Hybrid expert intelligent health analysis system based on large model
By combining traditional machine learning models with large language models, personalized health analysis solutions are generated, solving the problems of individual difference characterization and multi-source data processing in health and wellness platforms, and achieving efficient and personalized health management and prevention.
Patent Information
- Application Number
- CN202511561414.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-29
- Publication Date
- 2026-01-30
AI Technical Summary
Existing health and wellness information platforms are unable to fully depict individual differences and lack personalized disease prevention and intervention suggestions. Traditional machine learning models lack in-depth knowledge reasoning capabilities, while large language models lack targeted processing of multi-source data in health and wellness scenarios.
By combining traditional machine learning models (such as XGBoost and Random Forest) with large language models, personalized health analysis and intervention plans are generated through data preprocessing, machine learning prediction, and large language model analysis, while LoRA fine-tuning technology is used to reduce computing resource consumption.
It improves the accuracy of disease risk prediction and the intelligence level of personalized health analysis, and provides scientific and interpretable health guidance solutions, which are suitable for health management of the elderly and smart elderly care platforms.
Smart Images
Figure CN121439166A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of machine learning, in particular to a hybrid expert intelligent health analysis system based on a large model. BACKGROUND
[0002] With the acceleration of the aging process of society, the health management of the elderly and the prevention and control of chronic diseases have gradually become the core problems in the field of public health and health care. The existing health care information platform usually relies on rule base or simple statistical methods for risk assessment and health management, which is difficult to fully depict individual differences and provide personalized disease prevention and intervention suggestions. In recent years, machine learning methods have been widely used in medical diagnosis, disease prediction and other tasks. For example, models such as logistic regression, random forest and XGBoost can use structured data for classification and prediction, which to some extent improves the accuracy of disease risk assessment. However, such models can only output limited prediction results and lack deep knowledge reasoning and explanation capabilities, which are insufficient to meet the needs of health care in the whole life cycle health analysis and personalized care programs. At the same time, the development of large-scale pre-training language models has brought new opportunities for intelligent health management. Through fine-tuning on medical knowledge bases or disease guidelines, large language models can have strong medical knowledge integration and reasoning capabilities. However, existing large model applications are mostly focused on general question answering and medical literature analysis, lacking deep integration with traditional machine learning methods and targeted processing of multi-source data in the health care scene. Therefore, how to combine the efficient classification and prediction capabilities of traditional machine learning models with the knowledge reasoning capabilities of large language models to form a complete technical process that can not only predict disease risk probabilities but also generate personalized intelligent health analysis programs has become a problem that needs to be solved in the health care field. SUMMARY
[0003] Therefore, the present application provides a hybrid expert intelligent health analysis system based on a large model, which can balance the efficiency of traditional machine learning models and the knowledge reasoning advantages of large language models, improve the accuracy of disease risk prediction, and enhance the intelligence and personalization level of health care health analysis.
[0004] To achieve the above-mentioned purpose, the technical scheme of the present application is as follows: A hybrid expert intelligent health analysis system based on a large model, comprising: a data preprocessing module for cleaning and standardizing the collected multi-source health data, wherein the multi-source health data includes demographic information, physiological indicators, behavior data and environmental data; a machine learning prediction module for receiving preprocessed structured data and using one or more machine learning classification models to predict disease risk, outputting probability results of individuals in different disease risk dimensions; a large language model analysis module configured to receive the risk probability result output by the machine learning prediction module and perform knowledge reasoning and integration based on a large language model fine-tuned by a professional disease knowledge base to generate an individualized intelligent health analysis and intervention scheme; wherein the output of the machine learning prediction module serves as the input of the large language model analysis module, and together they constitute a hybrid expert intelligent analysis process.
[0005] wherein the machine learning prediction module comprises a hypertension detection model and an Alzheimer's detection model; The hypertension detection model is an XGBoost-based model, and its objective function includes a loss function and a regularization term, which is used to control the model complexity and prevent overfitting. The Alzheimer's detection model is a random forest-based model that outputs the final prediction result and probability by integrating multiple decision trees and using majority voting.
[0006] wherein the XGBoost hypertension detection model is trained by iteratively constructing decision trees, and when constructing the decision trees, the best split point is selected by evaluating the split gain.
[0007] wherein the random forest Alzheimer's detection model performs the following steps when training a single decision tree: performing self-sampling with replacement from the training data set to form a sub-data set; When splitting a node, randomly select a feature subset from all features and use Gini index or information gain as the splitting criterion; Repeat the split until the preset maximum depth of the tree or the minimum sample size threshold of the node is met.
[0008] wherein the large language model in the large language model analysis module is fine-tuned using the LoRA method; The LoRA method simulates parameter updates by connecting low-rank matrices `A` and `B` in parallel to the linear transformation layer of the large language model, where the product of `BA` is a low-rank matrix, and the rank `r` is much smaller than the dimension of the original parameter matrix. During training, the original parameters of the large language model are fixed, and only the low-rank matrices `A` and `B` are trained.
[0009] wherein the low-rank matrix `A` is initialized using a Gaussian distribution, the low-rank matrix `B` is initialized using a zero matrix, and the rank `r` is 8.
[0010] wherein the data preprocessing module performs standardization or normalization processing on numerical features.
[0011] The system is applied to the scenes of health management of the elderly, prevention of chronic diseases or smart elderly care platform, and provides disease risk prediction and personalized health guidance scheme for users.
[0012] Beneficial effects: 1. The system of the present application is a large model vertical application in the field of health and care, which realizes intelligent health analysis by combining traditional machine learning and intelligent health management of large language model. Specifically, first, the collected multi-source health data (including demographic information, physiological indicators, behavior data and environmental data) are preprocessed and cleaned. Then, for different application scenarios, a variety of machine learning models (such as logistic regression, random forest, XGBoost, etc.) are used to analyze and classify the input structured data, and output the probability results of individuals in different disease risk dimensions. On this basis, the risk prediction results are input into the large language model based on professional disease knowledge base for LoRA fine-tuning, and through its powerful reasoning and knowledge integration ability, personalized intelligent health analysis and intervention scheme are generated.
[0013] 2. The present application can take into account the efficiency of traditional machine learning model and the knowledge reasoning advantage of large language model, which not only improves the accuracy of disease risk prediction, but also enhances the intelligence and personalization level of health and care analysis. The system can be widely used in the scenes of health management of the elderly, prevention of chronic diseases, smart elderly care platform, etc., and provides scientific and explainable health guidance scheme for users.
[0014] 3. In the system of the present application, the precise prediction ability of traditional machine learning model and the deep knowledge reasoning advantage of large language model are effectively integrated, and through the series working process, the structured health data is first predicted by XGBoost, random forest and other efficient algorithms in multiple dimensions of disease risk probability, and then the prediction results are input into the large language model fine-tuned by professional medical knowledge, to generate health analysis and intervention scheme which is scientific and personalized, significantly improving the intelligence level of health and care management.
[0015] 4. The system of the present application introduces advanced model training and optimization strategies to ensure the accuracy and reliability of health risk assessment. Specifically, the XGBoost model prevents overfitting through the objective function containing regularization term and split gain calculation, and the random forest model enhances the generalization ability of the model through feature random selection and majority voting mechanism. This design enables the system to maintain stable and high-precision prediction performance when facing complex multi-source health data of the elderly.
[0016] 5.The system of the present application adopts LoRA low-rank adaptive technology to efficiently fine-tune large language models, and only a small number of parameters need to be trained to make them proficient in the professional knowledge of health and chronic disease management. This technical path not only greatly reduces the consumption of computing resources, making the intelligent system easy to deploy on regular servers, but also ensures that the system can generate interpretable and executable health guidance recommendations in real time based on the latest medical guidelines and individual risk profiles, thereby having high practical value in smart aging and other scenarios. BRIEF DESCRIPTION OF DRAWINGS
[0017] Figure 1 The system working principle diagram of the embodiment of the present application. DETAILED DESCRIPTION
[0018] The present application will be described in detail below with reference to the accompanying drawings and examples.
[0019] The present application provides a hybrid expert intelligent health analysis system based on a large model, which preprocesses and cleans the collected multi-source health data (including demographic information, physiological indicators, behavior data and environmental data); for different application scenarios, a plurality of machine learning models (such as logistic regression, random forest, XGBoost, etc.) are used to analyze and classify the input structured data, and output the probability results of individuals in different disease risk dimensions. On this basis, the risk prediction results are input into a large language model that has been fine-tuned based on a professional disease knowledge base, and through its powerful reasoning and knowledge integration capabilities, personalized intelligent health analysis and intervention schemes are generated.
[0020] The system of the present application comprises: A data preprocessing module for cleaning and standardizing the collected multi-source health data, including demographic information, physiological indicators, behavior data and environmental data; A machine learning prediction module for receiving preprocessed structured data and using one or more machine learning classification models to predict disease risk, outputting probability results of individuals in different disease risk dimensions; A large language model analysis module for receiving risk probability results output by the machine learning prediction module and performing knowledge reasoning and integration based on a large language model fine-tuned by a professional disease knowledge base to generate personalized intelligent health analysis and intervention schemes; Wherein, the output of the machine learning prediction module serves as the input of the large language model analysis module, and together they constitute a hybrid expert intelligent analysis process.
[0021] The principle of the embodiment of the present application is as follows Figure 1As shown, in the system of the present embodiment, the input structured data is subjected to feature analysis and classification by the XGBoost-based hypertension detection model and the random forest-based Alzheimer's detection model, and the probability results of the individual in different disease risk dimensions are output.
[0022] Specifically, for a data set containing n pieces of m-dimensional data, the XGBoost-based hypertension detection model can be represented as:
[0023] wherein: is a CART decision tree, k is the number of trees, is an input feature vector.
[0024] The objective function of the XGBoost-based hypertension detection model includes a loss function and a regularization term:
[0025]
[0026]
[0027] wherein: is an L1 regularization term is an L2 regularization term T is the number of leaf nodes w j is the weight of the leaf node.
[0028] The training process of the XGBoost-based hypertension detection model is an iterative process. First, the model is initialized. For each iteration, the first-order derivative and the second-order derivative of the current model are calculated, a new decision tree is constructed, the objective function is optimized, the optimal weight of each leaf node is calculated, and finally the new tree is added to the model. Repeat the steps until the specified number of iterations is reached or the early stopping condition is met.
[0029] When constructing the decision tree of the XGBoost-based hypertension detection model, the best split point is selected by evaluating the split gain. The calculation formula of the split gain is:
[0030] wherein: GL and GR are the first-order derivatives and of the left and right child nodes, respectively, HL and HR are the second-order derivatives and of the left and right child nodes, respectively, and γ is the regularization parameter.
[0031] Further, by reducing the learning rate (eta) and increasing the number of trees (n_estimators), increasing min_child_weight and gamma, using subsample and colsample_bytree for sampling, increasing the regularization parameters lambda and alpha, etc. to avoid the phenomenon of overfitting; increase max_depth to capture more complex patterns, reduce min_child_weight to allow more fine-grained splitting, adjust the learning rate and the number of trees to balance accuracy and training time, and improve model performance.
[0032] In the random forest-based Alzheimer's detection model, numerical features are standardized or normalized to ensure that each feature is within the same order of magnitude range.
[0033]
[0034] where, is the original feature value, is the mean of the feature, is the standard deviation of the feature.
[0035]
[0036] where, are the minimum and maximum values of the feature, respectively.
[0037] Information gain (Information Gain), Gini index (Gini Index), or feature importance ranking is used to select key features;
[0038] where, is the D entropy of the data set, is the subset with value v on feature A.
[0039] is constructed as:
[0040] where is the probability that the sample belongs to the ith class, and C is the number of classes.
[0041] Set the number of decision trees in the forest to 200, and for each decision tree, bootstrap sampling is performed from the training data set, that is, randomly sampling with replacement to form a sub-data set:
[0042] where, the sample is drawn with replacement from the training set D to form the sub-data set .
[0043] When splitting a node, six subset features are randomly selected from approximately 30 features in total, and then the Gini index or entropy value is used as the splitting criterion.
[0044]
[0045] Repeat the splitting until the maximum depth is reached. , The current depth of the tree. This is a preset maximum depth threshold or minimum sample number threshold. |D| is the number of samples contained in the current node. The preset minimum sample size threshold.
[0046] Multiple decision trees are integrated, and majority voting is used as the final output of the model.
[0047]
[0048] in Let B represent the prediction result of the b-th tree, where B is the number of predictions and I is the indicator function.
[0049] For a new input sample x, each tree Final classification results:
[0050] Simultaneously output the predicted probability:
[0051] in, It is the first Each tree predicts the probability of "category=1".
[0052] In this embodiment, based on Alzheimer's disease and hypertension datasets, LORA fine-tuning is used to fine-tune a large model. During LORA fine-tuning of the large model, several new weight matrices are inserted alongside the LLM matrix, operating in parallel with it. Due to the low-rank nature of the model, it is decomposed into a reduced-dimensional matrix. and the increasing dimension matrix ,in This allows for training the LLM with a very small number of parameters. During training, the LLM parameters are fixed, and only matrices A and B are trained. After the model training is complete, A and B can be directly added to the original parameters, thus avoiding additional inference latency during inference.
[0053]
[0054] At initialization, A is initialized with Gaussian and B is initialized with a zero matrix. Because r is usually a very small value, it does not bring significant increase of video memory. It is tested that the embodiment takes r=8 as the size of rank.
[0055] To sum up, the above is only the preferred embodiment of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A hybrid expert intelligent health analysis system based on a large model, characterized in that, The system comprises: a data preprocessing module for cleaning and standardizing the collected multi-source health data, including demographic information, physiological indicators, behavior data and environmental data; a machine learning prediction module for receiving the preprocessed structured data and using one or more machine learning classification models to predict disease risk, outputting the probability of individuals in different disease risk dimensions; a large language model analysis module for receiving the risk probability results output by the machine learning prediction module and performing knowledge reasoning and integration based on a large language model fine-tuned by a professional disease knowledge base to generate personalized intelligent health analysis and intervention programs; wherein the output of the machine learning prediction module serves as the input of the large language model analysis module, and together they constitute a hybrid expert intelligent analysis process.
2. The system of claim 1, wherein, The machine learning prediction module includes a hypertension detection model and an Alzheimer's detection model. The hypertension detection model is an XGBoost-based model whose objective function includes a loss function and a regularization term to control model complexity and prevent overfitting. The Alzheimer's detection model is a random forest-based model that integrates multiple decision trees and outputs the final prediction result and probability through majority voting.
3. The system of claim 2, wherein, The XGBoost hypertension detection model is trained by iteratively constructing decision trees, and when constructing the decision trees, the best split point is selected by evaluating the split gain.
4. The system of claim 2, wherein, The random forest Alzheimer's detection model performs the following steps when training a single decision tree: Perform resampling with replacement from the training data set to form a sub-data set; When splitting the node, randomly select a feature subset from all features and use Gini index or information gain as the splitting criterion; Repeat the split until the preset maximum depth of the tree or the minimum sample size threshold of the node is met.
5. The system of claim 1, wherein, The large language model in the large language model analysis module is fine-tuned using the LoRA method; The LoRA method simulates parameter updates by connecting low-rank matrices A and B in parallel to the linear transformation layer of the large language model, where the product of BA is a low-rank matrix, and the rank r is much smaller than the dimension of the original parameter matrix; During training, the original parameters of the large language model are fixed, and only the low-rank matrices A and B are trained.
6. The system of claim 5, wherein, The low-rank matrix A is initialized using a Gaussian distribution, the low-rank matrix B is initialized using a zero matrix, and the value of the rank r is 8.
7. The system of claim 1, wherein, The data preprocessing module performs standardization or normalization processing on numerical features.
8. The system of any one of claims 1 to 7, wherein, The system is applied to the scenarios of elderly health management, chronic disease prevention or smart elderly care platforms, and provides users with disease risk prediction and personalized health guidance programs.