Personalized AI intelligent health management system based on multi-source data fusion

Abstract

The invention relates to the technical field of health management systems, and particularly discloses a personalized AI intelligent health management system based on multi-source data fusion. Comprising a multi-modal health data dynamic acquisition and fusion module, a personalized dynamic health portrait construction module, a hierarchical risk early warning and root cause inference module, a self-adaptive personalized intervention strategy generation and closed-loop optimization module and a privacy protection and federated learning module which are connected in sequence. According to the system, standardized acquisition, quality verification and feature level fusion of different types of health data are realized through a multi-modal health data dynamic acquisition and fusion module, the problems of multi-source data dispersion and poor fusion effect in the prior art are solved, and a high-quality data basis is provided for subsequent health analysis; through an embedded dual adaptive calibration mechanism and a dynamic updating unit, portrait calibration is carried out in combination with a user individual historical baseline and a group generality mode, and the change of a user health state can be adapted in real time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of health management systems, in particular to a personalized AI intelligent health management system based on multi-source data fusion. BACKGROUND

[0002] With the rapid development of artificial intelligence technology and intelligent wearable devices, the health management field is gradually extending from traditional offline medical treatment to online personalized monitoring and intervention. At present, there are various health management related products on the market, which can realize basic health data collection and simple health prompt function, meeting the needs of some users for preliminary monitoring of health status. However, there are still many defects in the existing health management technology that need to be solved. Firstly, the processing capacity of multi-source health data is insufficient. Different sources and types of health data are often in a scattered state, lack effective fusion mechanism, which makes it difficult to fully exploit the value of data, and further affects the accuracy of subsequent health analysis. Secondly, the health portrait construction lacks dynamic and personalized calibration, which is mostly based on fixed templates or simple historical data statistics, and cannot adapt to the changes of user health status in real time, nor can it balance the individual differences and common characteristics of the group. Thirdly, the risk early warning and root cause analysis are disconnected. The existing system can only realize the preliminary judgment of risk level, and cannot accurately trace the key reasons for the risk, which is not conducive to the targeted measures taken by users. Fourthly, the intervention strategy lacks a closed-loop optimization mechanism. The generated intervention scheme is mostly a general suggestion, which does not fully consider the user's execution feedback and health status change, and it is difficult to realize continuous personalized optimization. Fifthly, there are short boards in data privacy protection. In the process of realizing cloud collaborative optimization, there is a risk of leakage of sensitive health data of users. In addition, the special management ability for specific chronic diseases is insufficient. The existing system cannot adapt to the personalized management needs of different chronic diseases, which leads to limited health management effect for this kind of users.

[0003] Based on the above deficiencies of the prior art, there is an urgent need for a personalized AI intelligent health management system that can realize effective fusion of multi-source data, accurately construct dynamic health portrait, accurately warn risks and trace root causes, adaptively generate optimized intervention strategies and protect data privacy and security, in order to improve the scientificity and effectiveness of health management. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a personalized AI intelligent health management system based on multi-source data fusion, which solves the problems in the background art.

[0005] In order to achieve the above object, the present application is realized by the following technical scheme: a multi-source data fusion personalized AI intelligent health management system, comprising a multi-modal health data dynamic acquisition and fusion module, a personalized dynamic health portrait construction module, a hierarchical risk early warning and root cause inference module, an adaptive personalized intervention strategy generation and closed-loop optimization module, and a privacy protection and federated learning module connected in sequence; The multi-modal health data dynamic acquisition and fusion module is used for acquiring multi-dimensional health data through an intelligent terminal, an environmental sensor and a manual input interface, generating a unified health state vector through space-time alignment and feature-level fusion , and outputting to the personalized dynamic health portrait construction module; The personalized dynamic health portrait construction module is based on user historical baseline data and group common mode adaptive calibration, and outputs a dynamic health portrait containing physiological homeostasis, behavior mode and psychological tendency , and transmits to the hierarchical risk early warning and root cause inference module; The hierarchical risk early warning and root cause inference module calculates multi-dimensional health risk probability in parallel , performs interpretable root cause tracing on high-risk events, and outputs the early warning result and root cause information to the adaptive personalized intervention strategy generation and closed-loop optimization module; The adaptive personalized intervention strategy generation and closed-loop optimization module generates a personalized comprehensive intervention scheme according to the early warning and root cause information , tracks intervention feedback and health state changes, and forms a closed-loop data for iterative optimization of the previous modules; The privacy protection and federated learning module is deployed between the local device and the cloud server, realizes local feature extraction and cloud global model aggregation, and guarantees data privacy and security.

[0006] Preferably, the multi-modal health data dynamic acquisition and fusion module comprises a data acquisition unit, a data quality verification unit and a feature fusion unit; The data acquisition unit is used for acquiring physiological signal data, behavior and environment data, and subjective report data to form a multi-source original data set; The data quality verification unit is used for AI-driven quality screening of the original data set, and the specific steps are as follows: S11, analyze the format of each data source, and distinguish real-time stream data and offline reporting data; S12, calculate the signal integrity and noise interference degree of each data source; S13, set a quality threshold , , if or This triggers a data re-collection or supplementary collection mechanism; S14. Perform spatiotemporal alignment on the data that passes the verification to unify the timestamp and data dimension; The feature fusion unit employs an adaptive weighted fusion algorithm based on attention weights to calculate the first... Feature vectors of data sources Fusion weights : ; in, The real-time signal-to-noise ratio of the data source, with a value range of 0-50dB. The value ranges from 0 to 1, representing the relevance to the current health context. , For adjustable hyperparameters, , This is used to balance the effects of signal-to-noise ratio and correlation. Total number of data sources; The unified health status vector It is generated by the feature fusion unit through the following formula: ; And it must meet the fusion effectiveness coefficient. Otherwise, return to the feature fusion unit to readjust the weights.

[0007] Preferably, the personalized dynamic health profile construction module includes an individual baseline calibration unit, a group reference calibration unit, and a profile dynamic update unit; The individual baseline calibration unit is used to calculate the current features. Compared with the user's personal historical baseline offset : ; in, The standard deviation of the user's historical data over the past 90 days; The value ranges from 0 to 3, with larger values ​​indicating more significant deviations from the baseline; an offset threshold is set. ,like Mark it as an abnormal offset and trigger key monitoring; The group reference calibration unit incorporates user characteristics and typical patterns of homogeneous groups. Comparative group consistency factor : ; in, The cosine similarity function is used. The closer the value is to 1, the higher the degree of fit with the group pattern; The dynamic health profile Features after dual calibration are scaled to a personalized ratio Combined into: ; in, It is a unit vector. Weights are assigned based on individual differences and can be dynamically adjusted according to users' health management needs; The portrait dynamic update unit collects new health data every 24 hours and recalculates. and Update dynamic health profile This ensures that the profile is consistent with the user's real-time health status.

[0008] Preferably, the hierarchical risk warning and root cause inference module includes a shared underlying encoder, a multi-dimensional risk prediction unit, a root cause tracing unit, and a hierarchical warning unit; The shared underlying encoder is used to create dynamic health profiles. Perform feature dimensionality reduction and enhancement to output a low-dimensional dense feature vector. ; The multi-dimensional risk prediction unit contains multiple independent risk prediction heads, used to calculate the first... Probability of health risks : ; in, This is the Sigmoid function, with an output range of 0-1; For the first Parameters of a risk prediction head; The trend score of this type of risk-related indicator within a 7-day time window (the value ranges from 0 to 1, and the larger the value is for a more obvious upward trend). The trend influence coefficient; all risk probabilities constitute a risk probability vector. ; The root cause tracing unit, for those identified as high-risk ( , For events involving gradient-weighted class activation mapping, a gradient-weighted method is used to calculate each feature dimension in the health profile. Contribution to the prediction of this risk : ; The top contributors Features ( ) and their corresponding data sources are output as key root causes; The tiered early warning unit sets three risk thresholds: low risk... Medium risk High risk Each triggers a different warning action: low risk only logs the information, medium risk sends a health alert, and high risk is linked to the medical service interface and notifies the guardian.

[0009] Preferably, the adaptive personalized intervention strategy generation and closed-loop optimization module includes a strategy matching engine, an intervention execution tracking unit, and a closed-loop learning unit; The policy matching engine employs a constraint optimization-based policy matching algorithm to obtain the optimal intervention plan by solving the following formula. : ; in, A set of feasible strategies that comply with current medical guidelines and user contraindications; This is a function representing the expected intervention effect, with a value range of 0-1; Execute a cost function for the user, with a value range of 0-1; A personalized moderating factor characterizing user compliance levels; the lower the compliance, the lower the compliance level. The larger the value; The intervention execution tracking unit is used to collect the execution data of the intervention plan in real time. The specific steps are as follows: S21. Receive the execution progress feedback from the user client; S22. Calculate the execution completion rate. , ; S23, if Analyze the reasons for non-implementation and generate a simplified intervention plan; S24. Record user feedback and ratings during the execution process. This will serve as a basis for subsequent strategy optimization. The closed-loop learning unit constructs an intervention-feedback experience pool and stores quadruples. ,in To intervene in the implementation feedback, This represents the change in risk probability; the calibration parameters for the personalized dynamic health profile construction module are applied every 7 days using experience pool data. Prediction header parameters of the hierarchical risk warning and root cause inference module Perform coordinated fine-tuning.

[0010] Preferably, the feature fusion unit of the multimodal health data dynamic acquisition and fusion module also supports dynamic iterative optimization of the fusion weights, with the following specific steps: S31. Set the fusion validity threshold. ; S32. Calculate the unified health status vector after fusion. effectiveness coefficient The formula is as follows: ; in, For data source noise interference, For data integrity; S33, if Maintain current weights ; S34, if Calculate the weight adjustment coefficient Update weights Recalculate ; S35. Repeat steps S32-S34 until... The final unified health state vector is output.

[0011] Preferably, the individual baseline calibration unit of the personalized dynamic health profile construction module also supports dynamic updating of historical baselines, with the following specific rules: The user's personal historical baseline is recalculated every 180 days. The mean and standard deviation of health data over the past 180 days were used; If a user's health status undergoes a significant change, triggering an emergency baseline update, the baseline will be reconstructed based on intensive monitoring data from the 30 days following surgery / diagnosis. ; After the baseline is updated, the personalization ratio is recalculated. ,in The standard deviation of the new baseline ensures that the calibration mechanism adapts to changes in the user's health status.

[0012] Preferably, the root cause tracing unit of the hierarchical risk warning and root cause inference module also supports root cause verification and secondary tracing mechanisms: For the output before One key root cause was identified, and the correlation between these root causes was verified by combining the user's historical health data, and the correlation degree was calculated. ; like Eliminate the root cause and add the next-order contribution feature; Relevance retention after secondary source tracing The root causes are identified, and a final list of root causes is generated, with the correlation strength marked.

[0013] Preferably, the personalized AI intelligent health management system based on multi-source data fusion also includes a system initialization and user modeling module; The module is used to perform the following operations during the system startup phase, based on the user-completed health questionnaire, initial physical examination report, and 7-day intensive monitoring data: S41. Establish the user's initial health baseline. Calculate the initial standard deviation ; S42. Based on user age, gender, medical history, and other information, match typical patterns of homogeneous groups. Initialize the group consistency factor calculation parameters; S43. Construct user risk profiles, marking high-risk disease types and contraindicated intervention items; S44. Initialize the personalized dynamic health profile building module. The prediction header parameters of the hierarchical risk warning and root cause inference module provide the initial configuration for system operation.

[0014] Preferably, the system supports specific optimization for the management of certain chronic diseases. In the specific optimization mode: The multimodal health data dynamic acquisition and fusion module supplements the collection of disease-specific data, enhancing the fusion weight of this type of data. To 1.2-1.5 times the usual weight; The personalized dynamic health profile building module strengthens disease control indicators in the physiological homeostasis dimension and adjusts the personalization ratio. ; The tiered risk warning and root cause inference module adds a specific prediction task for the risk of disease complications and optimizes the trend influence coefficient of the corresponding risk prediction head. ; The adaptive personalized intervention strategy generation and closed-loop optimization module's strategy matching engine prioritizes intervention plans recommended by chronic disease management guidelines, reducing the execution burden and cost functions. weight This will improve user compliance.

[0015] This invention provides a personalized AI-powered intelligent health management system that integrates multi-source data, offering the following advantages: 1. Improve the efficiency of multi-source health data utilization: This system realizes the standardized collection, quality verification and feature-level fusion of different types of health data through the multi-modal health data dynamic acquisition and fusion module, which solves the problems of scattered multi-source data and poor fusion effect in the existing technology, and provides a high-quality data foundation for subsequent health analysis.

[0016] 2. Achieve accurate and personalized dynamic health profile construction: By embedding a dual adaptive calibration mechanism and dynamic update unit, and combining the user's individual historical baseline with the group's common patterns for profile calibration, it can adapt to changes in the user's health status in real time, taking into account individual differences and group reference characteristics, effectively improving the accuracy and dynamic adaptability of the health profile.

[0017] 3. Improve the accuracy and interpretability of risk warning: The hierarchical risk warning and root cause inference module adopts a structure of shared underlying encoder and multiple independent risk prediction heads, realizing parallel and accurate prediction of multi-dimensional health risks; at the same time, through the root cause tracing and verification mechanism, it can accurately locate the key influencing factors of high-risk events, solving the deficiency of existing technologies in risk warning that only knows the result and not the cause, and providing a clear basis for the formulation of subsequent intervention strategies.

[0018] 4. Achieve adaptive closed-loop optimization of intervention strategies: The adaptive personalized intervention strategy generation and closed-loop optimization module generates personalized intervention plans that meet user needs through constrained optimization algorithms, and builds a closed-loop learning mechanism by combining intervention execution feedback and changes in health status. It continuously iterates and optimizes the parameters of the preceding modules to ensure that the intervention strategy can be dynamically adjusted according to the user's health status and execution, thereby improving the intervention effect and user compliance.

[0019] 5. Safeguarding user health data privacy and security: The privacy protection and federated learning module adopts a collaborative learning mode of local feature extraction and cloud-based global model aggregation, which realizes the security of data not leaving the domain, effectively solves the data leakage risk in the cloud collaborative optimization process in existing technologies, and improves the system's security and user trust.

[0020] 6. Adapt to specific chronic disease management needs: The system supports specific optimization modes for specific chronic diseases. By making targeted adjustments to the parameters and function configurations of each core module, it strengthens the specific adaptability of chronic disease-related data collection, profile construction, risk prediction and intervention strategies, broadens the system's applicable scenarios, and improves the health management effect for chronic disease users. Attached Figure Description

[0021] Fig. 1 This is a schematic diagram of the process of a personalized AI intelligent health management system based on multi-source data fusion as described in this invention; Fig. 2 This is a schematic diagram of the principle of a personalized AI intelligent health management system based on multi-source data fusion, as described in this invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figs. 1-2As shown, the present invention provides a technical solution: a personalized AI intelligent health management system for multi-source data fusion, comprising a multimodal health data dynamic acquisition and fusion module, a personalized dynamic health profile construction module, a hierarchical risk warning and root cause inference module, an adaptive personalized intervention strategy generation and closed-loop optimization module, and a privacy protection and federated learning module connected in sequence. The multimodal health data dynamic acquisition and fusion module is used to collect multidimensional health data through smart terminals, environmental sensors, and manual input interfaces, and generate a unified health state vector through spatiotemporal alignment and feature-level fusion. Output to the personalized dynamic health profile building module; The personalized dynamic health profile construction module adaptively calibrates based on users' historical baseline data and common group patterns, outputting a dynamic health profile that includes physiological homeostasis, behavioral patterns, and psychological tendencies. The data is then transmitted to the hierarchical risk warning and root cause inference module. The hierarchical risk warning and root cause inference module calculates multi-dimensional health risk probabilities in parallel. For high-risk events, explainable root cause tracing is performed, and the early warning results and root cause information are output to the adaptive personalized intervention strategy generation and closed-loop optimization module. The adaptive personalized intervention strategy generation and closed-loop optimization module generates a personalized comprehensive intervention plan based on early warning and root cause information. Tracking intervention feedback and changes in health status to form closed-loop data for iterative optimization of preceding modules; The privacy protection and federated learning module is deployed between local devices and cloud servers to achieve local feature extraction and cloud-based global model aggregation, ensuring data privacy and security.

[0024] More specifically, the multimodal health data dynamic acquisition and fusion module includes a data acquisition unit, a data quality verification unit, and a feature fusion unit; The data acquisition unit is used to collect physiological signal data (heart rate, blood pressure, etc.), behavioral and environmental data (movement trajectory, temperature and humidity, etc.), and subjective report data (symptom description, sleep quality, etc.) to form a multi-source raw dataset. The data quality verification unit is used to perform AI-driven quality screening on the original dataset. The specific steps are as follows: S11. Parse the formats of each data source and distinguish between real-time streaming data and offline data; S12. Calculate the signal integrity of each data source. (Value range 0-1, close to 1 if there are no missing data) and noise interference. (Value range: 0-1; the lower the noise, the smaller the value). S13, Set quality threshold , ,like or This triggers a data re-collection or supplementary collection mechanism; S14. Perform spatiotemporal alignment on the data that passes the verification to unify the timestamp and data dimension; The feature fusion unit employs an adaptive weighted fusion algorithm based on attention weights to calculate the first... Feature vectors of data sources Fusion weights : ; in, The real-time signal-to-noise ratio of the data source, with a value range of 0-50dB. The value ranges from 0 to 1, representing the relevance to the current health context. , For adjustable hyperparameters, , This is used to balance the effects of signal-to-noise ratio and correlation. Total number of data sources; The unified health status vector It is generated by the feature fusion unit through the following formula: ; And it must meet the fusion effectiveness coefficient. Otherwise, return to the feature fusion unit to readjust the weights.

[0025] The multimodal health data dynamic acquisition and fusion module comprehensively covers health data across three core dimensions: physiological, behavioral environment, and subjective reports, ensuring the integrity and diversity of the raw data. The data quality verification unit, leveraging AI-driven signal integrity, noise interference calculation, and threshold filtering mechanisms, effectively eliminates low-quality data with severe missing data or excessive noise. Spatiotemporal alignment further unifies the data format and dimensions, providing a reliable foundation for subsequent fusion. The feature fusion unit employs an attention-based adaptive weighting algorithm, dynamically adjusting the weights of each data source based on signal-to-noise ratio and health context relevance. Combined with a fusion effectiveness coefficient verification (≥0.9), the final generated unified health state vector... It can accurately aggregate effective information from multiple sources, highlighting the impact of high-value data while avoiding analytical errors caused by biases in a single data source, providing high-quality and reliable input data support for building personalized dynamic health profiles.

[0026] In this embodiment, for example, a 35-year-old male user uses this system for health management. The data acquisition unit collects physiological signal data such as heart rate and blood pressure in real time through a smartwatch, and collects behavioral and environmental data such as daily movement trajectory and ambient temperature and humidity through mobile phone positioning and environmental sensors. Subjective report data such as daily symptom description and sleep quality score are manually filled in through the system APP, forming a multi-source raw dataset. The data quality verification unit first parses the format of each data source, determining that heart rate and blood pressure are real-time streaming data, movement trajectory and temperature and humidity are semi-real-time streaming data, and symptom description and sleep quality score are offline data. Then, it calculates the signal integrity of each data source. With noise interference Heart rate data collected by smartwatches , Temperature and humidity data collected by environmental sensors , (Due to signal interruptions during certain periods, the data integrity was slightly below the threshold). The supplementary data acquisition mechanism was triggered to replenish the temperature and humidity data for that period, and the results were recalculated. , ,satisfy , The quality threshold was determined; then, all data that passed the verification were spatiotemporally aligned, unifying real-time and semi-real-time streaming data to timestamps of 15 minutes per record, and associating offline data with the daily timestamp of the corresponding date, unifying the data dimension to a 128-dimensional feature vector; hyperparameters were set for the feature fusion unit. , Calculate the real-time signal-to-noise ratio of each data source. Relevance to health context Its central rate data , motion trajectory data , Sleep quality score data , The fusion weights of each data source are calculated using the attention weight formula. The weights are 0.35, 0.28, and 0.37 respectively (the total weight of other data sources is 0.10); then, according to the formula... Generate a unified health state vector and calculate the fusion effectiveness coefficient: ; To meet the requirements for fusion effectiveness, Output to the personalized dynamic health profile building module.

[0027] More specifically, the personalized dynamic health profile construction module includes an individual baseline calibration unit, a group reference calibration unit, and a profile dynamic update unit; The individual baseline calibration unit is used to calculate the current features. Compared with the user's personal historical baseline offset : ; in, The standard deviation of the user's historical data over the past 90 days; The value ranges from 0 to 3, with larger values ​​indicating more significant deviations from the baseline; an offset threshold is set. ,like Mark it as an abnormal offset and trigger key monitoring; The group reference calibration unit incorporates user characteristics and typical patterns of homogeneous groups. Comparative group consistency factor : ; in, The cosine similarity function is used. The closer the value is to 1, the higher the degree of fit with the group pattern; The dynamic health profile Features after dual calibration are scaled to a personalized ratio Combined into: ; in, It is a unit vector. Weights are assigned based on individual differences and can be dynamically adjusted according to users' health management needs; The portrait dynamic update unit collects new health data every 24 hours and recalculates. and Update dynamic health profile This ensures that the profile is consistent with the user's real-time health status.

[0028] The personalized dynamic health profile construction module calculates the deviation between the current health characteristics and the user's historical baseline through the individual baseline calibration unit, accurately captures abnormal deviations in the individual's health status and triggers key monitoring, providing individual benchmark support for personalized analysis; The group reference calibration unit calculates the group consistency factor using cosine similarity, taking into account the fit between user characteristics and typical patterns of homogeneous groups, avoiding bias caused by judging solely based on individual baselines. The profile dynamic update unit recalibrates relevant parameters and updates the profile every 24 hours based on new data. The dynamic health profile generated by combining personalized proportions with dual calibration features fully preserves individual health differences while incorporating common group reference dimensions, achieving accurate and real-time mapping of users' physiological homeostasis, behavioral patterns, and psychological tendencies, providing high-quality core input for the stratified risk warning and root cause inference modules.

[0029] In this embodiment, the aforementioned case of a 35-year-old male user is used. This user's unified health status vector is output by the multimodal health data dynamic acquisition and fusion module. (128 dimensions, feature mean 0.72) The personalized dynamic health profile construction module first retrieves the individual's historical data from the past 90 days through the individual baseline calibration unit to calculate the individual's historical baseline. (128 dimensions, feature mean 0.68), standard deviation of historical data over the past 90 days According to the formula The calculated offset is: ; because Mark it as an abnormal offset and trigger key monitoring; The group reference calibration unit was matched with typical group patterns that were homogeneous with the user (35-year-old male, no history of chronic diseases, and light daily exercise). (128 dimensions, feature mean 0.70), calculate the population consistency factor using the cosine similarity function: ; This indicates that their current health characteristics are highly consistent with the patterns of their homogeneous group; based on the user's health management needs (focusing on fluctuations in individual physiological indicators), a personalized ratio is set. unit vector For a 128-dimensional vector of all ones, according to the formula: Calculated dynamic health profile (128 dimensions, feature mean 0.65); the profile dynamic update unit collects new health data at the same time the following day and recalculates... (Returned to normal range) Update based on new parameters This ensures that the profile remains consistent with the user's real-time health status.

[0030] More specifically, the hierarchical risk warning and root cause inference module includes a shared underlying encoder, a multi-dimensional risk prediction unit, a root cause tracing unit, and a hierarchical warning unit; The shared underlying encoder is used to create dynamic health profiles. Perform feature dimensionality reduction and enhancement to output a low-dimensional dense feature vector. ; The multi-dimensional risk prediction unit contains multiple independent risk prediction heads, used to calculate the first... Probability of health risks : ; in, This is the Sigmoid function, with an output range of 0-1; For the first Parameters of a risk prediction head; The trend score of this type of risk-related indicator within a 7-day time window (the value ranges from 0 to 1, and the larger the value is for a more obvious upward trend). The trend influence coefficient; all risk probabilities constitute a risk probability vector. ; The root cause tracing unit, for those identified as high-risk ( , For events involving gradient-weighted class activation mapping, a gradient-weighted method is used to calculate each feature dimension in the health profile. Contribution to the prediction of this risk : ; The top contributors Features ( ) and their corresponding data sources are output as key root causes; The tiered early warning unit sets three risk thresholds: low risk... Medium risk High risk Each triggers a different warning action: low risk only logs the information, medium risk sends a health alert, and high risk is linked to the medical service interface and notifies the guardian.

[0031] The hierarchical risk warning and root cause inference module reduces the dimensionality and enhances the features of dynamic health profiles through a shared underlying encoder, effectively eliminating redundant information and strengthening key health features, providing high-quality, low-dimensional input for risk prediction. The multi-dimensional risk prediction unit uses an independent prediction head combined with a 7-day trend score to achieve accurate probability quantification of multiple types of health risks, taking into account both the current state and the impact of changing trends. The root cause tracing unit uses a gradient-weighted activation mapping method to locate core impact features and data sources for high-risk events, solving the pain point of traditional early warnings that only know the risk but not the root cause. The graded early warning unit triggers differentiated response actions based on three-level risk thresholds, realizing hierarchical management of health risks. This avoids over-warning of low-risk events while ensuring timely handling of high-risk events, providing accurate and interpretable risk basis for the generation of subsequent adaptive personalized intervention strategies.

[0032] In this embodiment, the aforementioned case of a 35-year-old male user is used, and the updated dynamic health profile of this user is as follows. (128-dimensional) Input hierarchical risk warning and root cause inference module, share the bottom encoder and use a 3-layer fully connected network to perform feature dimensionality reduction and enhancement, output 64-dimensional low-dimensional dense feature vector (The mean value of the feature is 0.62); The multi-dimensional risk prediction unit includes three independent risk prediction heads: hypertension, diabetes, and cardiovascular disease, with a set trend influence coefficient. The system retrieves the user's heart rate, blood pressure, blood sugar, and other relevant indicators from the past 7 days to calculate a trend score, including the trend score for hypertension-related indicators. (Showing a clear upward trend), trend scores of diabetes-related indicators (Stable trend), trend score of cardiovascular disease related indicators (Slight upward trend), first head parameter for predicting hypertension risk: .

[0033] More specifically, the adaptive personalized intervention strategy generation and closed-loop optimization module includes a strategy matching engine, an intervention execution tracking unit, and a closed-loop learning unit; The policy matching engine employs a constraint optimization-based policy matching algorithm to obtain the optimal intervention plan by solving the following formula. : ; in, A set of feasible strategies that comply with current medical guidelines and user contraindications; This is the expected function for the intervention effect, with a value range of 0-1; the better the effect, the larger the value. The user is assigned a burden cost function, with a value ranging from 0 to 1; the lighter the burden, the smaller the value. A personalized moderating factor characterizing user compliance levels; the lower the compliance, the lower the compliance level. The larger the value; The intervention execution tracking unit is used to collect the execution data of the intervention plan in real time. The specific steps are as follows: S21. Receive the execution progress feedback from the user (fully executed / partially executed / not executed); S22. Calculate the execution completion rate. , ; S23, if Analyze the reasons for non-implementation (too difficult / time conflict, etc.) and generate a simplified intervention plan; S24. Record user feedback and ratings during the execution process. This will serve as a basis for subsequent strategy optimization. The closed-loop learning unit constructs an intervention-feedback experience pool and stores quadruples. ,in To intervene in the feedback process (including the completion rate) Feedback rating ), This represents the change in risk probability; the calibration parameters for the personalized dynamic health profile construction module are applied every 7 days using experience pool data. Prediction header parameters of the hierarchical risk warning and root cause inference module Perform coordinated fine-tuning.

[0034] The adaptive personalized intervention strategy generation and closed-loop optimization module uses the constraint optimization algorithm of the strategy matching engine to balance the expected intervention effect with the user's execution burden, generating an optimal plan that suits individual compliance, while adhering to medical guidelines and user contraindications. The intervention execution tracking unit monitors the execution progress and feedback in real time, dynamically simplifying plans with low completion rates to ensure the feasibility of the intervention. The closed-loop learning unit builds an intervention-feedback experience pool and periodically fine-tunes the core parameters of the preceding modules, forming a complete closed loop of plan generation, execution tracking, and feedback optimization. This ensures that the intervention strategy continuously iterates with the user's health status and execution performance, significantly improving the personalization and long-term intervention effect of health management.

[0035] In this embodiment, the aforementioned case of a 35-year-old male user is used. The user's hypertension risk warning result (risk probability 0.65) and root cause information (high-salt diet, lack of exercise) are input into this module.

[0036] The strategy matching engine filters the set of feasible strategies. The regimen included a daily low-salt diet (≤5g salt), 30 minutes of aerobic exercise 5 times a week, and daily blood pressure monitoring at 8 AM. All three measures complied with hypertension management guidelines and avoided contraindications related to seafood allergies. Combined with the user's previous moderate adherence to interventions (personalized adjustment factor), the regimen was effective. The expected intervention effects of each strategy were assessed. The execution cost is 0.82, 0.85, and 0.78 respectively. The values ​​are 0.3, 0.42, and 0.25 respectively, according to the formula. The calculated target values ​​are: 0.82-0.5×0.3=0.67, 0.85-0.5×0.42=0.64, and 0.78-0.5×0.25=0.655. A low-salt diet (≤5g salt per day) is selected as the optimal intervention program.

[0037] The intervention execution tracking unit continuously receives user feedback, and on day 7, it calculates the execution completion rate based on the 5-day execution period. User feedback rating A score (indicating moderate difficulty); the closed-loop learning unit will use quadruples. Store in the experience pool; 7 days later, retrieve the data from the experience pool and apply the calibration parameters of the personalized dynamic health profile building module. The parameter was fine-tuned from 0.82 to 0.84, which is the head parameter for hypertension prediction in the stratified risk warning and root cause inference module. Fine-tuning from 0.51 to 0.50 The value was fine-tuned from 0.15 to 0.14 to achieve closed-loop optimization of system parameters.

[0038] More specifically, the feature fusion unit of the multimodal health data dynamic acquisition and fusion module also supports dynamic iterative optimization of the fusion weights, with the following specific steps: S31. Set the fusion validity threshold. ; S32. Calculate the unified health status vector after fusion. effectiveness coefficient The formula is as follows: ; in, For data source noise interference, For data integrity; S33, if Maintain current weights ; S34, if Calculate the weight adjustment coefficient Update weights Recalculate ; S35. Repeat steps S32-S34 until... The final unified health state vector is output.

[0039] The dynamic iterative optimization mechanism for the fusion weights of the feature fusion units in the multimodal health data dynamic acquisition and fusion module is implemented by setting a fusion effectiveness threshold. Combined with data source noise interference With data integrity Calculate the effectiveness coefficient The fusion weights that do not reach the threshold are dynamically adjusted; this is done through weight adjustment coefficients. By increasing the proportion of high-quality data sources (high integrity, low noise) and iteratively optimizing until the effectiveness requirements are met, the problem of static weights being unable to adapt to dynamic fluctuations in data source quality is effectively solved, significantly improving the unified health state vector. The accuracy and reliability of the data provide higher-quality data support for the subsequent construction of personalized dynamic health profiles and risk warnings.

[0040] In this embodiment, the aforementioned case of a 35-year-old male user is used. The core data source for this user, after data quality verification, includes: heart rate data ( , , ), motion trajectory data ( , , ), sleep quality score data ( , , ), and the remaining auxiliary data sources in total (average ,average ).

[0041] According to the formula Calculate the initial effectiveness coefficient; molecule = ; Denominator = ,because This triggers iterative optimization of weights.

[0042] Calculate the weight adjustment factor ; in, ,have to , , , Update weights After normalization, , , , .

[0043] Recalculate: ; Stop iterating and output the final unified health state vector. .

[0044] More specifically, the individual baseline calibration unit of the personalized dynamic health profile construction module also supports dynamic updates of historical baselines, with the following specific rules: The user's personal historical baseline is recalculated every 180 days. The mean and standard deviation of health data over the past 180 days were used; If a user's health status undergoes a significant change (such as being diagnosed with a chronic disease or recovering from surgery), an emergency baseline update is triggered, rebuilding the baseline based on intensive monitoring data from the 30 days following the surgery / diagnosis. ; After the baseline is updated, the personalization ratio is recalculated. ,in The standard deviation of the new baseline ensures that the calibration mechanism adapts to changes in the user's health status.

[0045] The personalized dynamic health profile construction module's individual baseline calibration unit features a dynamic historical baseline update mechanism. This mechanism recalculates the individual's historical baseline every 180 days based on the past 180 days of health data, and urgently reconstructs the baseline based on 30 days of intensive monitoring data when the user's health status undergoes significant changes. This ensures the baseline always accurately reflects the user's current health level and avoids calculation errors caused by outdated baselines. Simultaneously, it utilizes personalized proportions... The coordinated adjustments allow individual differences in weighting to adapt to new health status fluctuations, ensuring that the calibration mechanism of the dynamic health profile continuously aligns with changes in user health, thereby further improving the accuracy and dynamic adaptability of the profile in depicting individual health status.

[0046] In this embodiment, the aforementioned case of a 35-year-old male user is used. This user had previously received intervention due to moderate risk of hypertension, and was diagnosed with essential hypertension (a significant change in health status) 3 months later. The system then triggered an emergency baseline update.

[0047] Based on intensive monitoring data (blood pressure, heart rate, and other indicators three times a day) for 30 days after diagnosis, an individual's historical baseline was reconstructed. The mean (systolic blood pressure 135 mmHg, diastolic blood pressure 85 mmHg) and standard deviation of the new baseline were calculated. (Original baseline standard deviation) ); due to the user's previous personalization ratio According to the formula ,by The constraints were ultimately adjusted to After the update, the individual baseline calibration unit is based on the new baseline. The dynamic health profile is generated by calculating the offset, combining the group reference calibration unit with a new personalized scaling fusion feature. It more accurately reflects the user's health status after being diagnosed with hypertension, providing a more adaptable core input for the generation of subsequent risk warning and intervention strategies.

[0048] More specifically, the root cause tracing unit of the hierarchical risk warning and root cause inference module also supports root cause verification and secondary tracing mechanisms: For the output before One key root cause was identified, and the correlation between these root causes was verified by combining the user's historical health data, and the correlation degree was calculated. ; like Eliminate the root cause and add the next-order contribution feature; Relevance retention after secondary source tracing The root causes are identified, and a final list of root causes is generated, with the correlation strength marked.

[0049] The newly added root cause verification and secondary tracing mechanisms in the root cause tracing unit calculate the correlation between key root causes and users' historical risk events. The system performs validity screening on the first N root causes of the initial output, removing false or weakly related root causes with a correlation score below 0.3 and supplementing them with high contribution features. Finally, it retains the root causes with high correlation scores and labels the correlation strength. This effectively solves the problem of root cause misjudgment that is prone to occur when relying solely on contribution score, significantly improves the reliability and interpretability of root cause tracing results, and provides clear and credible root cause basis for the accurate generation of subsequent adaptive personalized intervention strategies.

[0050] In this embodiment, the aforementioned case of a 35-year-old male user is used. One month after the user was diagnosed with essential hypertension, the system monitored his hypertension risk probability. (High risk) The root cause tracing unit initiates initial tracing and uses the gradient weighted class activation mapping method to calculate the top 3 key root causes in terms of contribution: Feature 1 (daily salt intake, contribution 0.62), Feature 2 (daily exercise duration, contribution 0.28), and Feature 3 (number of nighttime sleep apnea, contribution 0.10).

[0051] Subsequently, the root cause verification mechanism was activated, and the user's historical health data for the past two years was retrieved. It was found that there were a total of 20 high-risk events for hypertension in the past: feature 1 appeared 16 times with historical risk events, feature 2 appeared 12 times, and feature 3 appeared 5 times. According to the formula Calculate the correlation degree; , , .

[0052] because Remove feature 3 and add the next contribution feature (feature 4: morning resting heart rate, contribution 0.08); verify the correlation of feature 4: it co-occurred with historical risk events 7 times. The second source tracing has been completed.

[0053] The final root cause list is generated as follows: Feature 1 (daily salt intake, high association strength). Feature 2 (daily exercise duration, medium correlation strength) Feature 4 (morning resting heart rate, low to medium correlation strength) (and synchronize it to the intervention strategy generation module).

[0054] More specifically, the personalized AI intelligent health management system based on multi-source data fusion also includes a system initialization and user modeling module; The module is used to perform the following operations during the system startup phase, based on the user-completed health questionnaire, initial physical examination report, and 7-day intensive monitoring data: S41. Establish the user's initial health baseline. Calculate the initial standard deviation ; S42. Based on user age, gender, medical history, and other information, match typical patterns of homogeneous groups. Initialize the group consistency factor calculation parameters; S43. Construct user risk profiles, marking high-risk disease types and contraindicated intervention items; S44. Initialize the personalized dynamic health profile building module. The prediction header parameters of the hierarchical risk warning and root cause inference module provide the initial configuration for system operation.

[0055] The system initialization and user modeling module serves as the foundational startup unit for system operation. By integrating user health questionnaires, initial physical examination reports, and 7-day intensive monitoring data, it establishes the initial health baseline and standard deviation, matches typical patterns of homogeneous groups, constructs risk propensity profiles, and initializes key parameters for core modules. This provides standardized initial configurations tailored to individual user characteristics for subsequent modules such as personalized dynamic health profile construction, stratified risk warning, and root cause inference. It effectively avoids analytical biases caused by a lack of user-specific data support during the system startup phase, ensuring that the entire health management process accurately adapts to users' basic health characteristics from the outset, laying a solid foundation for subsequent dynamic optimization and personalized management.

[0056] In this embodiment, taking a 35-year-old male new user as an example, during the system startup phase, the user fills out a health questionnaire through the system APP, entering information such as age, gender, no history of chronic diseases, family history of hypertension in the father, daily light exercise, and seafood allergy; uploads physical examination reports from the past 3 months, including core indicators such as average systolic blood pressure of 125 mmHg, average diastolic blood pressure of 80 mmHg, and fasting blood glucose of 5.6 mmol / L; and completes 7 days of intensive monitoring while wearing a smartwatch, collecting data such as an average daily resting heart rate of 72 beats / min, an average daily exercise duration of 25 minutes, and an average nighttime sleep duration of 7.2 hours. The module performs initialization operations step by step: S41 establishes an initial health baseline based on 7 days of intensive monitoring data and core physiological indicators from physical examination reports. (Systolic blood pressure 125 mmHg, diastolic blood pressure 80 mmHg, resting heart rate 72 beats / min, etc.), calculate the initial standard deviation of the blood pressure dimension. ; S42 matches typical patterns of similar groups based on the user's characteristics: 35 years old, male, no history of chronic diseases, and light daily exercise. (This group consists of men aged 30-40 without chronic diseases, with a mean systolic blood pressure of 123 mmHg, a mean diastolic blood pressure of 79 mmHg, an average daily exercise duration of 30 minutes, and a sleep duration of 7.5 hours.) Initialize the basic parameters of the cosine similarity function for calculating the group consistency factor. S43, combined with family history of hypertension, construct a risk profile, mark the high-risk disease type as hypertension, and mark the contraindications for intervention as dietary recommendations containing seafood and related allergen intervention measures; S44, Initialize the personalized dynamic health profile building module. Initialize the initial parameters of the three risk prediction heads (hypertension, diabetes, and cardiovascular disease) in the hierarchical risk warning and root cause inference module (e.g., hypertension prediction head). , Diabetes prediction head , This completes all initial system configurations, supporting the formal operation of subsequent modules.

[0057] More specifically, the system supports specialized optimization for the management of specific chronic diseases (such as hypertension and diabetes). In the specialized optimization mode: The multimodal health data dynamic acquisition and fusion module supplements the collection of disease-specific data (daily average blood pressure, blood glucose fluctuation curves, etc.) to enhance the fusion weight of this type of data. To 1.2-1.5 times the usual weight; The personalized dynamic health profile building module strengthens disease control indicators in the physiological homeostasis dimension (such as blood pressure control targets and blood glucose target achievement rates) and adjusts the personalization ratio. ; The tiered risk warning and root cause inference module adds a specific prediction task for the risk of disease complications and optimizes the trend influence coefficient of the corresponding risk prediction head. ; The adaptive personalized intervention strategy generation and closed-loop optimization module's strategy matching engine prioritizes intervention plans recommended by chronic disease management guidelines, reducing the execution burden and cost functions. weight This will improve user compliance.

[0058] This specialized optimization model for specific chronic diseases involves targeted adjustments to the parameters and functions of four core modules: multimodal health data collection, personalized health profile construction, stratified risk warning and root cause inference, and adaptive intervention strategy generation. This strengthens the collection and utilization of chronic disease-specific data, the precise characterization of disease control indicators, the specialized early warning of complication risks, and the generation of guideline-adaptive intervention plans. It effectively addresses the shortcomings of traditional health management systems in adapting to personalized chronic disease management, significantly improving the accuracy, targeting, and user compliance of health management for chronic disease patients, and providing strong support for the long-term standardized management of chronic diseases.

[0059] In this embodiment, using the aforementioned case of a 35-year-old male user with hypertension, after the user is diagnosed with primary hypertension, the system automatically switches to the hypertension-specific optimization mode. The multimodal health data dynamic acquisition and fusion module supplements the collection of disease-specific data such as daily average blood pressure, 24-hour blood pressure fluctuation curves, and morning blood pressure peaks, and applies the conventional fusion weights to this type of data. Increased to 1.3 times (i.e.) The personalized dynamic health profile construction module strengthens disease control indicators in the physiological homeostasis dimension, such as systolic blood pressure control targets (<130 mmHg) and diastolic blood pressure control targets (<80 mmHg), and adjusts the personalization ratio. The tiered risk warning and root cause inference module adds specific prediction tasks for the risks of two complications: hypertensive nephropathy and stroke, and optimizes the trend influence coefficient of the corresponding risk prediction heads. .

[0060] The adaptive personalized intervention strategy generation and closed-loop optimization module's strategy matching engine prioritizes intervention protocols recommended in the "Guidelines for the Prevention and Treatment of Hypertension" and sets an execution burden cost function. weight The optimal intervention plan was selected based on "daily low-salt diet (≤5g), 30 minutes of moderate-intensity exercise (brisk walking) daily, and blood pressure monitoring after taking medication at 7 am".

[0061] Because the solution aligns with the guidelines and reduces the burden of implementation, the user intervention completion rate within one month is high. Feedback rating The score (out of 5) reduced the core risk probability of hypertension from 0.65 to 0.48, and improved the accuracy of complication risk prediction to 0.88 compared with the conventional model, achieving precise and long-term management of hypertension.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A personalized AI-powered intelligent health management system that integrates multi-source data, characterized in that: It includes a multimodal health data dynamic acquisition and fusion module, a personalized dynamic health profile construction module, a hierarchical risk warning and root cause inference module, an adaptive personalized intervention strategy generation and closed-loop optimization module, and a privacy protection and federated learning module, which are connected in sequence. The multimodal health data dynamic acquisition and fusion module is used to collect multidimensional health data through smart terminals, environmental sensors, and manual input interfaces, and generate a unified health state vector through spatiotemporal alignment and feature-level fusion. Output to the personalized dynamic health profile building module; The personalized dynamic health profile construction module adaptively calibrates based on users' historical baseline data and common group patterns, outputting a dynamic health profile that includes physiological homeostasis, behavioral patterns, and psychological tendencies. The data is then transmitted to the hierarchical risk warning and root cause inference module. The hierarchical risk warning and root cause inference module calculates multi-dimensional health risk probabilities in parallel. For high-risk events, explainable root cause tracing is performed, and the early warning results and root cause information are output to the adaptive personalized intervention strategy generation and closed-loop optimization module. The adaptive personalized intervention strategy generation and closed-loop optimization module generates a personalized comprehensive intervention plan based on early warning and root cause information. Tracking intervention feedback and changes in health status to form closed-loop data for iterative optimization of preceding modules; The privacy protection and federated learning module is deployed between local devices and cloud servers to achieve local feature extraction and cloud-based global model aggregation, ensuring data privacy and security.

2. The personalized AI intelligent health management system based on multi-source data fusion according to claim 1, characterized in that, The multimodal health data dynamic acquisition and fusion module includes a data acquisition unit, a data quality verification unit, and a feature fusion unit; The data acquisition unit is used to collect physiological signal data, behavioral and environmental data, and subjective report data to form a multi-source raw dataset; The data quality verification unit is used to perform AI-driven quality screening on the original dataset. The specific steps are as follows: S11. Parse the formats of each data source and distinguish between real-time streaming data and offline data; S12. Calculate the signal integrity of each data source. With noise interference ; S13, Set quality threshold , ,like or This triggers a data re-collection or supplementary collection mechanism; S14. Perform spatiotemporal alignment on the data that passes the verification to unify the timestamp and data dimension; The feature fusion unit employs an adaptive weighted fusion algorithm based on attention weights to calculate the first... Feature vectors of data sources Fusion weights : ； in, The real-time signal-to-noise ratio of the data source, with a value range of 0-50dB. The value ranges from 0 to 1, representing the relevance to the current health context. , For adjustable hyperparameters, , This is used to balance the effects of signal-to-noise ratio and correlation. Total number of data sources; The unified health status vector It is generated by the feature fusion unit through the following formula: ； And it must meet the fusion effectiveness coefficient. Otherwise, return to the feature fusion unit to readjust the weights.

3. The personalized AI intelligent health management system based on multi-source data fusion according to claim 2, characterized in that, The personalized dynamic health profile construction module includes an individual baseline calibration unit, a group reference calibration unit, and a profile dynamic update unit. The individual baseline calibration unit is used to calculate the current features. Compared with the user's personal historical baseline offset : ； in, The standard deviation of the user's historical data over the past 90 days; The value ranges from 0 to 3, with larger values ​​indicating more significant deviations from the baseline; an offset threshold is set. ,like Mark it as an abnormal offset and trigger key monitoring; The group reference calibration unit incorporates user characteristics and typical patterns of homogeneous groups. Comparative group consistency factor : ； in, The cosine similarity function is used. The closer the value is to 1, the higher the degree of fit with the group pattern; The dynamic health profile Features after dual calibration are scaled to a personalized ratio Combined into: ； in, It is a unit vector. Weights are assigned based on individual differences and can be dynamically adjusted according to users' health management needs; The portrait dynamic update unit collects new health data every 24 hours and recalculates. and Update dynamic health profile This ensures that the profile is consistent with the user's real-time health status.

4. The personalized AI intelligent health management system based on multi-source data fusion according to claim 3, characterized in that, The hierarchical risk warning and root cause inference module includes a shared underlying encoder, a multi-dimensional risk prediction unit, a root cause tracing unit, and a hierarchical warning unit. The shared underlying encoder is used to create dynamic health profiles. Perform feature dimensionality reduction and enhancement to output a low-dimensional dense feature vector. ; The multi-dimensional risk prediction unit contains multiple independent risk prediction heads, used to calculate the first... Probability of health risks : ； in, This is the Sigmoid function, with an output range of 0-1; For the first Parameters of a risk prediction head; The trend score of this type of risk-related indicator within a 7-day time window (the value ranges from 0 to 1, and the larger the value is for a more obvious upward trend). The trend influence coefficient; all risk probabilities constitute a risk probability vector. ; The root cause tracing unit, for those identified as high-risk ( , For events involving gradient-weighted class activation mapping, a gradient-weighted method is used to calculate each feature dimension in the health profile. Contribution to the prediction of this risk : ； The top contributors Features ( ) and their corresponding data sources are output as key root causes; The tiered early warning unit sets three risk thresholds: low risk... Medium risk High risk Each triggers a different warning action: low risk only logs the information, medium risk sends a health alert, and high risk is linked to the medical service interface and notifies the guardian.

5. A personalized AI intelligent health management system based on multi-source data fusion according to claim 4, characterized in that, The adaptive personalized intervention strategy generation and closed-loop optimization module includes a strategy matching engine, an intervention execution tracking unit, and a closed-loop learning unit. The policy matching engine employs a constraint optimization-based policy matching algorithm to obtain the optimal intervention plan by solving the following formula. : ； in, A set of feasible strategies that comply with current medical guidelines and user contraindications; This is a function representing the expected intervention effect, with a value range of 0-1; Execute a cost function for the user, with a value range of 0-1; A personalized moderating factor characterizing user compliance levels; the lower the compliance, the lower the compliance level. The larger the value; The intervention execution tracking unit is used to collect the execution data of the intervention plan in real time. The specific steps are as follows: S21. Receive the execution progress feedback from the user client; S22. Calculate the execution completion rate. , ; S23, if Analyze the reasons for non-implementation and generate a simplified intervention plan; S24. Record user feedback and ratings during the execution process. This will serve as a basis for subsequent strategy optimization. The closed-loop learning unit constructs an intervention-feedback experience pool and stores quadruples. ,in To intervene in the implementation feedback, This represents the change in risk probability; the calibration parameters for the personalized dynamic health profile construction module are applied every 7 days using experience pool data. Prediction header parameters of the hierarchical risk warning and root cause inference module Perform coordinated fine-tuning.

6. A personalized AI intelligent health management system based on multi-source data fusion according to claim 5, characterized in that, The feature fusion unit of the multimodal health data dynamic acquisition and fusion module also supports dynamic iterative optimization of the fusion weights, with the specific steps as follows: S31. Set the fusion validity threshold. ; S32. Calculate the unified health status vector after fusion. effectiveness coefficient The formula is as follows: ； in, For data source noise interference, For data integrity; S33, if Maintain current weights ; S34, if Calculate the weight adjustment coefficient Update weights Recalculate ; S35. Repeat steps S32-S34 until... The final unified health state vector is output.

7. A personalized AI intelligent health management system based on multi-source data fusion according to claim 6, characterized in that, The individual baseline calibration unit of the personalized dynamic health profile construction module also supports dynamic updates of historical baselines, with the following specific rules: The user's personal historical baseline is recalculated every 180 days. The mean and standard deviation of health data over the past 180 days were used; If a user's health status undergoes a significant change, triggering an emergency baseline update, the baseline will be reconstructed based on intensive monitoring data from the 30 days following surgery / diagnosis. ; After the baseline is updated, the personalization ratio is recalculated. ,in The standard deviation of the new baseline ensures that the calibration mechanism adapts to changes in the user's health status.

8. A personalized AI intelligent health management system based on multi-source data fusion according to claim 7, characterized in that, The root cause tracing unit of the hierarchical risk warning and root cause inference module also supports root cause verification and secondary tracing mechanisms: For the output before One key root cause was identified, and the correlation between these root causes was verified by combining the user's historical health data, and the correlation degree was calculated. ; like Eliminate the root cause and add the next-order contribution feature; Relevance retention after secondary source tracing The root causes are identified, and a final list of root causes is generated, with the correlation strength marked.

9. A personalized AI intelligent health management system based on multi-source data fusion according to claim 8, characterized in that, The personalized AI-powered intelligent health management system, which integrates multi-source data, also includes a system initialization and user modeling module. The module is used to perform the following operations during the system startup phase, based on the user-completed health questionnaire, initial physical examination report, and 7-day intensive monitoring data: S41. Establish the user's initial health baseline. Calculate the initial standard deviation ; S42. Based on user age, gender, medical history, and other information, match typical patterns of homogeneous groups. Initialize the group consistency factor calculation parameters; S43. Construct user risk profiles, marking high-risk disease types and contraindicated intervention items; S44. Initialize the personalized dynamic health profile building module. The prediction header parameters of the hierarchical risk warning and root cause inference module provide the initial configuration for system operation.

10. A personalized AI intelligent health management system based on multi-source data fusion according to claim 9, characterized in that, The system supports targeted optimization for the management of specific chronic diseases. Under the targeted optimization mode: The multimodal health data dynamic acquisition and fusion module supplements the collection of disease-specific data, enhancing the fusion weight of this type of data. To 1.2-1.5 times the usual weight; The personalized dynamic health profile building module strengthens disease control indicators in the physiological homeostasis dimension and adjusts the personalization ratio. ; The tiered risk warning and root cause inference module adds a specific prediction task for the risk of disease complications and optimizes the trend influence coefficient of the corresponding risk prediction head. ; The adaptive personalized intervention strategy generation and closed-loop optimization module's strategy matching engine prioritizes intervention plans recommended by chronic disease management guidelines, reducing the execution burden and cost functions. weight This will improve user compliance.

Images