Risk assessment and personalized scheme generation method based on multi-source health data

Abstract

The invention belongs to the technical field of health management, and particularly relates to a risk assessment and personalized scheme generation method based on multi-source health data. The method comprises the following steps: collecting health data of a user; cleaning, converting and standardizing the collected health data, and constructing a user health portrait and derivative features; inputting the processed health data into a pre-trained artificial intelligence model cluster which comprises a stacked integrated learning model and an interpretable random forest model which are respectively used for critical disease risk probability assessment and specific health risk classification assessment; dividing risk grade categories according to the risk assessment result; based on the health data and the risk level category, generating a personalized health intervention scheme by retrieving the enhancement generation architecture; and dynamically updating a risk assessment and intervention scheme according to follow-up visit data to form a self-adaptive health management closed loop.

Description

Technical Field

[0001] This invention belongs to the field of health management technology, specifically relating to a method for risk assessment and personalized solution generation based on multi-source health data. Background Technology

[0002] In existing technologies, health risk assessments typically rely on single physical examination or questionnaire data, lacking the integration of multi-source data, leading to incomplete assessment results. Traditional risk assessment models often suffer from black-box problems, poor interpretability, and difficulty in providing feature importance rankings. Health intervention plans are mostly generated based on generic templates, lacking personalized adaptation and prone to producing recommendations that conflict with the user's actual health condition. Furthermore, existing systems lack effective security verification mechanisms and dynamic update capabilities, failing to form a continuously optimized health management loop. Summary of the Invention

[0003] The purpose of this invention is to provide a risk assessment and personalized solution generation method based on multi-source health data. This method can comprehensively collect and process users' health data, use artificial intelligence models to conduct accurate risk assessments, and generate safe and personalized health intervention plans based on a retrieval-enhanced generation architecture, thereby achieving a dynamic closed loop of health management.

[0004] This invention provides a method for risk assessment and personalized treatment plan generation based on multi-source health data, the method comprising:

[0005] Collect user health data, which includes user basic information, health assessment data, health checkup data, and health questionnaire data; among which,

[0006] The health assessment data includes human body composition and morphological indicators, cardiovascular function and hemodynamic indicators, lung function indicators, and bone health indicators collected through non-invasive testing equipment.

[0007] The health checkup data includes biochemical indicators, routine examination data, and imaging and functional examination results;

[0008] The health questionnaire data includes users' lifestyle, medical history, and symptom information;

[0009] The collected health data is cleaned, transformed, and standardized, and user health profiles are built based on the processed health data. Derivative features, including trend features and cross features, are also constructed, and the processed data and health profiles are securely stored.

[0010] The processed health data is input into a pre-trained cluster of artificial intelligence models, which includes stacked ensemble learning models and interpretable random forest models.

[0011] The stacked ensemble learning model is used to assess the probability of a user's major disease risk based on the health data;

[0012] The interpretable random forest model is used to classify and assess a user's specific health risks based on the health data, and outputs a ranking of the importance of the top input features that contribute the most to the prediction.

[0013] Based on the risk assessment results, users' proactive health management goals are categorized into different risk level categories, including:

[0014] Category 1 Management Objective: Major disease risk assessment results indicate high risk;

[0015] The second category of management objectives: those whose major disease risk assessment results are of medium risk or who have sub-health issues;

[0016] The third category of management objectives: The risk assessment result for major diseases is low risk and there are no sub-health issues;

[0017] Based on users' health data and risk level categories, personalized health intervention plans are generated through a retrieval-enhanced generation architecture. The generation steps include:

[0018] The user's health profile is converted into a query vector, and an approximate nearest neighbor search is performed in a pre-built vector knowledge base to recall relevant knowledge fragments;

[0019] Based on the key entities in the health profile, graph queries are performed in a pre-constructed structured knowledge graph to retrieve relevant medical rules and constraints.

[0020] The retrieved knowledge fragments, medical rules and constraints, the user's complete health profile, preset output format requirements, and role definition instructions are combined to form structured prompts;

[0021] The structured prompts are input into a large language model that has been fine-tuned in the field of health management to generate a preliminary health intervention plan text;

[0022] The generated preliminary solutions are subjected to security and logical verification, including parsing and cross-checking the solution text using logical rules derived from the knowledge graph. If a suggestion in the solution conflicts with the logical rules, the suggestion is marked as unacceptable, and a regeneration or correction mechanism is triggered for the unacceptable suggestion until all suggestions pass the verification.

[0023] The risk assessment results and health intervention plans are updated based on the user's follow-up data to achieve a dynamic closed loop of health management.

[0024] The technical advantages of this invention are as follows:

[0025] By integrating multi-source health data and applying artificial intelligence model clusters, accurate assessments of users' major disease risks and specific health risks are achieved; the personalized and scientific nature of the generated health intervention plans is ensured through retrieval-enhanced generation architecture and security verification mechanisms; and an adaptive health management closed loop is formed through dynamic updates of follow-up data, improving the efficiency and effectiveness of health management. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the health management system architecture.

[0027] Figure 2 This is a schematic diagram of the artificial intelligence engine architecture. Detailed Implementation

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0029] Figure 1 This is a schematic diagram of the health management system architecture. This architecture is user-centric, building an ecosystem from comprehensive data perception to intelligent decision-making and personalized services. Specifically, it comprises four organically cooperating components: a data acquisition layer, a data processing and storage layer, an artificial intelligence engine core layer, and a service application layer.

[0030] 1. Data Acquisition Layer:

[0031] This layer serves as the perception foundation and data source for the entire system, and its design goal is to construct a holographic digital profile of the user's health status. The system systematically aggregates three key data types through standardized interfaces and protocols, which constitute the starting point for the execution of the method in this invention: the collection of basic user information and the collection of user health data.

[0032] Basic user information collection includes name, gender, age, employer, and contact information.

[0033] User health data collection includes the following three categories:

[0034] Health assessment data: Collected using a series of core non-invasive testing devices integrated into this methodology, aiming to capture real-time, objective physiological and functional parameters of the user. The collected data includes:

[0035] Human body composition and morphological indicators: height and weight, body fat percentage, muscle mass of limbs and trunk, waist-to-hip ratio.

[0036] Cardiovascular function and hemodynamic parameters: peripheral blood pressure, pulse pressure, pulse rate, ejection time, left ventricular load index, myocardial perfusion index, subendocardial myocardial viability, central arterial pressure, enhancement index, ankle-brachial index, and pulse wave velocity of the ankle arteries in both arms.

[0037] Lung function indicators: forced vital capacity, forced vital capacity in one second, and forced vital capacity in one second.

[0038] Bone health indicators: bone strength index, T-score, Z-score.

[0039] Health checkup data: As the clinical gold standard, it is securely imported via user-authorized electronic report uploads or application programming interfaces (APIs), providing crucial medical baselines and supporting evidence for risk assessment models. The data includes:

[0040] Biochemical indicators: blood glucose, blood uric acid, and five blood lipids (total cholesterol, triglycerides, high-density lipoprotein, low-density lipoprotein, and lipoprotein a).

[0041] Routine examinations: complete blood count, liver function tests, and kidney function tests.

[0042] Imaging and functional examination results: Chest CT scan results, electrocardiogram results, abdominal and carotid ultrasound results.

[0043] Health questionnaire data: This aims to quantify users' subjective feelings, lifestyles, behavioral patterns, and genetic background. The questionnaire design referenced multiple authoritative national guidelines and scales. The information collected includes:

[0044] Lifestyle and status: eating habits, exercise habits, psychological state, sleep status, and behavioral habits.

[0045] Medical history and symptoms: medical history, family history, and current symptoms and discomfort.

[0046] 2. Data Processing and Storage Layer:

[0047] After the data acquisition layer completes the collection of the massive and heterogeneous data, the data is securely transmitted to this layer for in-depth processing and persistent storage, providing high-quality data for the subsequent core layer of the artificial intelligence engine.

[0048] ETL (Extract, Transform, Load) and Standardized Pipelines: Run automated data extraction-transformation-loading processes to clean and transform data and map it onto a unified medical ontology to ensure semantic consistency of the data.

[0049] Advanced feature engineering: It not only calculates conventional derived indicators, but also constructs trend features and volatility features for time series data; it fuses multimodal data to create more predictive cross features.

[0050] Secure and compliant storage: A hybrid database architecture is adopted to physically isolate and pseudonymize personal identity information and health information. All data is strongly encrypted during transmission and storage.

[0051] 3. Core Layer of the Artificial Intelligence Engine:

[0052] This layer is the brain driving the intelligent operation of the entire system. It deploys a series of advanced algorithm models to perform in-depth analysis, reasoning, and decision-making on the processed high-quality data. The innovation of this layer lies in a cluster of health risk assessment models trained on more than 130,000 real-world health data sets, and an artificial intelligence architecture that ensures the scientific safety of the generated content. Figure 2 This is a schematic diagram of the artificial intelligence engine architecture.

[0053] 3.1. Health risk assessment using multi-model fusion:

[0054] This method employs an advanced model cluster that combines high-performance ensemble learning with interpretable algorithms to achieve accurate assessment of the risk of major diseases and sub-health issues. This model cluster is trained on over 130,000 rigorously cleaned and labeled multi-source health data sets.

[0055] Stacked ensemble learning models for major diseases:

[0056] One of the core innovations of this invention lies in its use of a two-layer stacked ensemble learning framework for risk probability quantification of major diseases with complex causes, such as stroke and cardiovascular disease. The core advantage of this framework is its ability to capture complex linear and nonlinear feature relationships in data from different algorithmic dimensions through heterogeneous base learners, and then perform higher-dimensional nonlinear fitting through meta-learners, thereby achieving generalization performance and prediction accuracy that surpasses that of a single model.

[0057] Stroke risk assessment:

[0058] Input data: Based on user basic information, peripheral blood pressure, pulse pressure, central arterial pressure, waist-to-hip ratio, and pulse wave velocity of the left and right arm ankle arteries from health assessment data, combined with blood glucose, blood lipids, and carotid ultrasound results from health checkup data, as well as family history, symptom and discomfort questionnaire data.

[0059] Model Structure and Training: This model employs a two-layer stacked ensemble learning framework. The first layer is the base learner layer, which deploys multiple heterogeneous models in parallel, including gradient boosting decision trees, logistic regression with L1 and L2 regularization, and random forests. Each base learner is trained independently on the training set and outputs its preliminary probability prediction of stroke risk. The second layer is the meta-learner, which uses the LightGBM (Lightweight Gradient Boosting Machine) model. Its training process involves using the predicted probabilities of all base learners in the first layer on the validation set as a new input feature matrix and fitting it with the true labels of the validation set to learn how to optimally combine and correct the predictions of the base learners. During model inference, new user data first generates first-layer predictions through all trained base learners. These predictions are then concatenated into a new feature vector, which is input into the meta-learner, which outputs the final risk probability.

[0060] Output: Based on the final output risk probability value, the system infers the high, medium, and low risk levels of stroke according to preset clinical thresholds.

[0061] Cardiovascular disease risk assessment:

[0062] Input data: Based on user basic information, height and weight, peripheral blood pressure, pulse pressure difference, ankle-brachial index, and pulse wave velocity of the ankle arteries in both arms from health assessment data, combined with blood lipids, carotid ultrasound results, electrocardiogram results from health check-up reports, as well as medical history and family history questionnaire data.

[0063] Model structure: It adopts the same two-layer stacked ensemble learning framework and training methodology as the stroke risk assessment model. The difference is that its base learner and meta learner are trained independently for the data features and labels of cardiovascular diseases.

[0064] Output: The inference suggests that the risk of developing cardiovascular disease over the next ten years is high, medium, or low.

[0065] Interpretable random forest models for specific risks:

[0066] Another core innovation of this invention lies in solving the black-box problem of traditional artificial intelligence models. We innovatively adopt an interpretable random forest algorithm, which provides high-precision risk classification while providing reliable feature importance ranking.

[0067] Interpretability mechanism: After the model is trained, the system quantifies the contribution of each input feature to the model's prediction results by calculating the average reduction in Gini impurity or ranking importance. The system automatically selects the top few features with the highest contribution and defines them as the key basis for this risk assessment.

[0068] COPD risk assessment:

[0069] Input data: Based on user basic information, forced vital capacity, forced vital capacity in one second, and one-second rate from health assessment data, combined with chest CT scan results from health checkup data and behavioral habits, symptoms, and discomfort data from health questionnaire data.

[0070] Model Structure and Output: An interpretable random forest model is used for construction. The model output includes not only binary classification results of COPD risk as high or low, but also a ranking of the importance of input features. For example, the system can clearly indicate to the user that the risk in this assessment mainly stems from a low forced vital capacity in one second (FVC) and a decrease in forced vital capacity in one second (FVC), thus achieving a transparent interpretation of the risk.

[0071] Fracture risk assessment:

[0072] Input data: Based on user basic information, bone strength index, T-score, and Z-score from health assessment data, combined with dietary habits, exercise habits, behavioral habits, medical history, symptoms, and discomfort data from health questionnaire data.

[0073] Model structure and output: An interpretable random forest model is used to infer the risk of fracture as high or low, and the main sources of risk can be identified, such as insufficient bone strength index.

[0074] Depression and anxiety risk assessment:

[0075] Input data: Based on data on mental state, sleep status, behavioral habits, symptoms, and discomfort from health questionnaires.

[0076] Model structure and output: An interpretable random forest model is used to construct the model, and the inference risk of anxiety or depression is classified as high or low.

[0077] 3.2. Accurate diagnosis of sub-health issues:

[0078] By combining knowledge graphs and clinical standards, a comprehensive analysis of users' multi-dimensional data is conducted to achieve the assessment of sub-health issues.

[0079] Musculoskeletal system problem assessment: The assessment results are obtained based on height and weight, body fat percentage, muscle mass of limbs and trunk, waist-to-hip ratio, bone strength index, T-score, and Z-score.

[0080] Cardiopulmonary system problem assessment: The assessment results are obtained based on five items: peripheral blood pressure, pulse pressure, central arterial pressure, pulse wave velocity of left and right arm ankle arteries, blood glucose, and blood lipids, combined with medical history, family history, symptoms and discomfort questionnaire data.

[0081] Metabolic system problem assessment: The assessment results are obtained based on five parameters: blood glucose, blood uric acid, blood lipids, and abdominal ultrasound.

[0082] Other issues assessment: Assessment results are obtained based on dietary habits, exercise habits, psychological state, and sleep status from the health questionnaire data.

[0083] 3.3. Proactive Health Management Goal Reasoning (Red-Yellow-Green Layering):

[0084] After completing the health risk assessment, the system initiates proactive health management goal inference. It categorizes health impacts and sudden risks into red, yellow, and green categories for management; this is a key step in achieving precise and tiered management.

[0085] Red-category management objectives (Category 1): These include high-risk outcomes identified in major disease risk assessments. The management objective is to promptly remind users to seek medical attention and prevent the occurrence of acute emergencies.

[0086] Yellow-category management objectives (Category II): These include medium-risk cases identified in major disease risk assessments and various clearly defined sub-health issues. The management objective is to remind users to regularly implement AI-generated health intervention plans and to schedule follow-ups.

[0087] Green-category management objectives (Category 3): These include users whose major disease risk assessment results are low-risk and who do not currently have any sub-health issues. The management objective is to encourage users to maintain their current healthy lifestyle.

[0088] 3.4. Solution Generation Driven by Both Knowledge and Data:

[0089] After clarifying the management objectives, the AI-powered health intervention plan generation function is activated. The core innovation of this invention lies in its adoption of an AI architecture that combines retrieval-enhanced generation with security verification. This architecture, through four core steps, ensures that while the plan is highly personalized and uses natural language, the content is absolutely scientific, safe, and reliable, fundamentally eliminating the content illusion problem inherent in large AI language models.

[0090] Semantic Query Formulation and Multi-Source Knowledge Retrieval: The knowledge foundation of this system consists of two parallel knowledge bases. The first is a vector knowledge base, storing a massive amount of authoritative medical guidelines and literature. Its construction process involves segmenting text and then converting it into high-dimensional vectors using a text embedding model. The second is a structured knowledge graph, which defines medical entities, attributes, and the relationships between them. During retrieval, the system converts the user's health profile into a query vector using a text embedding model. Then, it performs an approximate nearest neighbor search in the vector knowledge base to retrieve the most relevant knowledge fragments. Simultaneously, in the knowledge graph, it performs a graph query based on key entities in the user profile to retrieve all relevant medical rules and constraints.

[0091] Dynamic prompt word synthesis: The system dynamically and structurally combines the knowledge fragments retrieved in the previous step and the constraints that must be followed with the user's complete profile, the preset output format, and a role-playing instruction to form a super prompt word that is rich in content and clear in instructions.

[0092] Constrained Content Generation: The super-suggestion keyword is input into an AI-powered large language model fine-tuned in the field of health management. This model is strictly constrained; all content it generates must be based solely on the retrieval knowledge provided in the suggestion keyword, ensuring the professionalism and rigor of the solution.

[0093] Security and Logical Validation: The generated preliminary solution text is not directly presented to the user but first undergoes a separate automatic validation module. This module embeds a series of logical rules derived from a knowledge graph, which parse and cross-check the generated preliminary solution. For example, Rule 1: If the solution includes recommended food X, and the user's health profile records an allergy history to X or a similar food group, the suggestion is marked as unacceptable. Rule 2: If the user is assessed as having severe osteoporosis, and the intervention plan includes high-impact exercise, the exercise suggestion is marked as unacceptable. For any suggestion that fails validation, the system will feed it back to the aforementioned steps for regeneration or correction until all suggestions pass validation before the solution is finally released to the user.

[0094] Through the above process, this method can automatically generate multi-dimensional, personalized, safe, and reliable AI-driven health intervention plans, including exercise intervention plans, dietary intervention plans, psychological intervention plans, sleep intervention plans, behavioral correction intervention plans, and follow-up plans. The follow-up plan reminds users to attend regular follow-ups, collect health data again, and, based on the user's latest health data, reassess health risks and update the health intervention plan for the next cycle, thereby achieving proactive health management.

[0095] 4. Service Application Layer:

[0096] This layer is the final outlet for the system's value delivery, responsible for transforming the insights and decisions of the artificial intelligence engine into health management services that users can understand, perceive, and benefit from.

[0097] Immersive user interaction: Provide data visualization reports, risk interpretation, and gamified intervention tasks through mobile applications or web portals.

[0098] 24 / 7 Intelligent Companionship: Integrating an AI-powered chatbot, based on a large AI language model, this system provides users with real-time online Q&A and assists with health services.

[0099] Precision health education: Based on the user's health profile and current management stage, the most relevant health knowledge and skills guidance videos are precisely pushed through the popular science guidance module.

[0100] Adaptive Management: Activate the follow-up module. The system will feed back user follow-up data and implementation status to the core layer of the artificial intelligence engine to dynamically adjust the intervention plan for the next cycle, forming a continuously learning and evolving adaptive health management system.

[0101] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes will be obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for risk assessment and personalized treatment plan generation based on multi-source health data, characterized in that, The method includes: Collect user health data, which includes user basic information, health assessment data, health checkup data, and health questionnaire data; among which, The health assessment data includes human body composition and morphological indicators, cardiovascular function and hemodynamic indicators, lung function indicators, and bone health indicators collected through non-invasive testing equipment. The health checkup data includes biochemical indicators, routine examination data, and imaging and functional examination results; The health questionnaire data includes users' lifestyle, medical history, and symptom information; The collected health data is cleaned, transformed, and standardized, and user health profiles are built based on the processed health data. Derivative features, including trend features and cross features, are also constructed, and the processed data and health profiles are securely stored. The processed health data is input into a pre-trained cluster of artificial intelligence models, which includes stacked ensemble learning models and interpretable random forest models. The stacked ensemble learning model is used to assess the probability of a user's major disease risk based on the health data; The interpretable random forest model is used to classify and assess a user's specific health risks based on the health data, and outputs a ranking of the importance of the top input features that contribute the most to the prediction. Based on the risk assessment results, users' proactive health management goals are categorized into different risk level categories, including: Category 1 Management Objective: Major disease risk assessment results indicate high risk; The second category of management objectives: those whose major disease risk assessment results are of medium risk or who have sub-health issues; The third category of management objectives: The risk assessment result for major diseases is low risk and there are no sub-health issues; Based on users' health data and risk level categories, personalized health intervention plans are generated through a retrieval-enhanced generation architecture. The generation steps include: The user's health profile is converted into a query vector, and an approximate nearest neighbor search is performed in a pre-built vector knowledge base to recall relevant knowledge fragments; Based on the key entities in the health profile, graph queries are performed in a pre-constructed structured knowledge graph to retrieve relevant medical rules and constraints. The retrieved knowledge fragments, medical rules and constraints, the user's complete health profile, preset output format requirements, and role definition instructions are combined to form structured prompts; The structured prompts are input into a large language model that has been fine-tuned in the field of health management to generate a preliminary health intervention plan text; The generated preliminary solutions are subjected to security and logical verification, including parsing and cross-checking the solution text using logical rules derived from the knowledge graph. If a suggestion in the solution conflicts with the logical rules, the suggestion is marked as unacceptable, and a regeneration or correction mechanism is triggered for the unacceptable suggestion until all suggestions pass the verification. The risk assessment results and health intervention plans are updated based on the user's follow-up data to achieve a dynamic closed loop of health management.

2. The method according to claim 1, characterized in that, The construction and operation of the stacked ensemble learning model include: Gradient boosting decision trees, logistic regression with L1 and L2 regularization, and random forests are used as base learners. Each base learner independently processes the input health data and outputs a preliminary predicted probability. The LightGBM model is used as the meta-learner. The initial predicted probabilities output by all base learners are used as input features for refitting, and the final risk probability is output.

3. The method according to claim 1, characterized in that, The interpretable random forest model quantifies the contribution of each input feature to the model's prediction results by calculating the average reduction in Gini impurity or the importance of permutation, and outputs the importance ranking of the top few features with the highest contribution.

4. The method according to claim 1, characterized in that, The method also includes a sub-health problem determination step: based on the health data, combined with a pre-constructed knowledge graph and clinical standards, the sub-health status of the user's musculoskeletal system, cardiopulmonary system and metabolic system is determined.

5. The method according to claim 1, characterized in that, The security and logic verification of the generated scheme includes: If a user is assessed as having severe osteoporosis and the preliminary health intervention plan text includes high-impact exercise, then the recommendation for high-impact exercise is deemed unacceptable. If the plan includes recommended food X, and the user's health profile records a history of allergy to food X or food in the same category, then the dietary recommendation item is deemed unacceptable.

6. The method according to claim 1, characterized in that, The health intervention program includes at least one of the following: exercise intervention program, dietary intervention program, psychological intervention program, sleep intervention program, behavior correction intervention program, and follow-up program.

7. The method according to claim 1, characterized in that, The step of updating the risk assessment results and health intervention plan based on the user's follow-up data specifically includes: re-executing the risk assessment and plan generation steps based on the updated health data to form an adaptive health management closed loop.

8. The method according to claim 1, characterized in that, The major disease risks mentioned include the risk of stroke and the risk of cardiovascular disease.

9. The method according to claim 1, characterized in that, The specific health risks include COPD risk, fracture risk, and depression and anxiety risk.

10. The method according to claim 1, characterized in that, The generation of the personalized health intervention plan further incorporates the user's personal preferences and execution capabilities, and adapts the intensity, frequency, and form of the plan accordingly.

Images