Method for predicting early diabetes mellitus based on health data

Abstract

The invention relates to the technical field of diabetes prediction, in particular to a method for predicting early diabetes based on health data, and the method comprises the steps: carrying out the data collection and preprocessing of an individual health system, and obtaining diabetes risk related data; performing off-line calculation on the data by using a feature correlation analysis technology, and extracting a diabetes prediction feature set; performing risk assessment on the feature set through a preset diabetes prediction model to generate a diabetes risk prediction result; key risk factors are extracted, a risk factor classification model based on a random forest is constructed, risk attribution analysis is carried out, and main risk factors are identified; then, comprehensive assessment is carried out on the main risk factors in combination with confidence propagation analysis and risk path analysis technologies, and a core risk source is positioned; the individual diabetes risk portrait is constructed according to the core risk source, the relationship among the risk factors is analyzed, a personalized early intervention scheme is formulated, and the comprehensiveness, precision and interpretability of early diabetes prediction are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of diabetes prediction, and in particular to a method for predicting early diabetes based on health data. BACKGROUND

[0002] Diabetes is a major global public health problem. Early detection of prediabetes and implementation of intervention is the key to reducing the incidence of diabetes and the risk of complications. Traditional diabetes screening relies on a single physiological indicator (such as fasting blood glucose, postprandial blood glucose), which is difficult to fully capture the potential risks of individuals in multiple dimensions such as genetics, lifestyle, and metabolic indicators, and has problems such as high rate of missed diagnosis and insufficient prediction accuracy. With the popularity of wearable devices, electronic health records (EHR) and health monitoring technology, individual health data is showing a multi-modal trend (covering blood glucose, blood pressure, BMI, eating habits, exercise data, family history, etc.). How to efficiently integrate and mine the risk association features in these data has become a core challenge for early diabetes prediction.

[0003] In the prior art, the prediction method based on a single model or simple statistical analysis is difficult to handle the non-linear correlation and dynamic risk transmission mechanism between complex health data, resulting in weak attribution ability to key risk factors and inability to provide accurate guidance for personalized intervention. For example, the traditional logistic regression model is insufficient to describe the interaction of high-dimensional features, and a single machine learning model may ignore the causal relationship between features, resulting in poor interpretability of the prediction results. In addition, the existing method lacks deep analysis of the complex dependence relationship between risk factors, making it difficult to identify the core risk source driving the occurrence of diabetes, resulting in limited effect of the intervention scheme.

[0004] The above content is only used to assist in understanding the technical solutions of the present application and does not represent the acknowledgement of the above content as prior art. SUMMARY

[0005] The main purpose of the present application is to provide a method for predicting early diabetes based on health data, which aims to solve the technical problems of traditional early diabetes prediction methods, such as difficulty in handling complex associations of multi-dimensional health data, accurately attributing risk factors, and developing personalized intervention schemes, resulting in insufficient prediction accuracy, poor interpretability, and limited intervention effect.

[0006] To achieve the above purpose, the present application provides a method for predicting early diabetes based on health data, which comprises:

[0007] Data acquisition and data preprocessing of individual health system to obtain diabetes risk related data;

[0008] Using feature correlation analysis technology to perform offline calculation on the diabetes risk related data to obtain a set of diabetes prediction features;

[0009] perform risk assessment on the diabetes prediction feature set through a preset diabetes prediction model to obtain a diabetes risk prediction result;

[0010] perform key risk factor extraction on the diabetes risk prediction result to obtain a risk factor feature; construct a risk factor classification model based on a random forest, input the risk factor feature into the risk factor classification model to perform risk attribution analysis, and obtain a main risk factor;

[0011] perform comprehensive evaluation on the main risk factor by using a method combining confidence propagation analysis and risk path analysis technology to obtain a core risk source;

[0012] construct an individual diabetes risk portrait according to the core risk source, further understand the relationship between risk factors through the individual diabetes risk portrait, and develop a personalized early intervention scheme.

[0013] Optionally, the data collection and data preprocessing on the individual health system to obtain diabetes risk related data include:

[0014] perform multi-source data collection on the individual health system to obtain individual health data;

[0015] perform data cleaning, standardization and feature engineering processing on the individual health data to obtain diabetes risk related data.

[0016] Optionally, the offline calculation on the diabetes risk related data by using a feature correlation analysis technology to obtain a diabetes prediction feature set includes:

[0017] perform feature conversion and discretization on the diabetes risk related data to obtain health risk feature data;

[0018] obtain a risk feature frequent item set by using a feature correlation mining based on an improved Apriori algorithm to obtain strong correlation rules in the health risk feature data;

[0019] calculate the correlation strength of the risk feature frequent item set according to a feature importance evaluation algorithm, obtain a correlation rule with high correlation strength, and obtain a feature correlation degree;

[0020] perform sorting on the risk feature frequent item set corresponding to the feature correlation degree in descending order, and select a risk feature frequent item set with a higher feature correlation degree to obtain a diabetes prediction feature set.

[0021]

[0022] Optionally, the feature correlation degree algorithm formula is as follows:

[0023] ​

[0024] wherein R Fi, Fj is the correlation degree between features Fi and Fj, exp is an exponential function, and α and β are weight coefficients, and P Fi and P Fjare probabilities of features Fi and Fj, respectively, P(Fi ∩ Fj) is a probability of simultaneous occurrence of features Fi and Fj, and MI Fi, Fj is mutual information between features Fi and Fj, and γ is a correlation degree correction coefficient.

[0025] Optionally, the risk assessment on the diabetes prediction feature set by the preset diabetes prediction model to obtain a diabetes risk prediction result comprises:

[0026] data normalization and feature coding are performed on the diabetes prediction feature set to obtain model input data; the model input data is divided into a training set, a validation set and a test set;

[0027] a diabetes prediction model based on deep learning is constructed, wherein the diabetes prediction model comprises a feature extraction layer, a risk assessment layer and a result output layer;

[0028] the training set is input into the diabetes prediction model based on deep learning for model training, and the validation set is input into the model for parameter tuning through a diabetes prediction loss function as follows to generate an optimal prediction model;

[0029] wherein the diabetes prediction loss function is as follows:

[0030]

[0031] wherein, is the diabetes prediction loss function, is a model parameter, N is a sample number, is a true label of the i-th sample, wherein 1 represents a disease and 0 represents health, is a prediction probability of the i-th sample, λ is a regularization coefficient, M is a model layer number, θk is a model parameter of the k-th layer, and δ is a loss function correction coefficient;

[0032] the test set is input into the optimal prediction model for performance evaluation to obtain a diabetes risk prediction result, including a disease probability and a risk level.

[0033] Optionally, the key risk factors are extracted from the diabetes risk prediction result to obtain risk factor features; a risk factor classification model based on a random forest is constructed, and the risk factor features are input into the risk factor classification model for risk attribution analysis to obtain main risk factors, including:

[0034] Risk factors are extracted from high-risk samples in the diabetes risk prediction result to obtain risk factor features;

[0035] A risk factor classification model based on a random forest is constructed, and the risk factor features are input into the risk factor classification model for training to obtain risk factor importance ranking; the model is explained by SHAP value to generate risk factor contribution degree;

[0036] The risk factor contribution degree is quantitatively analyzed by a risk factor weight calculation function as follows to obtain a risk factor weight value;

[0037] The formula of the risk factor weight calculation function is as follows:

[0038]

[0039] In the formula, W f is the weight value of the risk factor f, T is the number of decision trees, Gini f, t is the Gini index reduction amount of feature f in the tth tree, Ginitotal t is the total Gini index reduction amount of the tth tree, ΔAccuracy f, t is the accuracy drop value of the tth tree after removing the feature f, ΔAccuracymax t is the maximum accuracy drop value of the tth tree, and ϵ is a weight correction coefficient.

[0040] The risk factor weight values are sorted in descending order to obtain main risk factors, including blood glucose level, BMI, family history, and eating habits.

[0041] Optionally, the high-risk samples in the diabetes risk prediction result are subjected to risk factor extraction to obtain risk factor features, including:

[0042] The high-risk individual data set is screened out by threshold division on the diabetes risk prediction result;

[0043] Feature correlation analysis was performed on the high-risk individual dataset to extract factors significantly associated with the onset of diabetes, resulting in a risk factor feature dataset; a risk factor feature database was constructed, and the risk factor feature dataset was saved to the risk factor feature database;

[0044] Obtain feature data from the risk factor feature database and combine it with the clinical knowledge base for specific analysis.

[0045] The risk factor characteristics were obtained through screening.

[0046] Optionally, the method of combining confidence propagation analysis and risk path analysis to comprehensively evaluate the main risk factors and obtain the core risk sources includes:

[0047] The main risk factors are standardized to obtain standard risk factor data;

[0048] Risk path analysis techniques are used to analyze the causal relationships between standard risk factors, construct a causal network diagram of risk factors, and obtain an initial risk path map.

[0049] By using a pre-defined confidence propagation analysis technique, the risk factors in the initial risk path map are probabilistically inferred to obtain the core risk sources.

[0050] Optionally, the step of using a preset confidence propagation analysis technique to perform probabilistic reasoning on the risk factors in the initial risk path map to obtain the core risk sources includes:

[0051] A confidence propagation analysis technique is constructed, which includes Bayesian network modeling, probabilistic inference, and maximum a posteriori probability estimation.

[0052] The initial risk path graph is probabilistically modeled using Bayesian network modeling to obtain the prior probabilities of nodes.

[0053] Confidence propagation is performed based on the prior probabilities of nodes. According to the conditional dependencies between risk factors, the posterior probability of each node is updated through a message passing algorithm to obtain the joint probability distribution of risk factors.

[0054] Maximum a posteriori probability estimation is performed on the joint probability distribution of risk factors to identify the combination of risk factors that contributes the most to the incidence of diabetes, thus obtaining the core risk source.

[0055] Optionally, the step of constructing an individual diabetes risk profile based on core risk sources, and further understanding the relationships between various risk factors through the individual diabetes risk profile, in order to develop personalized early intervention plans, includes:

[0056] A multi-dimensional analysis of core risk sources was conducted to construct an individual diabetes risk profile that includes physiological indicators, lifestyle, and genetic factors.

[0057] Personalized early intervention plans are developed based on individual diabetes risk profiles, including dietary adjustment recommendations, exercise plans, and blood glucose monitoring frequency.

[0058] By regularly tracking individual health data and intervention effects, and dynamically updating risk profiles and intervention plans, precise prevention and management of early-stage diabetes can be achieved.

[0059] This invention provides a method for predicting early diabetes based on health data. This method overcomes the limitations of traditional single-indicator detection by collecting and preprocessing individual multimodal health data.

[0060] Limitations: This approach comprehensively captures potential risk signals for diabetes, reducing the probability of missed diagnoses. It utilizes feature association analysis to mine strongly correlated features in high-dimensional data, forming a targeted diabetes prediction feature set, effectively filtering redundant information, improving model input quality, and laying the foundation for accurate predictions. A risk attribution model based on random forests is constructed, handling both the nonlinear and complex correlations in health data and identifying key risk factors through feature importance analysis, addressing the poor interpretability of traditional models and making prediction results more clinically valuable. Combining confidence propagation and risk path analysis techniques, it uncovers causal dependencies between risk factors, accurately identifying core risk sources driving diabetes, avoiding superficial correlation analysis, and providing targeted interventions. Based on core risk sources, it constructs individualized diabetes risk profiles, clearly presenting multi-factor interactions, supporting the development of differentiated early intervention plans, solving the homogenization problem of traditional intervention plans, and improving the targeting and effectiveness of preventative measures. Attached Figure Description

[0061] Figure 1 is a flowchart illustrating an embodiment of the method for predicting early diabetes based on health data according to the present invention.

[0062] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0063] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0064] Referring to Figure 1, which is a flowchart of an embodiment of the method for predicting early diabetes based on health data according to the present invention, an embodiment of the method for predicting early diabetes based on health data according to the present invention is presented.

[0065] In one embodiment, the method for predicting early diabetes based on health data includes the following steps:

[0066] Step S100: Data collection and preprocessing are performed on the individual health system to obtain diabetes risk-related data. The individual health system can be an integrated system for continuously acquiring individual physiological, behavioral, and environmental health information, and may include wearable health monitoring devices, electronic health record systems, mobile health applications, etc. This is achieved through multi-source health terminals.

[0067] The system collects raw data and performs preprocessing steps such as missing value imputation, outlier detection, unit standardization, and time alignment to ensure the integrity, consistency, and comparability of the input data, providing a reliable foundation for subsequent feature mining.

[0068] Step S200: Utilize feature association analysis technology to perform offline calculations on diabetes risk-related data to obtain a diabetes prediction feature set. The feature association analysis technology can employ methods such as mutual information, Spearman correlation coefficient, tree-based feature importance ranking, or chi-square test to perform nonlinear association mining on high-dimensional raw data. This retains combined features with stable predictive power for diabetes risk, eliminates redundant or low-correlation variables, reduces feature dimensionality, eliminates redundancy and noise, and improves the information density and generalization ability of the model input.

[0069] Step S300: A risk assessment of the diabetes prediction feature set is performed using a pre-defined diabetes prediction model to obtain diabetes risk prediction results. The pre-defined diabetes prediction model can be a trained machine learning model, such as a logistic regression model, gradient boosting tree model, or neural network classifier. Based on historical cohort data, it is trained using algorithms such as logistic regression, XGBoost, and support vector machines. Model parameters are determined through cross-validation and AUC optimization. The output is a continuous risk score or a binary classification label, achieving a preliminary quantitative assessment of individual diabetes risk and generating predictive outputs that can be used for subsequent key factor extraction.

[0070] Step S400: Extract key risk factors from the diabetes risk prediction results to obtain risk factor features. Construct a risk factor classification model based on random forest, and input the risk factor features into the risk factor classification model for risk attribution analysis to obtain the main risk factors. Key risk factor extraction can be based on model interpretation techniques such as SHAP values, LIME local approximation, or feature perturbation analysis. Attribution decomposition is performed on the output of the preset prediction model to extract several original feature variables that contribute the most to the prediction results. The prediction results of the black-box model are transformed into an interpretable list of original variable contributions, establishing a traceable mapping between predictions and factors.

[0071] Step S500: A comprehensive assessment of key risk factors is conducted using a combination of confidence propagation analysis and risk path analysis techniques to identify core risk sources. Confidence propagation analysis can be a probabilistic graphical model-based inference method that propagates uncertainty and correlation strength through conditional probability dependencies between nodes, assesses potential causal influence paths between variables, quantifies the transmission process of uncertainty among risk factors, and identifies which factors play a key role in probability propagation.

[0072] Significantly influential weights help eliminate spurious correlations. Risk path analysis techniques can be graph-based analysis methods based on causal inference to identify direct and indirect causal links between risk factors, reveal potential pathways in disease development, overcome the limitations of correlation analysis, and identify the key causal pathways that truly drive the occurrence of diabetes.

[0073] Step S600: Construct an individual diabetes risk profile based on the core risk sources. This profile helps to further understand the relationships between various risk factors, enabling the development of personalized early intervention plans. The individual diabetes risk profile can be presented as a multi-factor risk interaction network for the individual, either in the form of a visual graph or structured data. It includes the core risk sources and their interactions, visually demonstrating the causal structure of individual diabetes risk, supporting clinicians in quickly understanding the risk composition, and assisting in developing differentiated intervention strategies.

[0074] Taking early diabetes screening at a community health management center as an example, residents upload their exercise, sleep, and diet records for 30 consecutive days via smart bracelets and an app, simultaneously accessing the electronic medical record system to obtain their past physical examination data. The system automatically preprocesses the data, generating diabetes risk-related data; through feature association analysis, it identifies eight strongly correlated features, including BMI, fasting insulin, average nighttime heart rate, and post-dinner walking duration; it uses an XGBoost model to predict risk scores; it extracts the top 5 risk factors through SHAP analysis; and it uses a random forest model to rank and confirm insulin resistance index and insufficient nighttime activity as the most important factors; it constructs a Bayesian network to discover insulin resistance and increased liver glucose output as key pathways, and combines causal discovery algorithms to verify that insufficient nighttime activity, circadian rhythm disorder, and decreased insulin sensitivity are direct causal chains; finally, it generates an individual risk profile, showing that the resident's core risk sources are a sedentary lifestyle and insulin resistance tendency, rather than simply hyperglycemia. Based on this, the doctor did not recommend drug treatment, but instead developed a precise behavioral intervention plan of 30 minutes of brisk walking in the evening and standing at work for 1 hour after dinner. After three months, the resident's insulin sensitivity significantly improved, and blood sugar levels stabilized within the normal range.

[0075] Taking an enterprise employee health management platform as an example, the system integrates employee physical examination reports, work hour attendance data, cafeteria consumption records, and resting heart rate changes from wearable devices. Feature association analysis revealed a non-linear interaction effect between overtime hours and the frequency of high-sugar beverage consumption, which traditional models struggle to identify. Random forest attribution analysis confirmed overtime, sleep deprivation, elevated cortisol levels, and visceral fat accumulation as key pathways. Confidence propagation analysis showed that sleep deprivation had the highest propagation intensity within this pathway. The risk profile is characterized by three main factors: work stress, sleep disorders, and metabolic disorders.

[0076] This established a causal chain. Based on this, the platform provided personalized recommendations: optimizing work schedules, offering nighttime meditation audio, and distributing low-GI snack alternatives to the target employee group, instead of uniformly distributing blood sugar-lowering medications. One year later, the prediabetes conversion rate in this group decreased by 37%.

[0077] This embodiment provides a method for predicting early diabetes based on health data. Step S100 involves collecting and preprocessing data from an individual's health system to obtain diabetes risk-related data. Step S200 uses feature association analysis to perform offline calculations on the diabetes risk-related data to obtain a diabetes prediction feature set. Step S300 uses a pre-set diabetes prediction model to assess the risk of the diabetes prediction feature set to obtain a diabetes risk prediction result. Step S400 extracts key risk factors from the diabetes risk prediction result to obtain risk factor characteristics. Step S500 uses a combination of confidence propagation analysis and risk path analysis to comprehensively evaluate the main risk factors and obtain core risk sources. Step S600 constructs an individual diabetes risk profile based on the core risk sources, and further understands the relationships between various risk factors through the individual diabetes risk profile to develop personalized early intervention plans. The entire process forms a closed-loop chain from data collection, feature extraction, predictive modeling, attribution analysis, causal inference, profile generation, and intervention customization. This enables the early risk identification of diabetes to evolve from superficial correlation to causal drive, allowing intervention measures to directly target the core drivers of disease development and improve the accuracy and effectiveness of prevention strategies.

[0078] In one embodiment, data collection and preprocessing are performed on an individual health system to obtain diabetes risk-related data, including:

[0079] Individual health data is obtained by collecting data from multiple sources on an individual health system;

[0080] Individual health data is cleaned, standardized, and feature-engineered to obtain diabetes risk-related data.

[0081] In this context, an individual health system can be an integrated system for continuously acquiring individual physiological, behavioral, and environmental health information. It can provide multimodal health data input sources covering physiological indicators, lifestyle habits, and genetic background, supporting the comprehensiveness and dynamism of subsequent risk modeling. Individual health systems may include wearable health monitoring devices, electronic health record systems, and mobile health applications.

[0082] In this embodiment, multi-source data is collected from the individual health system to obtain individual health data. Individual health data can be directly obtained, unprocessed raw health information.

[0083] Data sets can serve as the initial input for data preprocessing, containing the most basic health information, but they also contain issues such as missing data, noise, and inconsistent formats. Multi-source data acquisition can achieve comprehensive coverage of multimodal health data, avoiding the dimensional gaps and information biases caused by a single data source.

[0084] Furthermore, data cleaning is performed on individual health data to obtain diabetes risk-related data. Data cleaning can identify and correct outliers, missing values, and timestamp misalignments in individual health data, thereby improving data integrity and accuracy and reducing model misjudgments and biases caused by data quality defects.

[0085] In this embodiment, individual health data is standardized to obtain diabetes risk-related data. Standardization can eliminate the interference of dimensional differences on distance-sensitive models (such as KNN and SVM), and improve the convergence stability and fairness of model training.

[0086] Furthermore, feature engineering is performed on individual health data to obtain diabetes risk-related data. Feature engineering can be a process based on domain knowledge and statistical methods to construct clinically interpretable derived variables or combined features from raw health data. It can explicitly encode nonlinear interactions and physiological mechanisms, transforming raw data into clinically interpretable risk representations, significantly improving the predictive efficacy and interpretability of subsequent models.

[0087] Taking a regional early diabetes screening program as an example, the method for predicting early diabetes based on health data in this embodiment can be as follows: The system synchronously collects raw data from resident-authorized smart bracelets, community health check-up centers (EHRs), home blood glucose meters, health management apps, and genetic testing reports to form individual health data. In the data cleaning stage, abnormal blood glucose values ​​(>20 mmol / L) triggered by showering are removed from the bracelets, and missing family medical history records are imputed using KNN. In the standardization stage, BMI, fasting blood glucose, and triglycerides are standardized to a Z-score. In the feature engineering stage, 12 high-order features are constructed, including the insulin resistance index, sedentary-obesity interaction score, and sleep fragmentation index (number of nighttime awakenings × total sleep duration). The final diabetes risk-related data contains a mixed set of raw variables and derived features, serving as input for subsequent feature association analysis. Compared to the old process that only uses raw physical examination data, this embodiment improves the model's AUC for prediabetes by 0.15 and reduces the false negative rate by 22%.

[0088] This embodiment provides a method for predicting early diabetes based on health data. It acquires raw individual health data covering physiological, behavioral, and genetic aspects through multi-source data collection, cleans the data to remove noise and missing data, standardizes and unifies the units of measurement, and then constructs a feature-based system.

[0089] Derivative variables explaining clinical mechanisms generate high-quality diabetes risk-related data. This process transforms low-order, heterogeneous, and noisy raw data into a structured, semantically rich input set with non-linear expressive capabilities. It fundamentally solves the problems of missed diagnoses and low accuracy caused by incomplete data, weak features, and noise interference in traditional screening, providing a solid, interpretable, and high signal-to-noise ratio data foundation for subsequent feature association analysis and risk modeling, and achieving systematic optimization from raw data collection to risk characterization construction.

[0090] In one embodiment, feature association analysis technology is used to perform offline calculations on diabetes risk-related data to obtain a diabetes prediction feature set, including:

[0091] By performing feature transformation and discretization on diabetes risk-related data, health risk feature data can be obtained.

[0092] Among these, diabetes risk-related data can be multi-dimensional feature variables that may be relevant to the potential development of diabetes, and can be used to carry multi-dimensional risk signals for individuals at the metabolic, behavioral, and genetic levels. In this embodiment, diabetes risk-related data includes physiological indicator data, behavioral habit data, and genetic and demographic data. By dividing continuous variables (such as blood glucose, BMI, and heart rate) into finite semantic intervals (such as low, medium, and high), and standardizing categorical variables (such as diet type and exercise frequency), the data is made suitable for symbolic association analysis, thus obtaining health risk feature data.

[0093] By utilizing feature association mining based on the improved Apriori algorithm, strong association rules in health risk feature data are obtained, resulting in frequent itemsets of risk features.

[0094] The feature association mining module based on the improved Apriori algorithm can be an enhanced association rule mining algorithm implementation unit used to discover high-frequency co-occurring feature combinations in discretized health risk feature data. In this embodiment, the feature association mining module based on the improved Apriori algorithm introduces a dynamic support threshold adaptive mechanism and a pruning strategy based on clinical priority to avoid rule omissions due to data sparsity and reduce the computational cost of low-value combinations. By applying the improved Apriori algorithm to iteratively generate candidate itemsets in the discretized feature space, and filtering frequent combinations that meet the dynamic support threshold, frequent itemsets of risk features are obtained.

[0095] The association strength of frequent itemsets of risk features is calculated based on the feature importance assessment algorithm, and association rules with high association strength are obtained to obtain the feature association degree.

[0096] Among them, the feature importance evaluation algorithm can be used to quantify frequent itemset pairs of risk features.

[0097] The calculation method for the contribution of diabetes risk prediction is based on information theory or model importance indicators. In this embodiment, information gain, mutual information, or Gini importance based on random forest are used to calculate the predictive power of each frequent itemset as a whole for the target variable (diabetes status). For each frequent itemset of risk features, the feature importance evaluation algorithm is input to calculate its overall predictive contribution to diabetes status, which is used as the association strength score for the combination.

[0098] The formula for the feature correlation algorithm is as follows:

[0099]

[0100] In the formula, R Fi, Fj Let exp be the correlation between features Fi and Fj, α and β be the weighting coefficients, and P be the weighting coefficient. Fi and P Fj Let Fi and Fj be the probabilities of features Fi and Fj respectively, and let P(Fi ∩ Fj) be the probability of features Fi and Fj appearing simultaneously. MI Fi, Fj Let γ be the mutual information between features Fi and Fj, and let γ be the correlation correction coefficient.

[0101] It should be noted that, This is used to quantify the statistical significance of the co-occurrence of two features. A value > 1 indicates that the co-occurrence probability of the two features is higher than that of independent events. For example, the elevation ratio of BMI > 25 and fasting blood glucose ≥ 6.1 mmol / L is 2, indicating that the probability of hyperglycemia in obese individuals is twice that of the general population, suggesting that clinical focus should be placed on the joint screening of these two features. When mining frequent itemsets using the Apriori algorithm, high-frequency co-occurring risk factor combinations are screened out, such as hypertension + hyperlipidemia, which may jointly exacerbate insulin resistance. It captures non-linear dependencies between features; even if the co-occurrence frequency is not high, if the information complementarity is strong, such as the high mutual information between sleep duration < 6 hours and insulin resistance, potential mechanistic associations can be identified. Insufficient sleep indirectly leads to insulin resistance by affecting leptin secretion. To compensate for the limitations of elevation ratios, only high-frequency co-occurrences are counted to discover low-frequency but crucial risk pairs, such as the weak co-occurrence but strong information dependence between gene variants and specific metabolic indicators.

[0102] The elevation weight α is used when clinical focus is on risk combinations prevalent in large populations, such as the high-frequency co-occurrence of obesity and dietary structure. Increasing α can make the elevation weight contribute more to the association, avoiding the omission of common risk links. In primary care screening scenarios, α > β is preferred to focus on easily detectable high-frequency risk factors. If the research scenario needs to explore deeper associations at the pathological mechanism level, such as the information dependence between fatty acid metabolism genes and triglycerides, increasing β can highlight the role of mutual information, which is suitable for precision medicine or etiological research.

[0103] The association correction coefficient γ is used to prevent misjudgments of association due to data sparsity. For example, if the co-occurrence sample size of a rare feature (such as a specific HLA genotype) and prediabetes is small, the lift and mutual information are close to 0. Setting γ=0.1 can force a lower limit on the association, avoiding the omission of genetic risk factors due to insufficient samples. Low-frequency risk factors with clear guideline evidence are retained, such as family history, which, although not high-frequency, has extremely high clinical value. The Sigmoid function is used to correct this. Mapping the result of the linear combination to the interval 0 to 1 makes the correlation R Fi, Fj It has probabilistic interpretation; the larger the value, the stronger the association.

[0104] Improvement Based on the frequency of feature co-occurrence, only high-frequency co-occurrence risk combinations can be identified, such as obesity and hyperglycemia frequently occurring together in the data. For example, suppose sleep apnea (Fi) and insulin resistance (Fj) co-occur in only 5% of the data, but physiologically, sleep apnea damages pancreatic β cells through hypoxia, and their mutual information is high because sleep apnea is known to significantly reduce the uncertainty of insulin resistance. In this case, the lift might be close to 1, the co-occurrence frequency is low, but the mutual information term MI is high. Fi, Fj Even after weighting with β, the correlation degree R can still be maintained. Fi, Fj Maintaining high values ​​helps avoid missing diagnoses of risk factors that are clearly relevant in clinical guidelines but are low-frequency in the data.

[0105] The frequent itemsets of risk features corresponding to the feature correlation are sorted in descending order, and the frequent itemsets of risk features with higher feature correlation are selected to obtain the diabetes prediction feature set.

[0106] In this embodiment, all frequent itemsets of risk features are sorted in descending order based on feature correlation scores. A stable ranking algorithm is used to ensure that itemsets with the same correlation are sorted a second time according to clinical priority (e.g., metabolic itemsets take precedence over behavioral itemsets). Higher-order itemsets (≥3 variables) are weighted and boosted to enhance the ranking priority of complex interaction patterns. The top N highly correlated itemsets are selected from the sorted frequent itemsets of risk features as the final input feature set.

[0107] Taking the discovery of combined features in community diabetes risk screening as an example, in a certain community screening project, traditional methods only identified high BMI and high blood sugar as independent risk factors. This approach discretizes residents' health data: BMI is divided into normal, overweight, and obese; dietary habits are divided into high sugar intake, high fat intake, and irregular eating; and sleep is divided into short-term fragmented and normal. Based on the improved Apriori algorithm, the frequent itemset obesity, high sugar intake, and fragmented sleep were found to have a support of 0.8%. Although the frequency is low, it was rated as having the highest correlation (0.72) by the feature importance evaluation algorithm (information gain), significantly higher than a single variable. This combination was retained as the core item of the diabetes prediction feature set. Subsequently, the random forest model, with this feature set as input, analyzed the diabetes...

[0108] The AUC in the pre-disease stage increased by 0.15, and this combination was identified by risk path analysis as the starting point of the causal chain of fat accumulation, insulin resistance, circadian rhythm disorder, and glucose metabolism imbalance, directly guiding the community intervention strategy to be upgraded from weight loss education to a diet-sleep synergistic intervention package.

[0109] This embodiment provides a method for predicting early diabetes based on health data. It utilizes feature association analysis to perform offline calculations on diabetes risk-related data, obtaining a diabetes prediction feature set. Through feature transformation and discretization, continuous variables are converted into semantically expressible intervals, enabling the formal identification of implicit nonlinear interaction patterns (such as obesity, hyperglycemia, and sleep disorders). An improved Apriori algorithm, through dynamic support adjustment and clinical knowledge pruning, effectively addresses the sparsity and long-tail distribution of health data, enhancing the sensitivity to mining low-frequency, high-risk combinations. Subsequently, a feature importance assessment algorithm is used to weight frequent itemsets, assigning them a predictive contribution at the pathological mechanism level (feature association), thereby filtering out statistically significant but biologically meaningless pseudo-associations. Finally, highly associated combination itemsets are selected to form the diabetes prediction feature set, transforming the model input from univariate linear features to a clinically interpretable multi-factor interaction structure. This process significantly improves the information density and pathological relevance of the feature set, providing a biologically reasonable input foundation for subsequent random forest attribution and causal path analysis, fundamentally solving the prediction drift and interpretability problems caused by coarse feature selection in traditional methods.

[0110] In one embodiment, a risk assessment is performed on a diabetes prediction feature set using a pre-defined diabetes prediction model to obtain diabetes risk prediction results, including:

[0111] Data normalization and feature encoding are performed on the diabetes prediction feature set to obtain the model input data.

[0112] The diabetes prediction feature set can be a high signal-to-noise ratio feature subset used for diabetes risk prediction, and can be used to carry screened diabetes-related risk variables. Data normalization is a preprocessing technique that scales continuous features with different dimensions and distributions to a uniform numerical range. It can be used to avoid gradient update imbalances caused by differences in dimensions during model training, improving convergence speed and stability. Feature encoding is a mapping process that converts categorical or textual variables into numerical vector forms. It can be used to integrate non-numerical health information into the mathematical space of deep learning models, supporting semantic-level feature fusion. In this embodiment, normalization is performed on continuous variables in the diabetes prediction feature set, and encoding mapping is performed on discrete variables, uniformly converting them into numerical vector forms. This achieves numerical unification of multi-source heterogeneous features, providing a basis for deep learning.

[0113] Degree-based learning models provide trainable input representations.

[0114] The model input data is divided into training set, validation set and test set.

[0115] The model input data can be a standardized numerical representation of a diabetes prediction feature set after data normalization and feature encoding. This ensures that information of different dimensions and data types is treated equally during model training, avoiding gradient explosion or convergence bias. In this embodiment, the model input data is randomly or hierarchically divided into three mutually exclusive subsets according to a preset ratio, used for training, optimization, and final evaluation, respectively. This establishes an independent data evaluation system, ensuring the generalization of the model training process and the reproducibility of the results.

[0116] A deep learning-based diabetes prediction model is constructed, which includes a feature extraction layer, a risk assessment layer, and a result output layer.

[0117] The diabetes prediction model can be a multi-layer neural network model based on a deep learning architecture. This model learns a non-linear representation of diabetes risk from standardized input and outputs a risk prediction result. It can be used to achieve automatic feature learning and risk probability output for high-dimensional, heterogeneous health data, improving the ability to model complex interaction patterns. In this embodiment, a three-layer neural network architecture is designed and initialized, responsible for feature representation learning, higher-order dependency modeling, and final prediction output, respectively. This constructs an end-to-end model with hierarchical feature learning capabilities, overcoming the limitations of traditional models in expressing non-linear interactions.

[0118] The training set is fed into a deep learning-based diabetes prediction model for training.

[0119] The training set can be a subset of diabetes prediction features used to train the parameters of the deep learning model. This subset allows the model to learn the statistical correlation between inputs and outputs, forming initial predictive capabilities. In this embodiment, the model parameters are iteratively updated using the backpropagation algorithm, causing the model output to gradually approximate the true labels in the training set. This allows the model to learn complex patterns of diabetes risk from the data, thus forming preliminary predictive capabilities.

[0120] The validation set is input into the model through the diabetes prediction loss function to fine-tune the parameters and generate the optimal prediction model.

[0121] The diabetes prediction loss function can be a mathematical function used to measure the difference between the model's predicted output and the true label, and can be used to guide the model training direction. In this embodiment, the loss is continuously calculated using the validation set during training. When the loss no longer decreases, the current model parameters are saved as the optimal version to avoid overfitting the model on the training set and ensure that the final model has the best generalization performance on unknown data.

[0122] The formula for the diabetes prediction loss function is shown below:

[0123]

[0124] In the formula, For diabetes prediction loss function, Here are the model parameters, and N is the number of samples. Let be the true label of the i-th sample, where 1 represents disease and 0 represents health. Let be the predicted probability of the i-th sample, λ be the regularization coefficient, M be the number of model layers, θk be the model parameters of the k-th layer, and δ be the loss function correction coefficient.

[0125] It should be noted that cross-entropy and losses The model is forced to output high probabilities for actual patients, ensuring that it prioritizes reducing missed diagnoses (misclassifying high-risk individuals as healthy) and misdiagnoses (misclassifying healthy individuals as high-risk). By adjusting the model's output threshold, such as classifying a predicted probability >0.7 as high risk, a trade-off between missed diagnoses and misdiagnoses can be struck, adapting to different screening scenarios, such as community screening focusing on reducing missed diagnoses and clinical diagnosis focusing on reducing misdiagnoses. The regularization coefficient λ is used to penalize complex parameter combinations, preventing the model from overfitting to noise in electronic health records (EHRs), such as misclassifying a temporary increase in blood glucose caused by stress as a persistent risk. For example, when λ=0.01, the model will be more inclined to select stable features such as BMI + fasting blood glucose, rather than noisy features such as random blood pressure fluctuations. In scenarios with poor EHR data quality (containing many missing values ​​and outliers), increasing λ strengthens feature selection and improves the model's predictive stability for new patients. Multi-layer neural networks can capture complex interactions between features, but too many layers can easily lead to overfitting. L2 regularization forces the model to learn more concise feature combinations by constraining the norm of the parameters in each layer, such as retaining only clinically interpretable interactions. When the cross-entropy loss approaches 0 or 1, it nears infinity, potentially leading to gradient explosion, such as a sample being misclassified by the model due to data errors. Setting δ = le-8 limits the loss value to a finite value, ensuring stable training and preventing individual outliers from dragging down the overall model. In medical data, outliers are often real clinical events, such as stress-induced hyperglycemia in the emergency room. The existence of δ allows the model to reasonably handle such extreme but reasonable data, rather than being completely dominated by them.

[0126] The test set is input into the optimal prediction model for performance evaluation to obtain diabetes risk prediction results, including the probability of disease and risk level.

[0127] The test set can be a subset of diabetes prediction features used to ultimately evaluate the overall performance of the model. It can also be used to provide unbiased estimates of the model's predictive ability, ensuring the evaluation results are accurate.

[0128] The results are reproducible. In this embodiment, test set data that was not involved in training and tuning is used as input to the optimal model to obtain its continuous probability and discrete level results, providing quantifiable and interpretable final risk assessment results as high-precision input for subsequent risk attribution analysis.

[0129] Taking a prediabetes screening clinic at a tertiary hospital as an example, this hospital introduced the deep learning prediction system described in this embodiment for prediabetes screening. The system receives 200 health features from EHR and wearable devices, which are normalized and embedded to form a unified input vector. The model uses a Transformer feature extraction layer to capture the nonlinear coupling between blood glucose fluctuation patterns and sleep structure, and a graph neural network risk assessment layer to identify the three-factor synergistic pathway of low nighttime activity, elevated insulin resistance index, and high fructose intake. FocalLoss is used during training to handle the imbalance problem where the proportion of prediabetes samples is less than 5%. An early stopping mechanism on the validation set ensures that the model achieves an AUC of 0.91 in the external queue. The test set outputs the probability and risk level of each patient's prediabetes. One patient has a probability of 0.87, indicating a high risk level, but their family history is negative. This result is passed to the subsequent random forest attribution module, which finds that the high risk is mainly due to circadian rhythm disorder and decreased insulin sensitivity caused by long-term nighttime work, rather than genetic factors. Based on this, the doctor developed a plan to adjust the patient's daily routine and provide morning light exposure, rather than recommending medication. Three months later, the patient's insulin sensitivity returned to normal levels.

[0130] For example, in the context of a regional public health big data platform, a provincial health big data platform integrated annual physical examination and wearable device data from 500,000 residents across the province and used the model in this embodiment for large-scale diabetes risk prediction. The model adopted a stratified time division: 2018-2020 as the training set, 2021 as the validation set, and 2022 as the test set, simulating real-time evolution. The model's output of disease probability and risk level was used for community-level hierarchical management. The system identified a high-risk subgroup with low BMI but high nighttime heart rate and high consumption of processed foods, with a disease probability of 0.78, but traditional models misclassified it as low risk. Subsequent risk path analysis revealed that this subgroup had a latent pathway of autonomic nervous system dysfunction, decreased metabolic flexibility, and ectopic fat deposition. Based on this, the platform updated its intervention guidelines, included this subgroup in the priority screening list, and pushed customized dietary and lifestyle recommendations, significantly reducing the proportion of this group progressing to diabetes within two years.

[0131] This embodiment provides a method for predicting early diabetes based on health data. It assesses the risk of a diabetes prediction feature set using a pre-defined diabetes prediction model, obtains model input data through data normalization and feature encoding, establishes an independent data evaluation system by dividing the data into training, validation, and test sets, and constructs a diabetes prediction system based on deep learning.

[0132] The model achieves automatic high-order representation learning of the diabetes prediction feature set. Through strict partitioning of the training, validation, and test sets and an early stopping mechanism, it ensures stable generalization ability on independent data. The diabetes prediction loss function is optimized to address sample imbalance, significantly improving the sensitivity to rare prediabetes states. Performance is evaluated by inputting the test set into the optimal prediction model, yielding diabetes risk prediction results, including the probability of disease and risk level. This improves prediction accuracy and robustness, providing high-quality, low-noise prediction input for subsequent risk attribution and confidence propagation-risk path analysis based on random forests. This supports the technical effectiveness of accurate identification of core risk sources and the scientific validity of intervention plans.

[0133] In one embodiment, key risk factors are extracted from the diabetes risk prediction results to obtain risk factor features; a risk factor classification model based on random forest is constructed, and the risk factor features are input into the risk factor classification model for risk attribution analysis to obtain the main risk factors, including:

[0134] Risk factors were extracted from high-risk samples in the diabetes risk prediction results to obtain risk factor characteristics.

[0135] The diabetes risk prediction results can be based on a diabetes prediction feature set inputted into a trained machine learning model, outputting a continuous risk score or a discrete risk level. These results can be used as a basis for screening high-risk samples to focus on the target population for subsequent attribution analysis. In this embodiment, by screening individual samples with diabetes risk prediction results above the high-risk threshold, model interpretation analysis is performed on their original input features to extract variables that significantly contribute to the prediction results, thus obtaining risk factor features. These may include high-contribution single variables, nonlinear interaction variable pairs, threshold-sensitive variables, etc.

[0136] Construct a risk factor classification model based on random forest.

[0137] Among them, the risk factor classification model based on random forests is an ensemble learning model composed of multiple decision trees, used to rank and classify the importance of risk factor features. It can involve retraining the random forest model on a subset of high-risk samples, with each tree built on randomly sampled features and sample subsets, and calculating feature importance through the reduction in Gini impurity during node splits. It can include feature importance evaluation based on the Gini index, feature ranking based on permutation importance, and factor weight calculation based on mean-reduced accuracy.

[0138] The risk factor features are input into the risk factor classification model for training, and the importance ranking of risk factors is obtained.

[0139] Specifically, the built-in mechanism of the random forest model calculates the reduction in Gini impurity resulting from splitting each feature across all trees and then aggregates and sorts them. The Gini index can be used as the splitting criterion to calculate the average importance score for each feature.

[0140] The model is interpreted using SHAP values ​​to generate the contribution of risk factors.

[0141] The SHAP value, based on the Shapley value model interpretation method of game theory, is used to quantify the marginal contribution of each feature to a single prediction result. Specifically, it can be used to calculate the SHAP value of the input features for each high-risk sample, quantifying the marginal impact of each factor on the predicted output of that sample.

[0142] The contribution of risk factors is quantitatively analyzed by using a risk factor weight calculation function to obtain the risk factor weight value.

[0143] The risk factor weighting function is a mathematical function that transforms the contribution of SHAP (Self-Assessment and Risk Factors) into a clinically interpretable weight. It can be achieved by standardizing the SHAP contribution and then adding clinical weight coefficients (such as epidemiological relative risk ratios or guideline recommendation weights) or expert scores to form a comprehensive weight value.

[0144] The formula for the risk factor weight calculation function is shown below:

[0145]

[0146] In the formula, W f Here, T represents the weight of risk factor f, and T represents the number of decision trees. (Gini coefficient) f, t Ginitotal represents the decrease in the Gini index of feature f in the t-th tree. t Let ΔAccuracy be the total reduction in the Gini index of the t-th tree. f, t ΔAccuracymax represents the decrease in accuracy of the t-th tree after removing feature f. t ϵ represents the maximum accuracy decrease for the t-th tree, and ϵ is the weight correction coefficient.

[0147] It should be noted that the number of decision trees, T, can be used to reduce the variance of a single tree by averaging the results of 100 trees. For example, if a tree overestimates the importance of exercise frequency due to sampling bias, averaging across multiple trees can offset the bias, making the weight calculation more stable. This avoids the random conclusions of a single model and ensures the repeatability of the weight ranking of major risk factors (such as BMI and blood glucose), which can serve as a basis for clinical guideline recommendations. The number of decision trees, T, can measure the contribution of feature f to sample splitting in that tree. For example, if fasting blood glucose is preferentially split in a tree, reducing the Gini impurity of child nodes by 0.3 (total reduction of 1.0), then the contribution of blood glucose in that tree is significant.

[0148] The contribution rate was 30%, reflecting its ability to directly distinguish high-risk groups.

[0149] Gini(f, t) is used to measure the direct contribution of feature f to rapidly distinguishing high-risk individuals for diabetes. For example, if glycated hemoglobin (HbA1c) ≥ 6.5% is used as a splitting condition, increasing the purity of diabetic patients in child nodes from 50% to 80%, its Gini value... f, t The levels were significantly higher than vague features such as "breakfast carbohydrate intake," indicating that HbA1c is a strong discriminant. (Ginitotal) t This is used to avoid overestimating the absolute contribution of secondary features in complex trees (with many splits) and underestimating the contribution of core features in simple trees (with fewer splits). For example, if tree t1 has a total Gini reduction of 100 (focusing on core metrics) and tree t2 has a reduction of 300 (including splits with noisy features), and a certain feature contributes 50 (50%) in t1 and 100 (33%) in t2, the former has a higher actual importance after normalization, reflecting its key role in a concise and efficient model. ΔAccuracy f, t This is used to measure the irreplaceability of feature f. Even if it is not a strong splitting feature, a significant drop in model performance after its removal indicates that it contains crucial implicit information. For example, a family history of diabetes may have a low splitting priority in a decision tree due to incomplete data (small Gini contribution), but after removal, the accuracy drops from 80% to 60% (Δ=20%), indicating that it carries key information about genetic risk and is an important basis for risk stratification. ΔAccuracymax t This standardizes the rate of accuracy decline, making the dependence of different features comparable within the 0-1 range and avoiding interference from absolute numerical differences. Some features (such as the insulin resistance index) may not be frequently used due to high testing costs, but their accuracy decline after removal is close to the maximum, suggesting that they are core indicators of deep pathological mechanisms and should be promoted for routine clinical testing.

[0150] The main risk factors are obtained by sorting the risk factor weights from largest to smallest.

[0151] In this study, all risk factors are ranked in descending order of their weights, and the top-ranked factors are selected as the primary risk factors. These factors may include metabolic dominant factors, behavioral dominant factors, and genetic-environment interaction factors. For example, in the case of precise attribution and stratified intervention for high-risk individuals with prediabetes, a health management platform screened 5,000 individuals at high risk of prediabetes, and the model predicted 500 high-risk individuals. The system only performed attribution analysis on these 500 individuals: using the SHAP value to calculate the individual contribution of 12 characteristics such as blood glucose, BMI, family history, and diet, it was found that the contribution of blood glucose fluctuations in some individuals far exceeded that of BMI; subsequently, the risk factor weight calculation function was used to multiply the mean SHAP value by the clinical guideline weight (e.g., family history × 1.8) to obtain the comprehensive weight; the ranking result was: blood glucose level (0.92) > BMI.

[0152] (0.78) > Family history (0.65) > Dietary habits (0.59). Based on this, the platform recommends continuous blood glucose monitoring services to those with the highest contribution to blood glucose levels, customized nutritionist intervention packages to those with both family history and poor diet, and exercise prescriptions to those whose BMI is the primary factor. Compared to traditional uniform intervention, this program improved intervention adherence by 41% and increased the rate of improvement in glycated hemoglobin by 33% after 3 months.

[0153] This embodiment provides a method for predicting early diabetes based on health data. By extracting key risk factors from diabetes risk prediction results, a risk factor classification model based on random forest is constructed. The importance of risk factor features is ranked and classified, and the contribution of risk factors is quantitatively analyzed using SHAP values ​​and risk factor weight calculation functions. Finally, the main risk factors are obtained, including blood glucose levels, BMI, family history, and dietary habits. This process shifts risk attribution from the average contribution of the population to the identification of core drivers in high-risk individuals, solving the core pain point of traditional models that know the risk is high but do not know why. It provides an operable, interpretable, and traceable decision-making basis for the precise formulation of personalized intervention plans, significantly improving the clinical translation efficiency and resource allocation rationality of early diabetes prevention.

[0154] In one embodiment, risk factors are extracted from high-risk samples in the diabetes risk prediction results to obtain risk factor features, including:

[0155] By applying thresholds to the diabetes risk prediction results, a dataset of high-risk individuals was selected.

[0156] The diabetes risk prediction results can be derived by inputting a diabetes prediction feature set into a trained machine learning model, outputting a continuous risk score or a discrete risk level. This can be used as a basis for screening high-risk samples, focusing on the target population for subsequent attribution analysis. In this embodiment, the diabetes risk prediction results are used to filter out a dataset of high-risk individuals through threshold division.

[0157] In this embodiment, the diabetes risk prediction results are thresholded to filter out a high-risk individual dataset. This high-risk individual dataset can be a set of individual samples with a significant risk of developing diabetes, selected based on a threshold set in the diabetes risk prediction results. It can be used as the input basis for local feature correlation analysis, avoiding the masking effect of the overall sample mean on the specific association patterns of high-risk subgroups. In this embodiment, by setting a threshold for the risk prediction probability, individual data exceeding this threshold are divided into independent subsets for subsequent analysis.

[0158] Furthermore, feature correlation analysis is performed on the high-risk individual dataset to extract factors significantly associated with the onset of diabetes, resulting in a risk factor feature dataset. This risk factor feature dataset can be a set of variables significantly associated with the onset of diabetes in this subgroup, extracted after feature correlation analysis of the high-risk individual dataset. It can be used to carry risk-driving signals specific to the high-risk group, serving as direct input for constructing a feature database and subsequent clinical screening. In this embodiment, the nonlinear correlation strength between each feature and diabetes status is calculated on the high-risk individual dataset, and significantly correlated variables are retained to form a feature subset.

[0159] In this embodiment, a risk factor feature database is constructed, and the risk factor feature dataset is saved to this database. The risk factor feature database can be a persistent data system for structured storage, versioned management, and standardized retrieval of the risk factor feature dataset. It can be used to achieve traceability, reusability, and cross-model retrieval of feature data, supporting the standardization and iterative optimization of subsequent analysis processes. In this embodiment, a table structure is established through a database management system, using fields such as individual ID, feature name, association strength, source model, and collection time as indexes, to persistently write and optimize the indexing of the risk factor feature dataset.

[0160] Furthermore, feature data is obtained from the risk factor feature database and combined with a clinical knowledge base for feature screening to obtain risk factor features. The clinical knowledge base can be a structured medical knowledge system composed of medical guidelines, evidence-based research, expert consensus, and epidemiological databases, used to guide the semantic rationality judgment of feature screening. It can provide evidence of pathophysiological mechanisms and clinical operability, filtering out statistically significant but medically meaningless pseudo-correlation features. In this embodiment, candidate features are extracted from the feature database and matched with operable factors and pathological mechanism labels in the clinical knowledge base, retaining variables that simultaneously meet the conditions of statistical relevance and clinical interpretability.

[0161] Taking the optimization of precision feature engineering in a regional diabetes prevention center as an example, a regional diabetes prevention center used a model to screen 100,000 residents for risk, predicting 1,200 high-risk individuals. Traditional methods directly used the characteristics of all 1,200 individuals for attribution, finding that seasonal weight fluctuations were statistically significant but clinically insignificant. This proposed solution first screened 387 individuals at extremely high risk based on a prediction probability > 0.8; mutual information analysis was performed on this subset, revealing that the standard deviation of nighttime blood glucose was significantly associated with diabetes progression in this group (information gain = 0.42), while the overall...

[0162] The indicator was not significant in the local analysis; the feature and candidate set were stored in the feature database; then, the ADA guideline knowledge base was called to filter out variables without intervention pathways, such as weekend drinking frequency, and retain 6 factors with pathological mechanism support, including nighttime blood glucose fluctuations, insulin resistance index, and prolonged sitting >8 hours / day; finally, a high-quality risk factor feature set was formed. This feature set was used for subsequent SHAP attribution and risk profile construction, upgrading the intervention plan from recommended weight loss to nighttime blood glucose monitoring, bedtime resistance training, and dinner carbohydrate control. After three months, the glycated hemoglobin of this group decreased by 2%, which was significantly higher than that of the control group.

[0163] This embodiment provides a method for predicting early diabetes based on health data. By screening high-risk individual datasets, performing feature correlation analysis, constructing a risk factor feature database, and combining it with a clinical knowledge base for feature screening, it can identify highly specific, non-global risk drivers in high-risk subgroups, achieve systematic management of feature data, and ensure that the final risk factor features have both statistical significance and medical interpretability.

[0164] In one embodiment, a method combining confidence propagation analysis and risk path analysis is used to comprehensively evaluate the main risk factors, thereby identifying core risk sources, including:

[0165] The main risk factors are standardized to obtain standard risk factor data.

[0166] The primary risk factors can be several key variables identified through random forest model analysis that contribute the most to the prediction of diabetes risk. These can be used as the input set for causal analysis, focusing on the variables with the greatest predictive influence and avoiding noise interference caused by an overly broad analysis scope. Primary risk factors can include metabolic dominant factors, behavioral dominant factors, and genetic-environment interaction factors. Standardizing the primary risk factors yields standardized risk factor data, which is a standardized dataset with the same dimensions and distribution characteristics formed by unifying the numerical scale of the primary risk factors. Standardized risk factor data can be obtained by Z-score standardization or Min-Max normalization methods, making its mean 0 and variance 1, or scaling it to the [0, 1] interval to eliminate modeling bias caused by differences in units, magnitudes, and distribution patterns. Standardized risk factor data can include Z-score standardized data, Min-Max normalized data, and robustly scaled data.

[0167] Risk path analysis techniques are used to analyze the causal relationships between standard risk factors, construct a causal network diagram of risk factors, and obtain an initial risk path map.

[0168] Among them, risk path analysis technology can be a graph structure analysis method based on causal inference, used to identify direct and indirect causal links between risk factors and reveal the potential of disease development.

[0169] Path analysis techniques can include structural equation modeling, Granger causality tests, and causal discovery algorithms. In this embodiment, risk path analysis is used to analyze the causal relationships between standard risk factors, resulting in a risk factor causal network graph. This graph can be a directed graph structure consisting of major risk factors as nodes and directional relationships determined by causal analysis as edges, used to characterize the causal dependency paths between factors. The risk factor causal network graph can be generated by learning the causal structure of standard risk factor data through structural equation modeling, PC algorithms, or FCI algorithms, outputting a set of edges with directionality and statistical significance, forming an acyclic directed graph. Risk factor causal network graphs can include structural equation-based causal graphs, constraint-based causal graphs, and score-based causal graphs.

[0170] By using a pre-defined confidence propagation analysis technique, the risk factors in the initial risk path map are probabilistically inferred to obtain the core risk sources.

[0171] The preset confidence propagation analysis technique can be a confidence propagation implementation method with defined parameters and algorithm flow, used to perform probabilistic inference on a fixed graph structure to identify core risk sources. The preset confidence propagation analysis technique can include Belief Propagation or junctiontree algorithms, setting message passing rules, convergence thresholds, and iteration counts to update the probability distribution of the initial risk path graph. In this embodiment, by using the preset confidence propagation analysis technique to perform probabilistic inference on the risk factors in the initial risk path graph, the core risk sources can be obtained. These can be the set of factors with the largest marginal impact and propagation strength identified in the initial risk path graph through confidence propagation analysis. Core risk sources can include single dominant risk sources, interactive superposition risk sources, and delayed effect risk sources, etc. Taking a community prediabetes intervention project as an example, a community conducted risk source tracing intervention on 200 people with impaired glucose tolerance. The system collected their main risk factors: fasting insulin, nighttime sleep duration, carbohydrate content at dinner, sedentary time, and family history. After Z-score standardization, a causal network graph was constructed using the PC algorithm, identifying two pathways: dinner carbohydrate percentage, postprandial blood glucose peak, insulin resistance and nighttime sleep duration, cortisol, and insulin sensitivity. After topological validation to remove weak edges, an initial risk path map was formed. Probabilistic reasoning using the belief propagation algorithm revealed that nighttime sleep duration had the highest marginal impact in the propagation process, and its influence path covered both insulin resistance and metabolic rhythm pathways, thus being identified as the core risk source. While family history is strongly associated with diabetes, it did not have a significant impact in the propagation chain and was excluded as a non-driving factor. Therefore, the intervention program focused on sleep improvement.

[0172] The group benefited from positive practices (such as light regulation and sleep hygiene education) rather than simply recommending dietary control. After three months, the group showed a 41% greater increase in insulin sensitivity compared to the control group.

[0173] This embodiment provides a method for predicting early diabetes based on health data. By combining confidence propagation analysis and risk path analysis techniques, it comprehensively assesses major risk factors to identify core risk sources. Through standardization, construction of causal network diagrams of risk factors, and probabilistic inference, it achieves dynamic quantification of risk propagation paths in the causal chain, accurately locating upstream factors that play a key driving role in disease evolution, rather than merely statistically correlated intermediate or downstream variables. This mechanism upgrades causal analysis from static structure identification to dynamic impact propagation assessment, effectively distinguishing between relevant and driving factors. It avoids misjudging confounding variables (such as age) or mediating variables (such as insulin resistance) as intervention targets, enabling intervention strategies to accurately target upstream nodes in disease evolution. This represents a paradigm shift from statistical correlation screening to causal-driven intervention, significantly improving the targeting and clinical translation efficiency of personalized prevention measures.

[0174] In one embodiment, a pre-defined confidence propagation analysis technique is used to perform probabilistic reasoning on the risk factors in the initial risk path map to obtain the core risk sources, including:

[0175] A confidence propagation analysis technique is constructed, which includes Bayesian network modeling, probabilistic inference, and maximum a posteriori probability estimation.

[0176] The confidence propagation analysis technique can be used to transform static causal structures into quantifiable and propagable probabilistic inference systems, enabling a reasoning framework for the joint assessment of multi-factor collaborative risks. It can be used to chain together three sub-modules: Bayesian network modeling, message-passing probabilistic inference, and maximum a posteriori probability estimation, forming an end-to-end causal probabilistic inference pipeline. In this embodiment, the confidence propagation analysis technique can utilize open-source probabilistic graph libraries such as pgmpy or PyMC3, defining the graph structure, specifying the CPT, calling the belief propagation function and MAP solver, and encapsulating them as standardized modules. Alternatively, it can be custom-implemented: building graph structure classes, message-passing engines, and MAP searchers based on C and / or Python, supporting custom prior distributions and convergence thresholds.

[0177] The initial risk path graph is probabilistically modeled using Bayesian network modeling to obtain the prior probabilities of nodes.

[0178] Bayesian network modeling can be used to formalize causal structures into computable probabilistic models, supporting subsequent message passing and joint distribution inference as probabilistic graphical models.

[0179] The initial risk path graph is a topological structure, with each risk factor as a node. The direction of the edges is defined according to their causal orientation, and the conditional probability table (CPT) for each node is estimated based on clinical priors and data distribution. In this embodiment, Bayesian network modeling can be based on priors set from clinical literature and cohort statistics: for example, positive family history = 0.3, BMI > 30 = 0.42, using empirical distributions to fill the CPT. Alternatively, a non-parametric Bayesian method can be used: for continuous factors (such as insulin levels), a Gaussian mixture model is used to fit the prior distribution, and for discrete factors, a Dirichlet prior is used.

[0180] Confidence propagation is performed based on the prior probabilities of nodes. According to the conditional dependencies between risk factors, the posterior probability of each node is updated through a message passing algorithm to obtain the joint probability distribution of risk factors.

[0181] The prior probabilities of nodes can be set based on epidemiological data or clinical expert experience, such as setting the prior probability of a positive family history to 0.3 and insufficient sleep to 0.45, based on the initial probability distribution of each risk factor independently determined by historical cohort statistics or literature consensus. It can be understood that the prior probabilities of nodes provide the initial state beliefs for the Bayesian network and are the starting point for probability propagation. In this embodiment, confidence propagation based on the prior probabilities of nodes can be performed using the Sum-Product algorithm: passing messages from leaf nodes to root nodes and then propagating back to calculate the marginal probability of each factor. Alternatively, it can employ...

[0182] LoopyBeliefPropagation: For graph structures with approximate loops, iterative propagation is performed in multiple rounds to obtain an approximate joint distribution within a finite number of steps.

[0183] Maximum a posteriori probability estimation is performed on the joint probability distribution of risk factors to identify the combination of risk factors that contributes the most to the incidence of diabetes, thus obtaining the core risk source.

[0184] The joint probability distribution of risk factors can be a probability distribution function of the common values ​​of all risk factors under given observation conditions. It can be understood that the joint probability distribution of risk factors integrates the prior probabilities and conditional dependencies of each node through a message passing algorithm to calculate the joint posterior probability of all factor combinations. This can be used to fully characterize the risk state space under the synergistic effect of multiple factors, providing a probabilistic basis for MAP estimation. In this embodiment, the maximum posterior probability estimation of the joint probability distribution of risk factors can be performed using a greedy search: fixing high-probability factors one by one, gradually expanding the combinations, and approximating the MAP solution within a finite search space.

[0185] Alternatively, the branch and bound method can be used: construct a state space tree, prune low-probability branches, and accurately solve high-dimensional MAP problems.

[0186] This embodiment transforms the initial risk path map into a computable probabilistic inference system by constructing a confidence propagation analysis technique. First, a Bayesian network model encodes the causal structure into a directed graph with conditional probability tables, setting node prior probabilities based on clinical priors and data distribution. Second, a message passing algorithm dynamically propagates conditional dependency information within the graph, updating the posterior probabilities of each risk factor and comprehensively characterizing their joint probability distribution, achieving accurate modeling of the multi-factor synergistic state. Finally, maximum posterior probability estimation searches for the factor combination that maximizes the probability of diabetes incidence in the high-dimensional joint distribution, identifying risk modules with statistical significance and biological synergy. This mechanism overcomes the limitations of traditional single-factor attribution, achieving a paradigm shift from identifying isolated high-risk factors to mining synergistic pathogenic patterns. It allows core risk sources to no longer rely on the importance of a single variable but rather manifest as clinically interventionizable risk combinations, supporting the development of multi-target synergistic intervention plans. This enables a clinical paradigm shift from broad-spectrum prevention to precise blocking, significantly improving the targeting, effectiveness, and cost-effectiveness of early intervention.

[0187] In one embodiment, an individual diabetes risk profile is constructed based on core risk sources. This profile is then used to further understand the relationships between various risk factors in order to develop personalized early intervention plans, including:

[0188] We conduct multi-dimensional analysis of core risk sources to construct an individual diabetes risk profile that includes physiological indicators, lifestyle, and genetic factors.

[0189] The individual diabetes risk profile can be a personalized risk representation model, presented in a dynamic network structure, incorporating physiological indicators, lifestyle, and genetic factors. It can provide interpretable and interactive visualization of the risk structure, supporting clinicians in developing and adjusting intervention strategies based on the dynamic dependencies between factors, thus transitioning from static profiling to adaptive management. This can include, but is not limited to, network topology-based risk profiles, time-series evolution-based risk profiles, and multi-dimensional heatmap-based risk profiles. In this embodiment, the core risk sources are obtained by comprehensively evaluating the main risk factors using confidence propagation and risk path analysis techniques. This can be a composite analysis method integrating probabilistic graphical reasoning and causal discovery, used to identify path structures with statistical significance and causal transmission force within the risk factor set. It can be used to further distinguish between direct driving factors and indirectly related factors based on the identified main risk factors, avoiding misclassification of mediating or confounding variables as intervention targets. Furthermore, the core risk sources are obtained by comprehensively evaluating the main risk factors using confidence propagation and risk path analysis techniques. This can be achieved by using Bayesian networks to calculate the posterior probability change magnitude of each factor.

[0190] By combining the direct causal in-degree identified by the PC algorithm, a comprehensive score is calculated: Score = α × propagation strength and β × causal in-degree. The Top-N factors are selected as core risk sources, or the average causal effect (ACE) of each factor is estimated using do-calculus in the causal graph. This ACE is then weighted and ranked with the node influence in the confidence propagation, retaining factors where both are above a threshold. This allows for dual verification of risk factors, ensuring that the identified core risk sources possess both strong statistical correlation and conform to biologically significant causal transmission logic, avoiding misjudgment of mediating variables or environmental confounding factors.

[0191] Personalized early intervention plans are developed based on individual diabetes risk profiles, including dietary adjustment recommendations, exercise plans, and blood glucose monitoring frequency.

[0192] Individual diabetes risk profiles can be built based on core risk sources and dynamically updated to develop personalized intervention plans. This can include, but is not limited to, synchronizing wearable device and EHR data quarterly, recalculating the levels and interaction weights of core risk sources, adjusting network parameters using incremental Bayesian update algorithms to avoid full retraining, or building a version control mechanism for the risk profile, generating a new version snapshot with each update, marking the implementation time of intervention measures and effect feedback (such as HbA1c changes), forming an intervention-response closed-loop record. This transforms the risk profile from a one-time assessment result into a continuously evolving dynamic management tool, supporting real-time adjustments to intervention plans as individual conditions change, improving the adaptability and effectiveness of long-term prevention.

[0193] By regularly tracking individual health data and intervention effects, and dynamically updating risk profiles and intervention plans, precise prevention and management of early-stage diabetes can be achieved.

[0194] Regularly tracking individual health data and intervention effectiveness can be achieved by continuously monitoring an individual's diabetes risk profile, combining changes in individual physiological, behavioral, and genetic attributes with the effectiveness of intervention measures, and dynamically adjusting the risk profile and intervention plan. This can include, but is not limited to, regularly collecting individual physiological indicators, lifestyle data, and genetic information, using confidence propagation and risk path analysis techniques to update the risk profile, and adjusting the intervention plan based on the updated risk profile to achieve precise prevention and management of early-stage diabetes.

[0195] Taking the long-term health management of patients with prediabetes as an example, the method for predicting early diabetes based on health data in this embodiment can be as follows: A 45-year-old prediabetic patient was identified as having high frequency of nighttime eating and poor sleep quality as core risk sources in the initial assessment, and risk mapping was performed.

[0196] The data showed a direct causal chain between the two factors through circadian rhythm disruption and increased insulin resistance. Based on this, doctors developed an intervention plan that restricted eating two hours after dinner and light exposure one hour before bedtime. Three months later, the patient uploaded new sleep monitoring data and a blood glucose trend chart, triggering an automatic profile update: it found a decrease in the level of nighttime eating nodes, but an increase in cortisol fluctuation nodes, and a new causal link between stress eating and cortisol. The profile was then updated to reflect the new pathway of stress, elevated cortisol, and recurrence of nighttime eating. The system automatically recommended adding a psychological stress management module (such as mindfulness training) and adjusting the blood glucose monitoring frequency from once a month to once every two weeks. This process did not require manual model reconstruction; the profile continuously optimized the intervention strategy through closed-loop feedback, achieving precise and dynamic management.

[0197] This embodiment provides a method for predicting early diabetes based on health data. By constructing an individual diabetes risk profile and developing a personalized early intervention plan based on the risk profile, and by regularly tracking individual health data and intervention effects, the risk profile and intervention plan can be dynamically updated to achieve the technical effect of precise prevention and management of early diabetes.

[0198] Furthermore, embodiments of the present invention also propose a storage medium storing a program for predicting early diabetes based on health data. When the program for predicting early diabetes based on health data is executed by a processor, it implements the steps of the method for predicting early diabetes based on health data as described above.

[0199] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for predicting early diabetes based on health data, characterized in that, The method includes: Data collection and preprocessing of individual health systems yielded diabetes risk-related data. Feature association analysis technology was used to perform offline calculations on diabetes risk-related data to obtain a diabetes prediction feature set. The risk assessment of the diabetes prediction feature set is performed using a pre-defined diabetes prediction model to obtain diabetes risk prediction results. Key risk factors were extracted from the diabetes risk prediction results to obtain risk factor characteristics; a risk factor classification model based on random forest was constructed, and the risk factor characteristics were input into the risk factor classification model for risk attribution analysis to obtain the main risk factors. A comprehensive assessment of the main risk factors was conducted using a combination of confidence propagation analysis and risk path analysis techniques to identify the core risk sources. Based on the core risk sources, an individual diabetes risk profile is constructed. Through this profile, the relationships between various risk factors can be further understood in order to develop personalized early intervention plans.

2. The method for predicting early diabetes based on health data as described in claim 1, characterized in that, The data collection and preprocessing of the individual health system to obtain diabetes risk-related data includes: Individual health data is obtained by collecting data from multiple sources on an individual health system; Individual health data is cleaned, standardized, and feature-engineered to obtain diabetes risk-related data.

3. The method for predicting early diabetes based on health data as described in claim 1, characterized in that, The method utilizes feature association analysis technology to perform offline calculations on diabetes risk-related data to obtain a diabetes prediction feature set, including: By performing feature transformation and discretization on diabetes risk-related data, health risk characteristic data is obtained; By utilizing feature association mining based on the improved Apriori algorithm, strong association rules in health risk feature data are obtained, resulting in frequent itemsets of risk features. The association strength of frequent itemsets of risk features is calculated based on the feature importance assessment algorithm. We obtain the feature correlation degree by selecting association rules with high correlation strength; The frequent itemsets of risk features corresponding to the feature correlation are sorted in descending order, and the frequent itemsets of risk features with higher feature correlation are selected to obtain the diabetes prediction feature set.

4. The method for predicting early diabetes based on health data as described in claim 3, characterized in that, The formula for the feature correlation algorithm is as follows: In the formula, R Fi, Fj Let exp be the correlation between features Fi and Fj, α and β be the weighting coefficients, and P be the weighting coefficient. Fi and P Fj Let Fi and Fj be the probabilities of features Fi and Fj respectively, and let P(Fi ∩ Fj) be the probability of features Fi and Fj appearing simultaneously. Fi, Fj Let γ be the mutual information between features Fi and Fj, and let γ be the correlation correction coefficient.

5. The method for predicting early diabetes based on health data as described in claim 1, characterized in that, The step of performing a risk assessment on a diabetes prediction feature set using a pre-defined diabetes prediction model to obtain diabetes risk prediction results includes: Data normalization and feature encoding are performed on the diabetes prediction feature set to obtain the model input data; the model input data is then divided into training set, validation set and test set. Construct a diabetes prediction model based on deep learning, which includes a feature extraction layer, a risk assessment layer, and a result output layer; The training set is input into the deep learning-based diabetes prediction model for model training, and the validation set is input into the model for parameter tuning using the following diabetes prediction loss function to generate the optimal prediction model. The formula for the diabetes prediction loss function is shown below: In the formula, For diabetes prediction loss function, Here are the model parameters, and N is the number of samples. Let be the true label of the i-th sample, where 1 represents disease and 0 represents health. Let be the predicted probability of the i-th sample, λ be the regularization coefficient, M be the number of model layers, θk be the model parameters of the k-th layer, and δ be the loss function correction coefficient. The test set is input into the optimal prediction model for performance evaluation to obtain diabetes risk prediction results, including the probability of disease and risk level.

6. The method for predicting early diabetes based on health data as described in claim 1, characterized in that, The key risk factors of the diabetes risk prediction results are extracted to obtain risk factor characteristics. A risk factor classification model based on random forest is constructed. Risk factor characteristics are input into the risk factor classification model for risk attribution analysis to obtain the main risk factors, including: Risk factors were extracted from high-risk samples in the diabetes risk prediction results to obtain risk factor characteristics. A risk factor classification model based on random forest is constructed. Risk factor features are input into the risk factor classification model for training to obtain the importance ranking of risk factors. The SHAP value is used to interpret the model to generate the contribution of risk factors. The contribution of risk factors is quantitatively analyzed using the following risk factor weight calculation function to obtain the risk factor weight value; The formula for the risk factor weight calculation function is shown below: In the formula, W f Here, T represents the weight of risk factor f, and T represents the number of decision trees. (Gini coefficient) f, t Ginitotal represents the decrease in the Gini index of feature f in the t-th tree. t Let ΔAccuracy be the total reduction in the Gini index of the t-th tree. f, t ΔAccuracymax represents the decrease in accuracy of the t-th tree after removing feature f. t Let ϵ be the maximum accuracy decrease for the t-th tree, and ϵ be the weight correction coefficient. The risk factors were sorted by weight from largest to smallest to obtain the main risk factors, including blood sugar level, BMI, family medical history, and dietary habits.

7. The method for predicting early diabetes based on health data as described in claim 6, characterized in that, The risk factors are extracted from high-risk samples in the diabetes risk prediction results to obtain risk factor characteristics, including: By applying thresholds to the diabetes risk prediction results, a dataset of high-risk individuals was selected. Feature correlation analysis was performed on the high-risk individual dataset to extract factors significantly associated with the onset of diabetes, resulting in a risk factor feature dataset; a risk factor feature database was constructed, and the risk factor feature dataset was saved to the risk factor feature database; Feature data from the risk factor feature database is obtained, and feature filtering is performed in conjunction with the clinical knowledge base to obtain risk factor features.

8. The method for predicting early diabetes based on health data as described in claim 1, characterized in that, The method, which combines confidence propagation analysis and risk path analysis, comprehensively evaluates the main risk factors to identify core risk sources, including: The main risk factors are standardized to obtain standard risk factor data; Risk path analysis techniques are used to analyze the causal relationships between standard risk factors, construct a causal network diagram of risk factors, and obtain an initial risk path map. By using a pre-defined confidence propagation analysis technique, the risk factors in the initial risk path map are probabilistically inferred to obtain the core risk sources.

9. The method for predicting early diabetes based on health data as described in claim 8, characterized in that, The process involves using a pre-defined confidence propagation analysis technique to perform probabilistic reasoning on the risk factors in the initial risk path map to obtain the core risk sources, including: A confidence propagation analysis technique is constructed, which includes Bayesian network modeling, probabilistic inference, and maximum a posteriori probability estimation. The initial risk path graph is probabilistically modeled using Bayesian network modeling to obtain the prior probabilities of nodes. Confidence propagation is performed based on the prior probabilities of nodes. According to the conditional dependencies between risk factors, the posterior probability of each node is updated through a message passing algorithm to obtain the joint probability distribution of risk factors. Maximum a posteriori probability estimation is performed on the joint probability distribution of risk factors to identify the combination of risk factors that contributes the most to the incidence of diabetes, thus obtaining the core risk source.

10. The method for predicting early diabetes based on health data as described in claim 1, characterized in that, The process involves constructing an individual diabetes risk profile based on core risk sources. This profile allows for a deeper understanding of the relationships between various risk factors, enabling the development of personalized diabetes treatment plans. Early intervention programs include: A multi-dimensional analysis of core risk sources was conducted to construct an individual diabetes risk profile that includes physiological indicators, lifestyle, and genetic factors. Personalized early intervention plans are developed based on individual diabetes risk profiles, including dietary adjustment recommendations, exercise plans, and blood glucose monitoring frequency. By regularly tracking individual health data and intervention effects, and dynamically updating risk profiles and intervention plans, precise prevention and management of early-stage diabetes can be achieved.

Images