Health maintenance big data intelligent management system

Abstract

The invention belongs to the technical field of computers, and particularly relates to a health-preserving big data intelligent management system. The method aims at solving the problems of data islands, individualized service deficiency and intervention lag in the field of health preservation and health care. The system comprises a data perception layer, a governance layer, a knowledge engine layer and an application service layer, and precise and real-time personalized health management is realized through multi-source heterogeneous data fusion and dynamic portrait construction in combination with integrated learning risk early warning, deep reinforcement learning intervention recommendation and a closed-loop feedback optimization mechanism. The system supports cross-platform interaction and multi-party cooperation, and user compliance and management efficiency are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computers, and particularly relates to a health preservation big data intelligent management system. BACKGROUND

[0002] With the deep integration of information technology and health science, big data technology is increasingly widely applied in the field of medical health, and gradually expands to disease prevention, health management and personalized intervention and other aspects. In particular, in the field of health preservation and health care, data analysis based on individual physiological data, living habits, environmental factors and other multi-source information has become an important means to realize precise and intelligent health management. In this field, by collecting and integrating long-term health data of users, health assessment models are constructed using data mining and machine learning technology, aiming to provide scientific, dynamic and personalized health preservation guidance services, so as to improve the health level and quality of life of the public.

[0003] Among them, health preservation big data management, as the core link connecting data collection and intelligent decision-making, its main goal is to realize efficient storage, cleaning, fusion and analysis of massive heterogeneous health data, and on this basis to support real-time response of personalized recommendation function. This technical direction not only involves the application of traditional database management and cloud computing architecture, but also emphasizes the automation and intelligentization of data processing process, in order to cope with the dynamic changes of user health status and the diversity of health preservation needs, and to ensure that the system can continuously output health management schemes with clinical significance and practical value.

[0004] Although the prior art has made certain progress in health data collection and basic analysis, there are still many bottlenecks: the scattered data sources lead to serious information island phenomenon, making it difficult to realize effective integration of cross-platform and multi-modal data; the data processing process relies on manual rule setting and lacks self-adaptive learning ability, so it is impossible to dynamically optimize model parameters according to user feedback; the modeling accuracy of the system to user behavior patterns is insufficient, resulting in low individuality and poor practicability of health preservation suggestions; at the same time, the existing architecture has limited performance in real-time and scalability, making it difficult to support stable operation under large-scale user concurrent access. The above problems are particularly prominent in the active health management scenario with high timeliness requirement, which seriously affects the actual application effect and user experience of the system, and therefore an health preservation big data intelligent management system with self-learning ability, supporting multi-source data fusion and being able to dynamically evolve is urgently needed. SUMMARY

[0005] The present application aims to provide a health maintenance big data intelligent management system to solve the technical problems of serious information island, lack of individualized service, lagging intervention strategy and difficulty in quantitatively evaluating the effect of health management in the current health maintenance field. The health management platform in the prior art relies on static questionnaires and periodic physical examination data, lacks the ability of continuous collection and fusion analysis of multi-dimensional data such as user daily life behavior, physiological index dynamic change, environmental factors and psychological state, resulting in rough health assessment results, generalized individualized suggestions, and inability to achieve precise, real-time active intervention. In addition, the massive heterogeneous data generated by various wearable devices, mobile applications and medical institutions cannot be effectively integrated to form a unified data-driven decision mechanism, resulting in low overall intelligence level of the system, poor user experience and insufficient compliance. Therefore, it is urgent to build a system-level solution that can realize the deep fusion of multi-source heterogeneous health maintenance data, dynamic modeling, intelligent reasoning and closed-loop feedback.

[0006] The technical scheme of the present application is a distributed system architecture composed of four logical levels including a data perception layer, a data governance layer, a knowledge engine layer and an application service layer. The data perception layer continuously collects the user's physical parameters, motion trajectory, sleep quality, diet record, emotional fluctuation, indoor and outdoor air quality, light intensity and geographic location information through the intelligent terminal device deployed on the user end, wearable physiological monitoring device, home environment sensor and third-party data interface. The above data enters the data governance layer in a streaming transmission mode. The layer first aligns the time stamp and normalizes the spatial coordinates of the received raw data, then uses an outlier detection algorithm based on a sliding window to identify and eliminate outlier data points, further uses an adaptive weighted interpolation method to compensate and repair the missing data segment, and ensures the integrity and time sequence consistency of the data sequence. The cleaned data is stored in a time series database, a graph database and an unstructured file storage system for supporting efficient retrieval and access of high-frequency sampling signals, entity relationship networks and multimedia logs respectively.

[0007] The knowledge engine layer as the core intelligent hub of the application contains a dynamic portrait module, a risk early warning module, an intervention recommendation module and a feedback optimization module. The dynamic portrait module constructs and continuously updates the user's full life cycle health profile based on the stored multi-modal data. The profile not only covers basic demographic characteristics and past medical history, but also extracts behavioral pattern characteristics through a long short-term memory neural network model to generate quantifiable physical fitness index, metabolic load score, psychological resilience level and environmental exposure cumulative value. The risk early warning module adopts an ensemble learning framework to combine gradient boosting tree models and Bayesian networks to jointly probabilistically predict the early occurrence of multiple chronic diseases. Specifically, the system takes each index in the dynamic portrait as an input feature vector, calculates the risk tendency score of each disease category through a pre-trained risk factor weight matrix, and triggers a graded warning signal according to the set dynamic threshold. The first level warning represents low risk normal monitoring, the second level warning starts the expert review process, and the third level warning automatically activates the emergency response protocol and sends alert information to the designated contact person.

[0008] The intervention recommendation module receives the output results from the risk early warning module and the current life scene context of the user, calls the pre-set rule engine and deep reinforcement learning strategy network to generate personalized health intervention schemes. The rule engine is built-in structured knowledge graph transformed from traditional Chinese medicine theory of treating disease before it occurs, modern nutrition guidelines and exercise rehabilitation medicine consensus, covering multiple dimensions such as diet taboos, rest adjustment, emotional counseling, acupoint massage and light exercise prescription. The deep reinforcement learning strategy network takes maximizing long-term health benefits as the objective function, collects user behavior feedback and subsequent physiological indicator changes as reward signals after each recommendation, and iteratively optimizes strategy parameters to realize the evolution from "thousands of people with different conditions" to "one person with one strategy". The generated intervention scheme includes specific execution actions, recommended time periods, expected duration and completion evaluation standards, and is converted into easy-to-understand guidance statements through natural language generation technology.

[0009] The application service layer provides a cross-platform integrated interaction interface for end users, supporting multiple access methods such as smartphones, tablets, smart speakers, and in-vehicle systems. The system dynamically adjusts the information presentation density and interaction mode according to the user's operation habits and information preferences, for example, providing a voice-dominated large font interface for elderly users and pushing socialized health reports with text and images for young users. At the same time, the application service layer opens standardized application programming interfaces, allowing medical institutions, community care centers, and commercial insurance institutions to access securely under authorization, enabling compliant sharing of health data and business collaboration. The feedback optimization module continuously tracks the actual execution of the intervention program, compares the planned actions with the actual behavior logs, calculates the compliance index, and injects the index into the reward function of the intervention recommendation module, prompting the system to preferentially recommend lightweight measures with high feasibility and high acceptance, forming a complete closed-loop control link of "perception-analysis-decision-execution-feedback".

[0010] Preferably, the body mass index calculation process in the dynamic image module is as follows: select five biological markers with Chinese medicine syndrome significance, such as tongue image, pulse waveform, daily range of basal body temperature, resting heart rate variability, and sweat electrolyte concentration, digitize their representations and map them to the membership space of nine basic body types, determine the main body type and its mixed degree through fuzzy clustering algorithm, and update the results every 24 hours to guide the selection of subsequent conditioning direction.

[0011] Preferably, the dynamic threshold setting mechanism in the risk warning module is as follows: the initial threshold is determined according to the percentile of large-scale epidemiological survey data, and then self-adaptively drifts according to the trend of individual historical baseline value; when the monitoring data of 7 consecutive days shows that a certain indicator shows a monotonic increasing or decreasing trend, the system determines that there is a structural deviation, at this time the original threshold is adjusted by 15% of the trend slope in the same direction to maintain the warning sensitivity.

[0012] Preferably, the scene context recognition in the intervention recommendation module relies on multi-sensor data fusion judgment. If the system detects that the user is between 08:00 and 18:00 on weekdays, the GPS positioning is stable in the office area, and the low-to-high frequency ratio of heart rate variability is continuously higher than 3.0, it is determined as a high-pressure work scene, at this time the system preferentially recommends breathing regulation training and neck relaxation exercises and other micro-intervention measures that can be completed at the workstation.

[0013] Preferably, the compliance index calculation formula in the feedback optimization module is the ratio of actual execution time to recommended time multiplied by action accuracy, where action accuracy is determined by the correlation coefficient of the action trajectory captured by the inertial sensor of the wearable device and the standard template, and a correlation coefficient less than 0.7 is considered as invalid execution.

[0014] Preferably, the data exchange between the modules in the knowledge engine layer follows a unified message queue protocol, all events are marked with a globally unique identifier and a digital signature, ensuring traceability and tamper-proofing of operations; the system performs a blockchain snapshot backup once a day at dawn, writes the hash value of the key decision log into the consortium chain for post-audit use.

[0015] Preferably, the application service layer has offline mode running capability, locally caches the intervention plan and data collection tasks of the last 3 days during network interruption, and automatically synchronizes to the cloud after connection recovery, and triggers the integrity verification process to prevent data loss.

[0016] Preferably, the data governance layer introduces a differential privacy protection mechanism, adds noise conforming to Laplace distribution to the original data when providing aggregate statistical reports externally, so that the data contribution of a single user cannot be inferred, ensuring that personal privacy is not leaked.

[0017] Preferably, the intervention recommendation module supports multi-objective optimization, and when facing multiple health goals such as blood glucose control, weight management and sleep improvement, a set of non-dominated solutions is generated by using a Pareto frontier search algorithm, and the user can select the compromise solution that best meets the current intention for execution.

[0018] Compared with the prior art, the advantages and positive effects of the present application are: The present system realizes the paradigm shift from passive response to active prevention, solves the fundamental defects of data fragmentation, decision experience and service static in the traditional health management mode by establishing a full-chain technical system covering "data collection-cleaning governance-intelligent deduction-individual intervention-effect feedback"; the system uses a method combining multi-modal sensor fusion and deep learning to significantly improve the accuracy and timeliness of health status recognition, and the average extension of disease risk early warning period is more than 6 months; the introduction of a dynamic strategy optimization mechanism based on reinforcement learning enables the intervention scheme to have self-evolution ability, and the user compliance is improved by 42% compared with the control group; by constructing a standardized data interface and privacy protection framework, the information barriers between individuals, families, communities and professional institutions are broken down, forming a new health management ecosystem with multi-party participation and collaborative governance, which provides a replicable and generalizable technical path for realizing universal health coverage. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 It is the overall technical scheme architecture diagram of the health preservation big data intelligent management system proposed by the present application. DETAILED DESCRIPTION

[0020] Please refer to Figure 1The system architecture of the health management big data intelligent management system is composed of four logical levels of data perception layer, data governance layer, knowledge engine layer and application service layer. Each level works collaboratively through a distributed computing framework to realize the whole-chain closed-loop management from multi-source heterogeneous data collection to personalized health intervention. The system is deployed in a hybrid cloud environment composed of edge computing nodes and cloud data centers. The edge nodes are responsible for local real-time data processing and low-latency response, while the cloud end undertakes large-scale model training, long-term trend analysis and knowledge extraction tasks across user groups. The overall technical process of the system starts with the continuous perception of multi-dimensional information such as user physiology, behavior and environment. Then the data governance layer completes cleaning, normalization and structured storage. Then the knowledge engine layer performs dynamic modeling, risk deduction and strategy generation. Finally, the application service layer provides executable personalized health guidance to users, and continuously optimizes the decision model based on feedback data.

[0021] The data perception layer, as the front-end data entry of the system, undertakes the function of collecting user's full-dimensional health-related data continuously and unobtrusively. This layer is composed of sensor arrays deployed on user's personal terminal devices and external third-party data interfaces. The intelligent terminal devices include smartphones, tablet computers and special health management hosts, which are equipped with high-precision accelerometers, gyroscopes, geomagnetic sensors and global positioning system modules to record the user's daily activity types, exercise intensity, gait characteristics and spatial movement trajectories. Wearable physiological monitoring devices include smart bracelets, chest strap heart rate monitors and ear clip blood oxygen probes to continuously obtain resting heart rate, heart rate variability, respiratory rate, body temperature, blood oxygen saturation and blood pressure trend values at a sampling frequency of 1 to 5 seconds. The home environment sensor network is distributed in the bedroom, living room and kitchen area, including PM2.5 / PM10 particulate matter concentration detectors, carbon dioxide concentration sensors, hygrometers, light intensity meters and noise decibel meters to achieve minute-level monitoring of the living micro-environment quality. Diet records are completed by capturing food images through the phone camera and combining voice input auxiliary labels. The system uses a convolutional neural network model to identify and analyze the food types, portion ratios and cooking methods in the images, and converts them into standardized nutritional component data. Emotional fluctuation information is derived from the extraction of tone features in the user's text content published on social platforms and voice calls. The natural language processing module performs sentiment polarity analysis on the text, and the voice signal is calculated for the psychological stress index through the Mel-frequency cepstrum coefficient and the fundamental frequency change rate. Geographic location information not only includes static coordinates, but also integrates place semantic labels such as "office", "gym", "hospital" or "park", which are obtained by matching GPS positioning data and point of interest databases. All the above data are encrypted and transmitted to the data governance layer through the Secure Sockets Layer protocol, and the Message Queue Telemetry Transport protocol is used to maintain long connections. In weak network environments, it automatically switches to batch compression upload mode to ensure the integrity and timeliness of the data stream. The data collection frequency is differentiated according to the differences in index types: physiological parameters are pushed in real time on demand, environmental data are reported every 30 seconds, behavior logs are aggregated and sent every 5 minutes, and third-party medical reports trigger synchronization requests immediately after updating.

[0022] The data governance layer receives the original data stream from the data perception layer and executes a series of automated processing procedures to ensure the data quality relied on for subsequent analysis. First, a timestamp alignment operation is performed. Due to the clock drift phenomenon in different devices, the system introduces a Network Time Protocol (NTP) server as the unified time reference to correct the timestamps of all data points to the UTC+8 standard time, with the error controlled within ±50 milliseconds. For spatial coordinate data, a Geographic Information System (GIS) projection conversion algorithm is used to map the longitude and latitude values in the WGS-84 coordinate system to the Mercator plane rectangular coordinate system, facilitating subsequent spatial clustering and path analysis. Outlier detection adopts a statistical method based on a sliding window. The window length is set to 15 minutes, and the rolling mean μ and standard deviation σ are calculated for each numerical indicator. If a data point x satisfies |x - μ|>3σ, it is determined as an outlier and removed. For skewed data with non-normal distribution, such as PM2.5 concentration or sentiment scores, the interquartile range method is used to redefine the threshold boundary, that is, when x<Q1 - 1.5IQR or x>Q3 + 1.5IQR, it is marked as an anomaly, where Q1 and Q3 are the first and third quartiles respectively, and IQR is the interquartile range. Missing data compensation adopts an adaptive weighted interpolation method, which constructs a weight matrix by comprehensively considering three dimensions: temporal proximity, physiological relevance, and scenario consistency. For example, when repairing a segment of missing heart rate data, the system not only refers to the observed values within 1 minute before and after (temporal weight), but also introduces the changing trends of respiratory rate and skin conductance as covariates (physiological weight), and determines whether it is in a sleep, exercise, or sedentary state to adjust the form of the interpolation function (scenario weight). Specifically, let the missing variable at the interpolation time be , and its estimated value is determined by the following formula: where, is the comprehensive weight of the th adjacent observation point, is its actual measured value, is the number of valid adjacent points. The weight is calculated by the formula: In the formula, the first term represents the time decay factor, and are adjustment coefficients, which control the overall weight scale and time sensitivity respectively; the second term is the contribution of physiological relevance, is the correlation strength coefficient, is the variable and is the Pearson correlation coefficient between them, represents the correlation with The auxiliary variable has a physiological coupling relationship; the third term is a scene consistency indicator function. Match gain coefficients to the scene. Let be the indicator function, at the current time Scene tags and neighboring points The value is 1 if the labels are the same, and 0 otherwise. After this processing, the system can effectively recover data gaps caused by temporary device disconnection or signal interference, ensuring the continuity of the time sequence.

[0023] After cleaning and repair, the data is categorized and imported into three heterogeneous database systems. High-frequency sampled physiological and environmental time-series data is written to a time-series database, which uses a columnar storage structure and supports millisecond-level time range queries and aggregation operations, suitable for quickly extracting waveform segments or calculating moving averages within specific time periods. Structured information with dense entity relationships, such as user-disease-symptom-drug association networks, constitution type evolution paths, and intervention causal graphs, is stored in a graph database, utilizing the relationship representation capabilities of nodes and edges to achieve efficient traversal and reasoning of complex knowledge. Unstructured data, including tongue and facial images, dietary photos, voice logs, and free text descriptions, is uniformly stored in a distributed file system, with content retrieval enabled by a metadata index table. Before being entered into the database, all data entries are appended with a globally unique identifier, acquisition device ID, data source type, encrypted hash fingerprint, and access permission tag, forming a complete data lineage tracing link. The database cluster is configured with master-slave replication and off-site disaster recovery backup mechanisms to ensure 24 / 7 uninterrupted service even in the event of a single point of failure.

[0024] The knowledge engine layer is the core intelligent hub of the application, integrating four functional units: dynamic portrait module, risk early warning module, intervention recommendation module and feedback optimization module, to realize the knowledge transition from raw data to executable decisions. The dynamic portrait module is responsible for building and continuously updating the user's full life cycle health profile. The profile maps the individual's physiological state, behavior preference and environmental interaction characteristics in the form of digital twin. The profile initialization stage enters basic demographic information, including age, gender, height, weight, occupation category and family history of genetic diseases. Subsequently, the profile is continuously enriched and derived through machine learning algorithms. Body mass index is one of the key quantitative indicators in the profile. The calculation process focuses on five biological markers with Chinese medicine syndrome differentiation significance: tongue image features are extracted by image segmentation algorithm, including tongue color, moss quality and crack features, which are converted into RGB color space three channel mean, gray level co-occurrence matrix contrast and local binary pattern texture entropy; pulse waveform is collected by radial artery pressure sensor, extracting main wave peak value, double beat wave height and rising branch slope as time domain features; the basic body temperature daily range refers to the difference between the morning wake-up temperature and the evening maximum temperature, reflecting the autonomic nervous regulation ability; the resting heart rate variability is represented by the standard deviation SDNN of RR interval sequence and the low frequency power LF / HF ratio to represent the vagus nerve tension; the sweat electrolyte concentration is measured by flexible patch sensor to determine the molar concentration of sodium, potassium and chloride ions. The above five-dimensional feature vector is mapped to the membership space of nine basic body types, namely, normal body, qi deficiency body, yang deficiency body, yin deficiency body, phlegm-dampness body, damp-heat body, blood stasis body, qi depression body and special body. The system uses fuzzy C-means clustering algorithm for body type identification, sets the number of clusters to 9, and selects weighted Euclidean distance as the distance measure. The weights are assigned according to the clinical expert experience, so that the features with greater influence on syndrome differentiation have higher discrimination. After the algorithm converges, the membership values of each body type are output, and the type corresponding to the maximum value is taken as the main body type, and the second largest value greater than 0.3 is recorded as the mixed body. The obtained body type conclusion is refreshed every 24 hours in the morning, serving as the basis for subsequent conditioning direction selection. The metabolic load score is a linear weighted sum of blood glucose fluctuation amplitude, blood lipid abnormality degree, uric acid level and visceral fat area, with weights of 0.3, 0.25, 0.2 and 0.25 respectively. The total score ranges from 0 to 10, and a score of 6 or above is defined as high load state. The psychological resilience level is based on three indicators: stability of daily emotional fluctuation curve, negative event response efficiency and mindfulness exercise participation. The support vector machine classifier is used to divide it into strong, medium and weak levels. The cumulative value of environmental exposure is calculated by the cumulative length of time in the environment with PM2.5>75μg / m3, noise>65dB(A) or carbon dioxide>1000ppm per month, in hours, to assess the risk of chronic toxicity damage.

[0025] The risk warning module uses an ensemble learning framework to jointly predict the early occurrence risk of multiple chronic diseases. The input feature vector is composed of indicators output by the dynamic portrait module, including body mass index, metabolic load score, psychological resilience level, environmental exposure cumulative value, historical physical examination abnormal items, recent physiological parameter trend slope, and lifestyle score. The system preloads two base models: the gradient boosting tree model is good at capturing nonlinear interaction effects and is suitable for processing discrete features and sparse data; the Bayesian network can explicitly model the causal dependence between variables and provide interpretable risk propagation paths. The two models run in parallel, each outputting independent risk tendency scores for five diseases: diabetes, hypertension, coronary heart disease, stroke, and depression, with scores ranging from 0 to 1, representing the likelihood of developing the disease within the next 6 months. The system further uses a stacking fusion strategy, taking the outputs of the two models as meta-features input into a logistic regression combiner to generate the final ensemble prediction results. A dynamic threshold setting mechanism is used to determine whether to trigger a warning signal. The initial threshold is determined based on the 90th percentile of national epidemiological survey data, for example, a diabetes risk score higher than 0.65 enters the attention interval. Thereafter, the system starts an adaptive drift correction program, which calculates the trend slope of each core indicator for the user over the past 7 days daily. If any indicator shows a continuous monotonic increase or decrease with a p-value < 0.05, it is determined that there is a structural bias. At this time, the original threshold is adjusted in the same direction by 15% of the absolute value of the trend slope, for example, if a user's fasting blood glucose is continuously rising with a slope of +0.12 mmol / L / day, the diabetes warning threshold is lowered to 0.65 x (1 0.12 x 0.15) = 0.6304, increasing the sensitivity to detect early signs of deterioration. The warning level is divided into three levels: level one corresponds to a risk score between the dynamic threshold and the threshold + 0.1, the system only records in the background and intensifies data sampling density; level two crosses the threshold + 0.1 but does not reach 0.8, starting the expert review process, generating a preliminary evaluation report by the AI-assisted diagnosis system, and pushing it to the signed doctor's end waiting for manual confirmation; level three reaches or exceeds 0.8, automatically activating the emergency response protocol, synchronously pushing red alert information to the user himself, emergency contacts, and community health service centers through the application service layer, including risk type, current score, key abnormal indicators, and recommended treatment department.

[0026] The intervention recommendation module receives the output results of the risk warning module and the current life scene context of the user, and generates a personalized health intervention scheme. Scene context recognition relies on multi-sensor data fusion judgment, and the system analyzes the combination mode of time, location, activity state and physiological parameters in real time. For example, when the system detects that the time is between 08:00 and 18:00 on weekdays, the GPS positioning is stable in the range of the labeled office building, the accelerometer shows a long time static state, and the low frequency to high frequency ratio of heart rate variability is continuously higher than 3.0, it is determined as a high pressure work scene; if the sleep latency is more than 30 minutes at night, the deep sleep ratio is less than 15%, and there are multiple awakenings in the early morning, it is marked as insomnia problem state; if it appears near the park green land in the morning on weekends, the step frequency is accelerated, and the heart rate reaches more than 60% of the reserve heart rate, it is identified as an outdoor exercise period. The rule engine has a structured knowledge graph built-in, which is transformed from the theory of traditional Chinese medicine prevention, modern nutrition guidelines and consensus of sports rehabilitation medicine, containing more than 12000 entity nodes and 38000 semantic relationships. The knowledge graph covers diet rules, such as "people with phlegm-dampness constitution should avoid eating fatty and greasy food"; rest adjustment suggestions, such as "people with liver fire should go to sleep before 11pm"; emotional counseling methods, such as "anxiety state recommends mindfulness breathing training"; acupoint massage prescription, such as "headache can press the Hegu and temple acupoints"; and light exercise programs, such as "patients with knee joint degeneration are suitable for swimming rather than running". Each rule is accompanied by applicable conditions, priority score and evidence level label. The deep reinforcement learning strategy network takes maximizing long-term health benefits as the objective function, the state space S is composed of the user's current dynamic portrait vector, the action space A contains all possible intervention measures, and the reward signal R is defined as the improvement degree of key physiological indicators in the next 24 hours minus the user's subjective discomfort feedback. The network uses a double deep Q network architecture, including an online network and a target network, the experience replay pool capacity is set to 1 million transition samples, and the exploration rate of the ε-greedy strategy is exponentially decayed from the initial 0.9 to 0.1. After each recommendation, the system collects the user's behavior feedback, including whether to view the suggestion, the actual execution time, the action completion degree recorded by the wearable device, and the next morning self-evaluation state, as the reward signal to update the Q value function, and promote the strategy to gradually converge to a high-compliance and high-benefit recommendation combination. The generated intervention scheme includes specific execution actions, recommended execution time, expected duration, resource links (such as audio guidance, video tutorials) and completion evaluation standards, and is converted into oral and context-adapted guidance sentences through natural language generation technology, for example, "You have a lot of work pressure today, please do 5 minutes of abdominal breathing exercise at 3pm, and we have prepared the guidance audio for you."

[0027] The application service layer provides a cross-platform integrated interface for end users, supporting multiple access methods such as smartphones, tablets, smart speakers, and car systems. The system dynamically adjusts the information presentation density and interaction mode according to the user's age group, educational background, device usage habits, and information preferences. For users over 65 years old, the interface automatically switches to a large font version dominated by voice, with button sizes not less than 48 pixels, and key reminders broadcast through smart speakers, supporting dialect recognition and simple instruction responses; for young white-collar groups, weekly health reports are pushed, including data visualization charts, achievement badges, and social sharing functions. Information display follows the principle of gradual disclosure, with only today's key tasks and urgent notifications displayed on the homepage, and deep data requiring step-by-step clicking to expand, avoiding cognitive overload. The system has offline mode running capability, with the last 3 days of intervention plans and data collection task lists cached locally during network interruption, using device built-in storage space to save pending synchronization raw sensor data packets, and automatically initiating incremental synchronization requests after connection is restored, triggering integrity verification processes, confirming successful transmission by comparing local and cloud log hash values, and preventing data loss. The application service layer opens standardized application programming interfaces, using OAuth 2.0 authentication mechanisms to allow medical institutions, community care centers, and commercial insurance institutions to securely access with user explicit authorization, realizing compliant sharing and business collaboration of health data. For example, a signed family doctor can access the patient's monthly trend report through the API, the insurance company can provide premium discounts based on 6 consecutive months of high adherence records, and the community center can issue regional health tips based on group environmental exposure data.

[0028] The feedback optimization module continuously tracks the actual execution of the intervention plan by the user, forming a key feedback channel of the closed-loop control. The system compares the planned actions with the actual behavior logs, and calculates the compliance index as the core indicator to measure the effectiveness of the intervention. The compliance index is defined as the ratio of the actual execution time to the recommended time multiplied by the action accuracy rate, where the action accuracy rate is determined by the correlation coefficient obtained by matching the motion trajectory captured by the inertial sensor of the wearable device with the pre-stored standard template using dynamic time warping algorithm. If the correlation coefficient is less than 0.7, it is considered as invalid execution. For example, a recommended shoulder and neck stretching training lasting 10 minutes, if the user only completes 6 minutes and the motion is severely deformed (correlation coefficient 0.62), the compliance index is (6 / 10) x 0.62 = 0.372. The index is updated daily and fed back into the reward function of the intervention recommendation module, so that the priority of low compliance measures in future recommendations is significantly reduced. The system further analyzes the potential reasons for the compliance difference, establishes a multi-factor regression model to identify the key obstacles affecting the execution willingness, such as time conflict, insufficient physical strength, device limitation or psychological resistance, and adjusts the subsequent strategy accordingly, prioritizing the recommendation of lightweight measures with high feasibility, short cycle and low threshold, such as splitting a 30-minute walk into 3 times of 10-minute micro-movements, or replacing complex recipes with pre-made healthy meal ordering links. Individualized feedback reports are generated at the end of each month, showing the compliance trend, health indicator improvement and cost-benefit analysis, helping users establish a positive incentive cycle.

[0029] The data exchange between the modules in the knowledge engine layer follows a unified message queue protocol. All events are marked with a globally unique identifier and an elliptic curve digital signature to ensure traceability and tamper resistance. The message body is encapsulated in a lightweight JSON format, containing fields such as timestamp, source module, target module, data payload, and check code. The system performs a blockchain snapshot backup at 02:00 every morning, writing the SHA-256 hash values of all key decision logs for the day to a consortium chain based on Hyperledger Fabric. The participating nodes include designated regulatory agencies, partner hospitals, and third-party audit units, providing post-responsibility tracing and compliance review. The data governance layer introduces differential privacy protection mechanisms. When providing aggregated statistical reports to the outside world, noise conforming to the Laplace distribution is added to the original data. The noise scale parameter Δf is set according to the sensitivity of the query function, making it impossible to reverse-infer the data contribution of a single user, satisfying the compliance requirements of GDPR and the Personal Information Protection Law. The intervention recommendation module supports multi-objective optimization. When users face multiple health goals such as blood sugar control, weight management, and sleep improvement, the system converts each goal into a constraint condition and a target function, and uses the Pareto frontier search algorithm to generate a set of non-dominated solutions. Each solution represents a different trade-off, such as "strict sugar control + moderate weight loss + sleep maintenance" or "stable blood sugar + significant weight loss + sleep improvement". Users can choose the most suitable compromise solution based on their current life priorities, embodying the people-oriented design concept.

[0030] The system breaks through the limitations of traditional platforms, which are limited to data display and simple reminders, and realizes fundamental changes from passive recording to active intervention, from general groups to individual exclusivity, and from isolated services to ecological linkage. The multi-modal sensor fusion technology refines the health status perception to the minute level, and the deep learning model extends the disease risk warning period to more than 6 months, achieving a qualitative leap compared with the traditional annual physical examination mode. The strategy optimization mechanism based on reinforcement learning enables the intervention scheme to have the ability of continuous evolution. Clinical trial data shows that the average compliance of users is improved by 42% compared with the control group, mainly due to the significant enhancement of the scene adaptability and execution friendliness of the recommended measures. The standardized API interface and privacy protection framework break down the information barriers between individuals, families, communities, and professional institutions, forming a new health management collaboration paradigm driven by data, with clear rights and responsibilities, and compatible incentives. The system has been piloted in three cities, covering 26,000 residents, and has generated 4.8 million personalized intervention schemes, helping to discover 1,173 cases of early risk of chronic diseases, 89% of which have been clinically diagnosed, verifying the feasibility and social value of the technical route. In the future, by connecting more types of biological sensors and expanding the knowledge graph to cover more diseases, the system's universality and accuracy can be further improved, providing solid technical support for the Health China strategy.

Claims

1. A health and wellness big data intelligent management system, characterized in that, include: The data perception layer is used to continuously collect users' raw data through smart terminal devices, wearable physiological monitoring devices and home environment sensors deployed on the user end, and send the raw data to the data governance layer in a streaming manner. The data governance layer is used to perform timestamp alignment and spatial coordinate normalization on the received raw data. The cleaned data is classified and stored in time-series databases, graph databases and unstructured file storage systems. The knowledge engine layer integrates a dynamic profiling module, a risk warning module, an intervention recommendation module, and a feedback optimization module. The dynamic profiling module builds and continuously updates the user's full life cycle health record based on multimodal data and generates cumulative values. The risk warning module uses an integrated learning framework to perform joint probability prediction of the early occurrence risk of chronic diseases and calculates risk propensity scores. The intervention recommendation module receives risk warning results and the user's current life scenario context, calls the rule engine and deep reinforcement learning strategy network to generate health intervention plans, and the feedback optimization module continuously tracks the user's actual implementation of the intervention plan and calculates the compliance index. The application service layer provides a cross-platform integrated interactive interface for end users, with open and standardized application programming interfaces. Most organizations can securely access the interface with authorization, enabling compliant sharing of health data and business collaboration, thereby forming a complete closed-loop control chain.

2. The health preservation big data intelligent management system according to claim 1, characterized in that, The body mass index calculation process in the dynamic portrait module is as follows: Five biomarkers with TCM diagnostic significance are selected, namely, tongue image, pulse waveform, daily basal body temperature variation, resting heart rate variability, and sweat electrolyte concentration. These biomarkers are digitally represented and mapped to the membership space of nine body constitution types. The main body constitution type and its degree of combination are determined by fuzzy clustering algorithm. The results are updated every 24 hours to guide the selection of subsequent conditioning directions.

3. The health preservation big data intelligent management system according to claim 2, characterized in that, The dynamic threshold setting mechanism in the risk warning module includes: the initial threshold is determined based on the percentile of large-scale epidemiological survey data, and then adaptive drift correction is performed according to the changing trend of individual historical baseline values; when the monitoring data for 7 consecutive days shows that a certain indicator shows a monotonically increasing or decreasing trend, the system determines that there is a structural shift, and at this time the original threshold is adjusted in the same direction by 15% of the trend slope to maintain the warning sensitivity.

4. The health preservation big data intelligent management system according to claim 3, characterized in that, The scenario context recognition in the intervention recommendation module relies on multi-sensor data fusion judgment. If the system detects that the user is between 08:00 and 18:00 on a weekday, the GPS positioning is stable in the office area, and the low-frequency to high-frequency ratio of heart rate variability is consistently higher than 3.0, it is determined to be a high-pressure work scenario. In this case, micro-intervention measures such as breathing regulation training and neck relaxation exercises that can be completed at the workstation are given priority.

5. The health preservation big data intelligent management system according to claim 4, characterized in that, The compliance index calculation method in the feedback optimization module includes: multiplying the ratio of actual execution time to recommended execution time by the action accuracy to obtain the compliance index, wherein the action accuracy is determined by the correlation coefficient between the action trajectory captured by the inertial sensor of the wearable device and the standard template, and a correlation coefficient lower than 0.7 is considered invalid execution.

6. The health preservation big data intelligent management system according to claim 5, characterized in that, Data exchange between modules within the knowledge engine layer follows a unified message queue protocol. All events are tagged with globally unique identifiers and digital signatures to ensure traceability and tamper-proof operation. The system performs a blockchain snapshot backup once a day at midnight, writing the hash values ​​of key decision logs into the consortium blockchain for post-audit purposes.

7. The health preservation big data intelligent management system according to claim 6, characterized in that, The application service layer has the ability to run in offline mode. During network interruption, it locally caches the intervention plans and data collection tasks of the most recent 3 days. Once the connection is restored, it automatically synchronizes to the cloud and triggers an integrity verification process to prevent data loss.

8. The health preservation big data intelligent management system according to claim 7, characterized in that, The data governance layer introduces a differential privacy protection mechanism, which adds noise conforming to a Laplace distribution to the original data when providing aggregated statistical reports to the outside world, so that the data contribution of an individual user cannot be reversed and personal privacy is protected from being leaked.

9. A health preservation big data intelligent management system according to claim 8, characterized in that, The intervention recommendation module supports multi-objective optimization. When faced with multiple health goals such as blood glucose control, weight management, and sleep improvement, it uses the Pareto front search algorithm to generate a set of non-dominated solutions, allowing users to choose the compromise solution that best suits their current needs.

10. A health preservation big data intelligent management system according to claim 9, characterized in that, The rule engine incorporates a structured knowledge graph derived from traditional Chinese medicine's theory of disease prevention, modern nutrition guidelines, and consensus on sports rehabilitation medicine. It covers multiple dimensions, including dietary dos and don'ts, lifestyle adjustments, emotional guidance, acupoint massage, and mild exercise prescriptions. Each rule comes with applicable conditions, priority scores, and evidence level tags to provide precise intervention suggestions for users with different physical conditions and health statuses.

Images