example, the causal strength of humidity on sleep is 0.8, while the causal strength of noise on sleep is 0.4. The causal feature chain provides an intuitive causal basis for health risk analysis and intervention optimization, and can help formulate targeted health management strategies. The causal map can be updated in real time to adapt to the dynamic changes in the user's health status, ensuring that the results of the causal analysis are always accurate and reliable.

[0080] In this embodiment, the heart rate, activity level, sleep quality and indoor humidity of elderly user D in a retirement community are collected in real time by smart devices to generate health status data. The community needs to analyze the causal relationship between the variables in the user's health status to provide a basis for health risk prediction and intervention design.

[0081] Causal graph construction: Humidity, heart rate, activity level, and sleep quality are defined as variable nodes in the causal graph. Data analysis found that there are potential causal relationships between humidity and sleep quality, sleep quality and heart rate, and activity level and heart rate, and causal edges are constructed.

[0082] Causal strength calculation: Simulate the intervention operation of increasing humidity to 50% and calculate the conditional probability distribution of sleep quality . Calculate the probability distribution of sleep quality in natural state According to the formula , the causal strength of humidity on sleep quality was calculated, and the result was 0.7, indicating that humidity has a significant positive impact on sleep quality.

[0083] Causal feature chain generation: Based on the calculation results of causal strength, a causal chain "humidity → sleep quality → heart rate" is generated. This chain shows that humidity indirectly affects heart rate fluctuations by affecting sleep quality.

[0084] Through causal analysis results, the community identified "humidity", a key factor in health management, and designed targeted health intervention measures, including maintaining humidity between 40%-60% through smart humidifiers, thereby improving users' sleep quality and reducing the health risks of heart rate fluctuations.

[0085] like Figure 3 As shown, based on health status data and causal characteristic data, health risk analysis is performed, the rapid fluctuation trend of data is captured by using short-term time series models, and the long-term trend of health changes is analyzed in combination with long-term trend prediction models to generate health risk data; based on the health risk data, a health intervention plan is generated; The present invention uses health status data and causal feature data to perform health risk analysis, aiming to dynamically identify potential risks of health changes from users' multidimensional health data and design personalized health intervention plans. This method combines the analytical capabilities of short-term time series models with long-term trend prediction models to capture the rapid fluctuation trends and long-term change patterns of health data. Through the collaborative analysis of these two types of models, accurate health risk data is generated, and scientific intervention plans are further provided based on risk data.

[0086] Short-term time series models are mainly used to capture short-term and drastic fluctuations in health status data. For example, a rapid increase in the heart rate of an elderly user within a few minutes may indicate heart problems; long-term trend prediction models identify slow changes in health status, such as gradually worsening hypertension or chronic fatigue, by analyzing the long-term evolution trend of the time series. The combination of the two can achieve a comprehensive assessment of health risks.

[0087] Health risk data is a multi-dimensional dynamic quantitative result that is used to describe the risk source and degree of health changes of users. The health intervention plan generated based on risk data proposes personalized and executable intervention suggestions through intelligent decision-making algorithms.

[0088] The input of health risk analysis includes health status data and causal feature data: Health status data describes the user's current overall health status and is composed of multi-dimensional feature vectors, such as heart rate stability, behavioral activity, and environmental adaptability. Causal feature data is extracted from the causal analysis model to describe the key causal relationships and their strengths between health variables, such as the causal chain of "humidity drop → sleep interruption → heart rate fluctuation".

[0089] Health risk analysis is achieved through the following two types of models: Short-term time series model: Captures short-term dramatic changes in health status data, such as sudden heart rate fluctuations or acute mood swings. Uses Convolutional Neural Network (CNN) to extract local features from short time windows of health status data. The model slides and analyzes data features within a fixed time period (such as heart rate changes in the last 5 minutes) to identify sudden abnormal patterns. Outputs short-term risk indicators, such as heart rate fluctuations exceeding the normal range in the past 10 minutes, marked as "high risk".

[0090] Long-term trend prediction model: Analyze the long-term trend of health status data and identify slowly developing health risks, such as the worsening of chronic diseases. Use the Long Short-Term Memory (LSTM) network to model the time series of health status data and capture long-term dependencies. LSTM learns long-term trends through memory units, such as the user's sleep time gradually decreasing in the past month. Output long-term risk trends, such as predicting that the user's fatigue risk will gradually increase in the next week.

[0091] Comprehensive risk assessment: Integrate short-term risk indicators with long-term trend forecast results to form comprehensive health risk data. Use a weighted algorithm to conduct a combined analysis of short-term and long-term risks, and assign weights based on the importance of health risks and the strength of causal feature data. For example, when the causal feature chain "lower humidity → interrupted sleep → heart rate fluctuations" has a high causal strength, the weight of this chain in the comprehensive risk assessment is significantly increased.

[0092] Based on health risk data, the system automatically generates personalized health intervention plans. The implementation steps are as follows: (1) Determination of intervention targets: Identify the main risk points and priority intervention targets based on health risk data. For example, if the short-term risk indicator shows abnormal heart rate fluctuations, and the long-term trend forecast shows that the decrease in humidity causes sleep problems, then the priority intervention is the environmental humidity. Identify the variables that need to be adjusted (such as humidity) and the target range (such as increasing humidity to 40%-60%).

[0093] (2) Generation of personalized intervention plans: Figure 4As shown in the figure, a conditional generative adversarial network (Conditional GAN) is used to generate multiple candidate health intervention plans. The generator designs possible intervention measures (such as increasing exercise and adjusting environmental conditions) based on health risk data, and the discriminator evaluates the rationality and effectiveness of each plan and finally selects the optimal intervention plan. It includes specific intervention measures (such as adjusting humidity and increasing social activities), execution steps (such as taking a 20-minute walk every morning), and expected effects (such as reducing heart rate fluctuations).

[0094] Through the collaborative analysis of short-term time series models and long-term trend prediction models, it is possible to accurately identify acute risks (such as sudden heart rate fluctuations) and chronic risks (such as long-term fatigue), and provide a comprehensive health risk assessment. Generate personalized intervention plans based on health risk data to ensure that intervention measures target the source of risk and improve the effectiveness of health management. Short-term and long-term risk models can update health risk data in real time, dynamically adjust intervention plans, and adapt to changes in user health status. The introduction of causal feature data gives the intervention plan a scientific basis, ensuring that intervention measures directly target the core causes of health risks.

[0095] In this example, the health status of elderly user E in a retirement community needs to be monitored in real time. The system collects their heart rate data (physiological characteristics), activity level (behavioral characteristics), indoor humidity (environmental characteristics), and voice emotion data (psychological characteristics). Based on this data, the system needs to identify the user's health risks and generate a personalized intervention plan.

[0096] Short-term risk analysis uses a convolutional neural network to analyze heart rate fluctuations in the last 5 minutes and finds that the heart rate suddenly increases from 60 beats per minute to 110 beats per minute, marking it as high risk.

[0097] Long-term trend prediction: By analyzing the activity volume of the past month through the long short-term memory network, it was found that the number of users' daily steps gradually decreased, and the long-term risk was increased fatigue.

[0098] The comprehensive risk assessment combines the causal characteristic chain of "reduced humidity → decreased sleep quality → heart rate fluctuations" to evaluate the weight of the impact of humidity on heart rate fluctuations, and finally generates health risk data, showing that the main risks are "insufficient environmental humidity" and "reduced activity".

[0099] Health intervention plan generation: Increase indoor humidity to 40%-60%; increase daily walking time to 30 minutes. Through the conditional generative adversarial network, the following intervention measures are generated: start the smart humidifier to increase the humidity to 50%; remind users to take a 30-minute walk every morning and record the number of steps through a smart bracelet; remind caregivers to talk to users to relieve psychological pressure and observe changes in users' emotions.

[0100] Through health risk analysis and the implementation of intervention plans, retirement communities have improved users’ indoor environmental conditions, increased their daily activity levels, effectively reduced the risk of heart rate fluctuations, and improved their overall health.

[0101] Preferably, the health risk analysis comprises the following steps: Using a short-term time series model to capture the rapid fluctuation trend of health feature data, the short-term time series model is a convolutional neural network; Analyzing the long-term trend of health changes based on a long-term trend prediction model, wherein the long-term trend prediction model is a long short-term memory network; Health risk data are generated by integrating the rapid fluctuation trend of the short-term time series model and the trend analysis results of the long-term trend prediction model.

[0102] In this invention, in order to comprehensively evaluate the health status of users and dynamically identify potential risks, a joint health risk analysis method based on short-term time series model and long-term trend prediction model is designed. This method captures the short-term fluctuation trend and long-term trend of user health changes through the analysis of multi-dimensional health feature data, and generates health risk data that can quantify the degree and source of health risks.

[0103] The short-term time series model uses a convolutional neural network (CNN) to extract fast fluctuation features in the local range of the time series, such as a sudden increase in heart rate, sleep interruptions, and drastic changes in mood swings. The long-term trend prediction model is based on the long short-term memory network (LSTM), which uses its memory ability for time series to identify long-term evolution trends in health changes, such as chronic fatigue or a gradual decline in activity. The analysis results of the two models are weighted and integrated to generate overall health risk data to provide support for subsequent health management.

[0104] The input of health risk analysis is dynamic health feature data, which is generated by the fusion of multimodal health data and covers the user's physiological (such as heart rate, blood pressure), behavioral (such as activity level, sleep time), environmental (such as humidity, air quality) and psychological (such as mood swings) characteristics.

[0105] A convolutional neural network is used as a short-term time series model, and its powerful feature extraction capability is used to analyze local changes in health feature data. The time series data is divided into windows of fixed length (for example, each window contains the health feature data of the last 10 minutes). Within each window, CNN extracts fast-fluctuating features through convolution operations, such as a sudden increase in the user's heart rate in the past 10 minutes, a sudden drop in activity, etc. The convolution kernel size and step size are set according to the length of the time window and the data sampling frequency to ensure the accuracy of local feature extraction. A short-term fluctuation risk indicator is generated to quantify the degree of abnormality of health features in a short period of time. For example, a "heart rate fluctuation risk value of 0.85" indicates a high risk of abnormal heart rate fluctuation.

[0106] The long short-term memory network (LSTM) is used as a long-term trend prediction model to capture the long-term dependencies of time series through its gating mechanism. The complete time series health feature data (such as activity data in the past 30 days) is input, and the LSTM gradually updates the hidden state to learn the long-term change trend. The health trend prediction value is generated through the output layer, for example, it is predicted that the activity will gradually decrease in the next week. The long-term trend risk indicator is generated to quantify the long-term trend of health changes. For example, "fatigue trend risk value of 0.75" indicates a high risk of fatigue in the future.

[0107] The analysis results of short-term volatility risk and long-term trend risk are combined into overall health risk data. Based on the weight adjustment of causal feature data, short-term and long-term risk indicators are weighted. For example, if the causal feature data shows that "humidity change → sleep interruption → heart rate fluctuation" has a high causal strength, the weight of the humidity-related part of the short-term risk indicator is amplified. Generate health risk data, including quantitative results of multidimensional risk sources and risk levels. For example, "the comprehensive health risk value is 0.9, the main risk sources: insufficient humidity and reduced activity."

[0108] The generated health risk data can intuitively reflect the main risk sources and risk levels of the user's health status, and provide a scientific basis for subsequent health intervention plans. For example, insufficient humidity may be the main risk source at present, so the intervention plan will give priority to adjusting the ambient humidity.

[0109] Through the joint analysis of short-term time series models and long-term trend prediction models, acute risks (such as sudden increase in heart rate) and chronic risks (such as long-term fatigue) in health changes can be accurately identified. The model updates short-term and long-term risk data in real time, adapts to the dynamic changes in the user's health status, and provides support for real-time health management. Comprehensive risk data not only provides the overall risk level, but also quantifies the risk contribution of each health feature, providing specific guidance for the design of intervention measures. The introduction of causal feature data makes the generation of risk data more scientific and can reveal the source of risk in a targeted manner.

[0110] In this example, the health status of an elderly user F in a retirement community needs real-time monitoring and risk assessment. The system collects heart rate data (physiological characteristics), activity level (behavioral characteristics), indoor humidity (environmental characteristics), and voice emotion data (psychological characteristics). Through health risk analysis, the user's main health risks are identified to provide support for subsequent intervention.

[0111] Short-term risk analysis: The heart rate data of the past 10 minutes is used as a time series input. Through the convolutional neural network, the rapid fluctuation of heart rate from 60 beats per minute to 100 beats per minute is identified, and the short-term fluctuation risk value is calculated as 0.85.

[0112] Long-term trend analysis: The activity data of the past 30 days is used as a time series input. Through the LSTM model, the long-term downward trend of activity from 5,000 steps per day to 2,000 steps per day is identified, and the long-term risk value is calculated as 0.75.

[0113] Comprehensive risk assessment: Combined with the causal characteristic data, it is found that the causal chain of "lack of humidity → decreased sleep quality → reduced activity" has a high causal strength. The comprehensive assessment results in health risk data, with a comprehensive health risk value of 0.9, and the main sources of risk are lack of humidity and reduced activity.

[0114] Based on health risk data, the retirement community prioritizes adjusting the humidity of the user's living environment to 40%-60%, and formulates an activity plan for users to walk 3,000 steps a day. Through these intervention measures, the risk of heart rate fluctuations and long-term fatigue in users has been successfully reduced.

[0115] Preferably, the generation of the health intervention plan is achieved by a conditional generative adversarial network, which includes a generator and a discriminator, wherein: The generator generates multiple health intervention candidates based on the health risk data; The discriminator evaluates the effectiveness and pertinence of candidate health intervention plans and selects the optimal health intervention plan as the final output.

[0116] The present invention realizes the generation of health intervention plans through conditional generative adversarial network (CGAN). CGAN is an extended model based on generative adversarial network (GAN), which allows samples that meet specific needs to be generated under specific conditions. In the present invention, CGAN uses health risk data as conditional input to generate personalized health intervention candidate plans, and finally selects the optimal health intervention plan through adversarial training between the generator and the discriminator.

[0117] The role of the generator is to generate multiple candidate health intervention plans based on health risk data, aiming to cover all possible intervention measures and their combinations. The discriminator evaluates the effectiveness and pertinence of the generated intervention plans, and continuously optimizes the generator's generation ability through adversarial training, so that the final output health intervention plan has higher pertinence and practicality.

[0118] The input for generating health intervention plans is health risk data. Health risk data is generated by a combination of short-term time series models and long-term trend prediction models, and includes the main sources of health risks and risk levels of users. For example, a user's health risk data may show "lack of humidity (high risk)", "heart rate fluctuations (medium risk)", and "reduced activity (low risk)".

[0119] The generator receives health risk data as conditional input. The health risk data is represented in vector form, and each dimension corresponds to a health risk and its weight. For example, the risk weight of insufficient humidity is 0.8, and the risk weight of heart rate fluctuation is 0.6. The generator uses the risk data as constraints and generates multiple health intervention candidate solutions through a neural network. For example, based on the high risk of insufficient humidity, the generator may generate the following candidate solutions: Increase humidity to 40%-60%; increase sleeping time; improve air circulation.

[0120] The generated candidate solutions are represented as multi-dimensional vectors, with each dimension corresponding to a specific intervention measure and its parameters.

[0121] The discriminator receives the candidate solutions generated by the generator and evaluates them in combination with the health risk data. For example, if the candidate solution includes "starting the air purifier", the discriminator will evaluate whether this measure is effective in mitigating the risk of insufficient humidity. The discriminator scores the candidate solutions based on their effectiveness and pertinence, and feeds the results back to the generator for optimization. Effectiveness measures the ability of the candidate solution to address the main sources of risk, and pertinence measures whether the solution meets the user's personalized needs. The discriminator selects the best candidate solution as the final output.

[0122] Through adversarial training between the generator and the discriminator, the generator's generation ability is continuously optimized, making the generated candidate solutions more in line with the user's health risk characteristics. In each round of training, the generator generates candidate health intervention solutions, the discriminator scores them, and uses the scoring results to adjust the generator's parameters. After multiple rounds of iterations, the generator is able to generate high-quality intervention solutions that meet actual needs. After adversarial training of the conditional generative adversarial network, the health intervention solutions finally generated have the following characteristics: Efficiency: The program content directly targets the main sources of health risks; Personalization: The program fully considers the user's personalized health needs and living habits; Executability: The program content is clear and easy for users to implement. For example, "take a 30-minute walk every day and adjust the humidity to 50%" is a clear and specific intervention plan.

[0123] The generator can generate multiple candidate plans based on health risk data, covering all possible intervention measures to ensure the comprehensiveness of the intervention. Through the adversarial evaluation of the discriminator, the health intervention plan finally output is more targeted and can effectively alleviate the main health risks of users. The conditional generative adversarial network can update the health intervention plan in real time to adapt to the dynamic changes in the user's health status. Through personalized and targeted design, the health intervention plan can more efficiently alleviate health risks and improve the overall health level of users.

[0124] In this embodiment, the health status of elderly user G in a retirement community requires real-time intervention. The health risk data shows that the main sources of health risks include "insufficient humidity (high risk, weight 0.8)" and "heart rate fluctuation (medium risk, weight 0.6)". A set of scientific and effective health intervention plans needs to be generated to help reduce health risks.

[0125] Health risk data is used as input to the generator, including insufficient humidity (weight 0.8) and heart rate fluctuation (weight 0.6).

[0126] Candidate generation: The generator generates the following candidates: Solution 1: Adjust the humidity to 45%-55%; remind users to drink 2000ml of water every day; Solution 2: Turn on the air purifier and adjust the humidity to 50%; do some light gymnastics to relieve heart rate fluctuations; Plan 3: Perform deep breathing exercises for 15 minutes every night and adjust the indoor humidity to 40%-60%.

[0127] Solution evaluation: The discriminator evaluates the effectiveness and pertinence of each candidate solution. For example, the measure of "alleviating heart rate fluctuations through light gymnastics" in Solution 2 has no direct effect on improving humidity, resulting in a low score.

[0128] Optimal solution selection: The discriminator selects solution 1 as the final output solution because it more efficiently alleviates the risks of insufficient humidity and heart rate fluctuations at the same time.

[0129] According to Health Intervention Plan 1, the retirement community adjusted the indoor humidity of users to 45%-55% and reminded users to increase their water intake, successfully reducing the risk of users' heart rate fluctuations while improving their overall health status.

[0130] Collect users' intervention execution feedback data, optimize the health intervention plan based on the feedback data, and generate an optimized health intervention plan; share the optimized health intervention plan to relevant terminal devices to achieve closed-loop optimization of health management.

[0131] Preferably, the intervention execution feedback data includes the user's physiological feedback data, behavior execution data, environmental response data and psychological feedback data, wherein: Physiological feedback data includes real-time physiological indicators of heart rate, blood pressure and blood oxygen saturation; Behavioral execution data include users’ completion of intervention tasks; Environmental response data include dynamic changes in indoor humidity, air quality, and noise levels; Psychological feedback data includes emotional change characteristics obtained through voice sentiment analysis.

[0132] The intervention execution feedback data in the present invention is used to dynamically monitor and evaluate the actual implementation effect of health intervention measures, and comprehensively reflect the user's response during the intervention process through multi-dimensional feedback data. These feedback data are divided into four categories: physiological feedback data, behavioral execution data, environmental response data, and psychological feedback data. Each category provides feedback information from the dimensions of physiological state, behavioral habits, environmental adaptability, and mental health status, providing a scientific basis for the optimization and adjustment of intervention plans.

[0133] Physiological feedback data: Real-time collection of physiological indicators such as the user's heart rate, blood pressure, and blood oxygen saturation to evaluate whether the intervention measures have achieved the expected health improvement effect. For example, through real-time heart rate monitoring, it can be determined whether the user's exercise intervention has improved cardiovascular function.

[0134] Behavior execution data: records the user's actual completion of health intervention tasks, which is used to evaluate the feasibility and compliance of the intervention. For example, the completion rate of the user's daily walking goal can be recorded through a step counter.

[0135] Environmental response data: monitors the dynamic changes in humidity, air quality, and noise levels in the user's living environment to determine the effectiveness of environmental interventions. For example, a humidity sensor can be used to determine whether the adjustment of a humidifier reaches the target humidity range.

[0136] Psychological feedback data: Based on voice sentiment analysis technology, the user's emotional fluctuation characteristics are monitored to evaluate the impact of intervention measures on the psychological state. For example, it can be used to determine whether the improvement in the user's mood is related to social intervention.

[0137] By comprehensively analyzing these feedback data, the present invention can identify the effectiveness and shortcomings of the intervention plan, and further optimize the intervention strategy to improve the health management effect.

[0138] Physiological feedback data: collected in real time by wearable devices worn by users, including heart rate sensors, blood pressure measuring devices and blood oxygen monitoring devices. Real-time recorded physiological indicators include the fluctuation curve of heart rate, systolic and diastolic blood pressure values, and dynamic changes in blood oxygen saturation.

[0139] Behavior execution data: Behavior tracking devices (such as smart bracelets, sports recorders) record users’ behavior activities, such as the total number of steps taken per day, task completion rate (for example, the proportion of users who have completed their drinking water goals), exercise duration, etc.

[0140] Environmental response data: Indoor environmental sensors monitor humidity, air quality (PM2.5 concentration) and noise levels in real time. Dynamically changing environmental indicators, such as indoor humidity change curve (%RH), real-time concentration value of PM2.5 in the air ( ) and noise level (dB).

[0141] Psychological feedback data: The voice emotion analysis module collects user voice samples and performs emotion recognition. Voice analysis generates emotion change features, such as the frequency of positive emotions, the proportion of negative emotions, and the range of emotion fluctuations.

[0142] The multi-dimensional feedback data is integrated into a health status feedback vector to quantify the intervention effect. For example, the heart rate fluctuation amplitude, humidity adjustment range, walking goal completion rate, etc. are integrated into a unified feedback evaluation vector. A mapping relationship between feedback data and health status improvement is established to determine the impact weight of each type of feedback data. For example, humidity has a higher weight on sleep quality improvement, while noise level has a lower weight on exercise effect.

[0143] The feedback data is used to determine whether the intervention has achieved the expected goal. For example, if the target humidity is 40%-60% and the actual humidity feedback is 45%, the humidity intervention is considered effective. The behavior execution data is used to determine whether the user has completed the intervention task as required. For example, if the user's daily walking goal is 5,000 steps and the user actually completes 4,000 steps, the completion rate is 80%.

[0144] Analyze abnormalities in feedback data. For example, if the heart rate fluctuates abnormally and the humidity adjustment does not meet the standard, it is determined that the humidity intervention measures are insufficient and the humidifier operation strategy needs to be optimized. Optimize the intervention plan based on feedback data. For example, adjust the time or intensity of humidity intervention, or increase the completion rate of exercise tasks by increasing the user's behavioral incentives.

[0145] Comprehensive feedback data can provide the following support for the optimization of health management: Dynamically adjust intervention goals: adjust exercise intensity or diet plan in real time based on physiological feedback data.

[0146] Improve environmental adaptability: Optimize indoor humidity, air quality and other parameters through environmental response data.

[0147] Enhance user compliance: Identify bottlenecks in intervention task completion through behavioral performance data and design personalized incentives.

[0148] Improve mental health: Identify the causes of emotional changes through psychological feedback data, increase social interaction or psychological counseling and other measures.

[0149] Through data collection in four dimensions, namely physiological, behavioral, environmental and psychological, the user's response during the health intervention process is fully reflected. By real-time monitoring and analyzing feedback data, the health intervention plan is dynamically optimized to improve the intervention effect and user satisfaction. Individual differences are identified through feedback data, and health intervention strategies are customized to improve the accuracy and scientificity of health management. Problems in the intervention process can be quickly identified, such as environmental conditions not meeting standards or insufficient user task completion rates, to ensure timely remedial measures.

[0150] In this example, an elderly user H in a retirement community participated in a health intervention program, including environmental adjustment (increasing indoor humidity), exercise goals (walking 5,000 steps per day), and psychological counseling (daily voice interaction). Feedback data is needed to evaluate the intervention effect and optimize the intervention plan.

[0151] Physiological feedback data: By wearing the device to monitor heart rate and blood pressure in real time, it was found that the user's heart rate stability improved after the intervention (from a fluctuation range of 10 beats / minute to 5 beats / minute), but the blood pressure was still high (systolic pressure remained above 145 mmHg).

[0152] Behavior execution data: The walking goal completion rate of users recorded by smart bracelets was 70%, which did not reach the goal of 5,000 steps per day.

[0153] Environmental response data: Monitoring by environmental sensors revealed that humidity adjustment did not reach the target range (actual humidity was 35%, target humidity was 40%-60%).

[0154] Psychological feedback data: Through voice sentiment analysis, it was found that the frequency of users' positive emotions in evening interactions was low (only 30%), suggesting that the effect of improving emotions was limited.

[0155] By analyzing the feedback data, the health intervention plan was optimized: the indoor humidity target range was adjusted to a more relaxed 40%-50% to reduce the dependence on humidity adjustment equipment; simpler walking goals were added for users, and more incentives were added; in response to psychological and emotional issues, the interaction time of caregivers was increased every evening. After the optimization, the user's heart rate was further stabilized, the walking completion rate increased to 90%, and the frequency of positive emotions increased to 60%.

[0156] Preferably, the optimized health intervention plan is shared to relevant terminal devices through blockchain technology, and the blockchain technology includes building a smart contract to dynamically manage access rights to the health intervention plan.

[0157] The present invention utilizes the distributed storage, non-tamperability and high transparency of blockchain technology to safely and efficiently share the optimized health intervention plan to relevant terminal devices. In this process, smart contracts are used to dynamically manage access rights to health intervention plans, ensuring accurate distribution and privacy protection of plans, while achieving multi-party collaboration and real-time updates.

[0158] Blockchain technology stores health intervention plans and related data through distributed ledgers to ensure the integrity and security of data when shared to multiple terminals. Smart contracts, as an automatically executed code logic, are responsible for dynamically managing access rights to health intervention plans, ensuring that different terminal devices can only access data related to their roles. For example, user family members can only view the summary information of the intervention plan, while caregivers can access detailed intervention measures and implementation instructions.

[0159] Through the encrypted storage of blockchain and the access rights management of smart contracts, the security and privacy of health intervention plans in the sharing process are ensured. Smart contracts realize the real-time distribution and dynamic update of health intervention plans, ensuring that users and care teams always have the latest intervention task information. Through the distributed ledger technology of blockchain, health intervention plans can be shared consistently among users, care teams and family members, improving collaboration efficiency. Smart contracts can automatically record and verify the execution of intervention tasks, reduce manual intervention, and improve the efficiency of health management.

[0160] like Figure 5 As shown, a system for implementing the health management method based on multidimensional data analysis comprises: The data processing module is used to collect the user's physiological data, behavioral data, environmental data and psychological data, pre-process the collected data, and use an anomaly detection model based on time series analysis to perform an anomaly analysis on the pre-processed health data to generate abnormal marking data; collect four types of user data, namely physiological, behavioral, environmental and psychological, perform pre-processing operations such as cleaning and standardization on the data, and use an anomaly detection model based on time series analysis to mark possible abnormal points (such as sudden heart rate fluctuations or humidity abnormalities) to provide high-quality input data for subsequent analysis.

[0161] The health status modeling module is used to extract the user's health characteristics based on the pre-processed health data and abnormal marking data, generate health characteristic data, combine the health characteristic data to generate health status data, and analyze the health status data through the causal analysis model based on the Bayesian network, extract the causal characteristics of health changes, and generate causal characteristic data; through in-depth analysis of the pre-processed health data and abnormal marking data, extract the user's dynamic health characteristics, and generate a multidimensional data representation reflecting the health status. The causal analysis model based on the Bayesian network reveals the causal relationship between variables in the health status (such as the indirect effect of humidity on heart rate fluctuations), generates causal characteristics, and provides a scientific basis for subsequent risk assessment.

[0162] The health risk analysis module is used to capture the rapid fluctuation trend of data based on health status data and causal characteristic data using a short-term time series model, and analyze the long-term trend of health changes in combination with a long-term trend prediction model to generate health risk data, and generate health intervention plans based on the health risk data; it combines short-term time series models (used to capture the rapid fluctuation trend of health data, such as a sharp increase in heart rate) and long-term trend prediction models (used to analyze the long-term trend of health changes, such as the evolution of chronic fatigue) to generate health risk data and design targeted health intervention plans based on the risk data.

[0163] The health intervention optimization module is used to collect the user's intervention execution feedback data, optimize the health intervention plan based on the feedback data, generate the optimized health intervention plan, and share the optimized health intervention plan to the relevant terminal devices to achieve closed-loop optimization of health management. By collecting the user's intervention execution feedback data in real time (including physiological, behavioral, environmental and psychological multi-dimensional feedback), the intervention plan is dynamically adjusted and optimized, and finally a personalized optimized intervention plan is generated, which is safely and efficiently shared to the user, nursing team or family terminal through blockchain technology to achieve closed-loop health management.

[0164] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A health management method based on multidimensional data analysis, characterized in that: The following steps are involved: Collect the user's physiological data, behavioral data, environmental data and psychological data, pre-process the collected data, and generate pre-processed health data; Performing anomaly analysis on the health data using an anomaly detection model based on time series analysis to generate anomaly labeled data; Based on the preprocessed health data and abnormal marked data, extract the user's health characteristics and generate health characteristic data; Combine health feature data to generate health status data; analyze health status data through a causal analysis model based on a Bayesian network, extract causal features of health changes, and generate causal feature data; Based on health status data and causal characteristic data, health risk analysis is carried out, using short-term time series models to capture the rapid fluctuation trend of data, and combining long-term trend prediction models to analyze the long-term trend of health changes and generate health risk data; Generate health intervention plans based on health risk data; Collecting intervention execution feedback data from users, optimizing the health intervention plan based on the feedback data, and generating an optimized health intervention plan; Share the optimized health intervention plan to relevant terminal devices to achieve closed-loop optimization of health management.

2. The health management method according to claim 1, characterized in that: The anomaly detection model is a sliding window model based on time series analysis. By setting the size of the sliding window and dynamically adjusting the threshold, local anomaly detection is performed on the preprocessed health data and anomaly marker data is generated.

3. The health management method according to claim 1, characterized in that: The health status data is generated by combining dynamic health feature coding data and abnormal marking data, and health status modeling is performed using a recursive neural network. The recursive neural network takes time series data as input and outputs health status data.

4. The health management method according to claim 3, characterized in that: The extraction of dynamic health feature encoding data is achieved through a dynamic attention mechanism, which assigns weights based on the importance of time series features, where the weights The calculation of includes the following steps: Calculate time points The feature importance score of , the feature importance score is calculated by It is obtained by weighted summing of the eigenvalue of and the eigenvalues ​​of adjacent time points; Calculate time points The feature weight Specifically: ， in, For time point The corresponding feature weights; For time point The feature importance score is calculated by the trend of the feature value change of the time series; is the index of the feature in the time series; is the total length of the time series; represents the natural exponential function; According to the feature weight at each time point And the corresponding eigenvalues, generate dynamic health feature encoding data.

5. The health management method according to claim 1, characterized in that: The extraction of the causal features is achieved through a causal analysis model based on a Bayesian network, where the causal strength The calculation of includes the following steps: Constructing a causal graph of health status data, wherein the causal graph includes variable nodes in the health status data and causal relationships between the nodes; For Node and nodes Causal strength The calculation formula for causal strength is: ， in, For Node and nodes the causal strength between them; For the node After the intervention node The conditional probability distribution of ; For Node Probability distribution in the absence of intervention; Based on the causal strength calculation results, variable relationships with high causal strength are extracted to generate causal feature chain data, where the causal feature chain is used to describe the key causal relationships between variables in the health status data.

6. The health management method according to claim 1, characterized in that: The health risk analysis includes the following steps: Using a short-term time series model to capture the rapid fluctuation trend of health feature data, the short-term time series model is a convolutional neural network; Analyzing the long-term trend of health changes based on a long-term trend prediction model, wherein the long-term trend prediction model is a long short-term memory network; Health risk data are generated by integrating the rapid fluctuation trend of the short-term time series model and the trend analysis results of the long-term trend prediction model.

7. The health management method according to claim 1, characterized in that: The generation of the health intervention plan is achieved through a conditional generative adversarial network, which includes a generator and a discriminator, wherein: The generator generates multiple health intervention candidates based on the health risk data; The discriminator evaluates the effectiveness and pertinence of candidate health intervention plans and selects the optimal health intervention plan as the final output.

8. The health management method according to claim 1, characterized in that: The intervention execution feedback data includes the user's physiological feedback data, behavior execution data, environmental response data and psychological feedback data, wherein: Physiological feedback data includes real-time physiological indicators of heart rate, blood pressure and blood oxygen saturation; Behavioral execution data include users’ completion of intervention tasks; Environmental response data include dynamic changes in indoor humidity, air quality, and noise levels; Psychological feedback data includes emotional change characteristics obtained through voice sentiment analysis.

9. The health management method according to claim 1, characterized in that: The optimized health intervention plan is shared to relevant terminal devices through blockchain technology, and the blockchain technology includes building a smart contract to dynamically manage access rights to the health intervention plan.

10. A system for implementing the health management method based on multidimensional data analysis according to any one of claims 1 to 9, characterized in that: include: A data processing module, used to collect physiological data, behavioral data, environmental data and psychological data of the user, pre-process the collected data, and perform anomaly analysis on the pre-processed health data using an anomaly detection model based on time series analysis to generate anomaly labeled data; The health status modeling module is used to extract the user's health characteristics based on the preprocessed health data and abnormal marker data, generate health characteristic data, combine the health characteristic data to generate health status data, and analyze the health status data through a causal analysis model based on a Bayesian network to extract the causal characteristics of health changes and generate causal characteristic data; The health risk analysis module is used to capture the rapid fluctuation trend of data based on health status data and causal characteristic data using a short-term time series model, and analyze the long-term trend of health changes in combination with a long-term trend prediction model, generate health risk data, and generate health intervention plans based on the health risk data; The health intervention optimization module is used to collect users' intervention execution feedback data, optimize the health intervention plan based on the feedback data, generate an optimized health intervention plan, and share the optimized health intervention plan to relevant terminal devices to achieve closed-loop optimization of health management.

Images