Multidimensional glucose index-driven interpretable ai weight management decision and intervention method

Abstract

This application discloses a multidimensional glucose index-driven interpretable AI-based weight management decision-making and intervention method, including the following steps: S1, acquiring multimodal data related to weight management for the target individual, including at least continuous glucose monitoring data, gut microbiota metabolomics data, behavioral data, and basic clinical indicators. Then, by integrating continuous glucose monitoring-derived indicators, gut microbiota metabolomics data, behavioral data, and basic clinical indicators, a multimodal standardized feature set is constructed, significantly improving the predictive accuracy of weight and body fat percentage changes. The SHAP value interpretability tool is used, providing both global and local interpretability, quantifying the positive and negative contributions of each feature to the prediction results, automatically prioritizing influencing factors based on the interpretability analysis results, and supporting dynamic report updates based on individual implementation status, thereby improving the long-term effectiveness of weight management.

Description

Technical Field

[0001] This application relates to the technical field of weight management, and more particularly to an interpretable AI-driven weight management decision-making and intervention method driven by multidimensional glucose indicators. Background Technology

[0002] Weight management is a key intervention strategy for preventing and improving chronic metabolic diseases such as obesity, diabetes, and cardiovascular disease. Traditional weight management methods mainly rely on dietary calorie control and physical activity guidance, such as setting a daily calorie deficit and recommending exercise duration. However, due to significant differences among individuals in genetic background, metabolic characteristics, gut microbiota composition, and lifestyle habits, standardized protocols often have inconsistent effects and are difficult to implement for precise and sustainable weight control.

[0003] Some studies have attempted to use machine learning models to predict weight changes. However, most existing models rely solely on dietary and exercise data and basic clinical indicators, neglecting the crucial regulatory role of individual insulin resistance and blood glucose fluctuations on energy metabolism and fat accumulation. These metabolic-related characteristics can reflect individual differentiated response patterns to diet and exercise, and are important early biomarkers for the effectiveness of weight management.

[0004] Meanwhile, most existing weight prediction models can only output prediction results, but cannot quantify the contribution of different physiological and behavioral characteristics to the prediction results. Clinicians or individual users find it difficult to understand the decision-making basis of the model, resulting in low trust and poor adherence to intervention recommendations. For example, the model may indicate that the predicted weight loss in the next month is not significant, but it cannot clearly point out whether the main limiting factor is excessive blood sugar fluctuations, insufficient short-chain fatty acid levels, or insufficient exercise intensity, thus making it difficult to formulate targeted improvement measures.

[0005] In addition, most existing methods output static dietary or exercise recommendations without prioritizing influencing factors based on interpretability analysis results or dynamically adjusting intervention strategies based on individual real-time execution feedback. This leads to situations where users may be overwhelmed by multiple recommendations in practice, or gradually lose effectiveness and adherence because the plan is not updated according to their own status.

[0006] Application content

[0007] This application aims to address, at least to some extent, the technical problems in the related art.

[0008] To achieve the above objectives, this application proposes a multidimensional glucose index-driven, interpretable AI-based weight management decision-making and intervention method, comprising the following steps:

[0009] S1. Obtain multimodal data related to weight management for the target individual, including at least continuous glucose monitoring data, gut microbiota metabolomics data, behavioral data and basic clinical indicators. The collection period shall be no less than 7 days, and the collection frequency of continuous glucose monitoring data shall be no less than once every 5 minutes. All data shall be associated with a unique individual identifier to establish an individual-specific data file.

[0010] S2. Preprocess the multimodal data to remove invalid, redundant and interference data. Extract core features and construct a standardized feature set according to the characteristics of different types of data. Continuous features are uniformly standardized and classification features are encoded. The feature set is validated for effectiveness.

[0011] S3. Based on the feature set, a multimodal fusion weight management effect prediction model is constructed and trained using machine learning algorithms. The core prediction targets are changes in weight and body fat percentage. The training set, validation set and test set are divided by stratified sampling, and the model parameters are optimized by cross-validation.

[0012] S4. Use interpretability tools to quantify the contribution of each feature to the prediction of weight management effect, taking into account both global and local interpretability. Global interpretability is achieved by ranking feature importance and global analysis of SHAP values ​​to quantify the contribution of each feature to the prediction of weight management effect. Local interpretability is achieved by analyzing the contribution of individual features to clarify the correlation between each feature of a specific individual and the prediction result.

[0013] S5. Input the preprocessed individual features into the trained interpretable AI prediction model to obtain short-term to long-term predicted values ​​of weight and body fat percentage changes. Combine the interpretability analysis results to generate an actionable intervention report containing key influencing factors, intervention priorities, and specific measures. The report can be dynamically updated based on individual implementation.

[0014] Specifically, the continuous glucose monitoring data in step S1 includes continuous blood glucose time series data. The calculated blood glucose fluctuation parameters include one or more of the following: blood glucose variability, area under the blood glucose curve, percentage of time when blood glucose reaches the target, and nighttime blood glucose trend. Blood glucose variability is quantified by standard deviation, coefficient of variation, and maximum blood glucose fluctuation amplitude. A percentage of time when blood glucose reaches the target is ≥70% is considered normal.

[0015] Specifically, the gut microbiome data in step S1 includes the abundance of microbial genes related to short-chain fatty acid synthesis, and the concentration of short-chain fatty acids in serum and feces; the short-chain fatty acids include at least propionic acid and butyric acid.

[0016] Specifically, the behavioral data in step S1 includes physical activity data monitored by smart wearable devices and dietary data obtained through a diet record APP. The physical activity data includes at least daily steps, duration of moderate and vigorous intensity exercise, and resting metabolic rate. The dietary data includes at least daily total calories and macronutrient intake. The basic clinical indicators include age, gender, body mass index, body fat percentage, and insulin resistance index, wherein an insulin resistance index ≥2.5 is defined as insulin resistance.

[0017] Specifically, step S2 includes: interpolating and imputing outliers in continuous glucose monitoring data and calculating blood glucose fluctuation characteristics; imputing missing values ​​in gut microbiota metabolomics and behavioral data using the mean or median and removing outliers using the Z-score method; applying Min-Max standardization to continuous features and one-hot encoding or label encoding to categorical features; and screening significant features through correlation analysis to ensure the effectiveness of the feature set.

[0018] Specifically, in step S3, the machine learning algorithm is gradient boosting tree, random forest or neural network, and the ratio of training set, validation set and test set is 7:2:1; the model parameters are optimized by grid search and five-fold cross-validation until the average absolute error of the model on the validation set is ≤5%.

[0019] Specifically, in step S4, global interpretability involves: calculating the SHAP value of each feature, drawing a SHAP summary map and a feature importance ranking map, quantifying the contribution of glucose-related features to weight change, and identifying the top three core features in terms of contribution. Local interpretability involves: generating a SHAP force map for a single individual, showing the positive and negative contribution direction and absolute value of each feature to the prediction result of that individual, and outputting a decision path map.

[0020] Specifically, in step S5, the short-term prediction is 1-2 weeks, the medium-term prediction is 1-3 months, and the long-term prediction is 3-6 months; the key influencing factors include at least the amplitude of blood glucose fluctuation, short-chain fatty acid concentration, daily steps, duration of moderate-intensity exercise, insulin resistance index, and total calorie intake; the intervention priority is ranked according to the average size of the absolute value of SHAP, with core influencing factors receiving priority intervention and secondary factors receiving auxiliary intervention.

[0021] In summary, the beneficial effects of the multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method of this application are as follows:

[0022] 1. By integrating continuous glucose monitoring-derived indicators, gut microbiota metabolomics data, behavioral data, and basic clinical indicators, a multimodal standardized feature set is constructed, which makes up for the shortcomings of traditional models that rely solely on behavioral and basic indicators, and significantly improves the predictive accuracy of changes in weight and body fat percentage.

[0023] 2. The SHAP value interpretability tool is adopted, which provides both global and local interpretability. It quantifies the positive and negative contributions of each feature to the prediction results and their magnitude, enabling users to understand the basis of the prediction and identify key constraints. This overcomes the problems of low trust and poor compliance caused by the lack of interpretability of existing models.

[0024] 3. Based on the interpretability analysis results, the system automatically prioritizes the influencing factors and generates an actionable intervention report that includes key influencing factors, intervention priorities, and specific measures. It also supports dynamic updates of the report based on individual implementation, avoiding the shortcomings of static recommendations that have poor adaptability and leave users at a loss, thus improving the effectiveness of long-term weight management. Attached Figure Description

[0025] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0026] Figure 1 The flowchart is for the interpretable AI weight management decision-making and intervention method driven by multidimensional glucose index in this application.

[0027] Figure 2 The flowchart for training the machine learning model of the interpretable AI weight management decision-making and intervention method driven by multidimensional glucose index in this application is shown.

[0028] Figure 3 This is a flowchart illustrating the feature standardization and coding process of the interpretable AI weight management decision-making and intervention method driven by multidimensional glucose indicators in this application. Detailed Implementation

[0029] To make the technical means, inventive features, objectives, and effects of this application easier to understand, the application is further described below with reference to specific illustrations. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0030] The present application will now be described in further detail with reference to the accompanying drawings.

[0031] like Figures 1-3 As shown in the embodiments of this application, the multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method includes the following steps:

[0032] S1. Obtain multimodal data related to weight management for the target individual, including at least continuous glucose monitoring data, gut microbiota metabolomics data, behavioral data, and basic clinical indicators. The collection period should be no less than 7 days, and the frequency of continuous glucose monitoring data collection should be no less than once every 5 minutes. All data should be associated with a unique individual identifier to establish an individual-specific data file.

[0033] Specifically, using the Abbott Continuous Glucose Monitoring (CGM) device for blood glucose data collection enables non-invasive, continuous 14-day blood glucose monitoring, effectively avoiding the limitations of traditional finger-prick blood sampling. After monitoring, core blood glucose fluctuation parameters are extracted, including the coefficient of variation, area under the blood glucose curve, time to reach target blood glucose levels, and nighttime blood glucose trends. Key event indicators are also extracted, including the duration of hyperglycemia and the slope of nighttime blood glucose decline, comprehensively quantifying the amplitude, duration, and diurnal variation characteristics of blood glucose fluctuations.

[0034] Behavioral data is monitored using smart devices and recorded via an app, with verification by a nutritionist. The diet record is recorded by the app to document the subject's 24-hour dietary situation. Subjects can record daily dietary details by taking photos, typing text, etc. The completeness and accuracy of the records are then verified by a professional nutritionist. After verification, daily core dietary indicators are calculated, including daily calorie intake and dietary fiber intake, to accurately quantify the impact of dietary structure on weight changes.

[0035] Physical activity monitoring uses smart bracelets to monitor subjects in real time without them noticing. The bracelets collect daily steps and duration of moderate to high intensity exercise. The built-in accelerometer and heart rate sensor can accurately distinguish exercise intensity, ensuring the objectivity and continuity of physical activity data.

[0036] Basic clinical indicators: Core physiological and clinical indicators of subjects were collected, including height, weight, body fat percentage (measured by bioelectrical impedance analysis), fasting blood glucose, and fasting insulin. The insulin resistance index was calculated by fasting blood glucose and insulin concentration. Subjects were followed up for 6 months, and their weight was remeasured every 3 months. An insulin resistance index ≥2.5 was defined as insulin resistance. This index was significantly pathologically associated with abnormal weight and blood glucose fluctuations.

[0037] S2. Preprocess the multimodal data to remove invalid, redundant and interference data. Extract core features and construct a standardized feature set according to the characteristics of different types of data. Continuous features are uniformly standardized and categorical features are encoded. The feature set is then validated for effectiveness.

[0038] Specifically, for CGM data: if the missing percentage of CGM data for a certain individual is >10%, the data is deemed unqualified and is removed. For locally missing data with a missing percentage of ≤10%, cubic spline interpolation is used to fill in the missing data. This method can effectively preserve the temporal continuity and trend characteristics of blood glucose data and avoid the impact of interpolation errors on subsequent analysis.

[0039] Behavioral data: Missing values ​​were imputed using the MICE algorithm with 5 iterations. The distribution characteristics of the missing data were simulated through multiple iterations to ensure the rationality of the imputed data and avoid bias caused by a single imputation method.

[0040] Outlier handling: The focus is on removing outliers from the target variable ΔBMI (Body Mass Index). Samples with ΔBMI exceeding ±3 standard deviations are removed. These samples represent extreme weight loss or gain, which do not conform to the normal weight change pattern. Removing them can effectively avoid interference from extreme data on model training and ensure the model's generalization ability.

[0041] Continuous variables, including GV, AUC, short-chain fatty acid concentration, daily calorie intake, and daily steps, were processed using Z-score standardization to map the data to a standard normal distribution (mean 0, standard deviation 1) to eliminate model bias caused by differences in the dimensions of different features.

[0042] Categorical variables: mainly gender, which are digitized using one-hot encoding to convert the categorical variables into binary vectors, avoiding the rank bias caused by label encoding and ensuring the model's accurate identification of categorical features.

[0043] S3. Based on the feature set, a multimodal fusion weight management effect prediction model is constructed and trained using machine learning algorithms. The core prediction targets are changes in weight and body fat percentage. The training set, validation set and test set are divided by stratified sampling, and the model parameters are optimized by cross-validation.

[0044] Specifically, multiple algorithms are used to build the model, including ensemble learning, deep learning and baseline models, and a multimodal fusion strategy is adopted to improve the model performance. The ensemble learning uses the XGBoost algorithm with the core parameters set as follows: learning rate 0.01, maximum depth 6, and early stopping method. This algorithm is good at capturing nonlinear correlations and high-order interactions between features and has good adaptability to tasks such as weight change that are affected by complex interactions of multiple factors.

[0045] The deep learning algorithm uses a long short-term memory network with the following core parameters: 2 LSTM layers (64 hidden units), a dropout rate of 0.3 (to prevent overfitting), and an Adam optimizer (learning rate lr=1e-4). This algorithm can effectively capture the temporal features of CGM data and explore the potential correlation between dynamic changes in blood glucose fluctuations and weight changes.

[0046] Baseline Model Comparison: Two baseline models, Logistic Regression and Random Forest, were set up for performance comparison. The superiority of the constructed model was verified by comparing it with the baseline models.

[0047] Two approaches, early fusion and late fusion, are employed to achieve efficient fusion of multi-dimensional data:

[0048] Early integration: CGM dynamic indicators, metabolomics data, behavioral data, and clinical indicators are concatenated into a single input matrix (120 dimensions), which is directly input into the model for training, realizing synchronous feature learning of multi-source data.

[0049] Late-stage fusion: Two sub-models are trained separately: an XGBoost model (input metabolomics + clinical indicators) and an LSTM model (input CGM time series data). Then, the weights of the two sub-models are optimized based on the AUC value of the validation set, and a weighted fusion is performed to give full play to the advantages of different models.

[0050] Model optimization employs a combined optimization strategy of hyperparameter tuning and feature selection to ensure optimal model performance.

[0051] Hyperparameter tuning: Bayesian optimization is adopted, with 50 iterations. The root mean square error or AUC value of the test set is used as the objective function to iteratively optimize the hyperparameters of various models. Compared with traditional grid search, Bayesian optimization can find the optimal parameter combination more efficiently and improve optimization efficiency.

[0052] Feature selection: Combining recursive feature elimination and SHAP value sorting, the top 20% of core features with the highest contribution are retained, further reducing feature dimensionality, reducing the interference of redundant information on the model, and improving the interpretability of the model.

[0053] S4. Utilize interpretability tools to quantify the contribution of each feature to the prediction of weight management effectiveness, taking into account both global and local interpretability. Global interpretability is achieved through feature importance ranking and global analysis of SHAP values ​​to quantify the contribution of each feature to the prediction of weight management effectiveness. Local interpretability is achieved through individual feature contribution analysis to clarify the correlation between each feature of a specific individual and the prediction results.

[0054] Specifically, the global interpretation first calculates the SHAP value of all samples. Based on the average of the absolute values ​​of the SHAP values, it determines the global contribution of each feature and selects the key predictors that have the greatest impact on weight change, such as GV, propionic acid concentration, and insulin resistance index (HOMA-IR). Then, it draws a feature importance ranking chart to show the top 10 core features in terms of contribution, intuitively presenting the strength of each feature's influence on weight change prediction. Finally, it generates a feature dependency chart to quantify the non-linear relationship between features and weight change.

[0055] For individuals at high risk of weight change, the local interpretation generates a personalized and interpretable report, identifying key influencing factors (including positive and negative factors) and providing targeted intervention recommendations: move dinner to before 6:00 PM to reduce nighttime carbohydrate intake and lower nighttime blood sugar fluctuations; supplement with 15g of inulin daily to promote the abundance of butyrate-producing bacteria in the gut and further increase butyrate concentration; maintain a daily step count of over 8000 steps and appropriately increase the duration of moderate-to-vigorous intensity exercise.

[0056] S5. Input the preprocessed individual features into the trained interpretable AI prediction model to obtain short-term to long-term predicted values ​​of weight and body fat percentage changes. Combine the interpretability analysis results to generate an actionable intervention report containing key influencing factors, intervention priorities, and specific measures. The report can be dynamically updated based on individual implementation.

[0057] Specifically, by inputting preprocessed individual features into a trained, interpretable AI prediction model, short- to long-term predicted values ​​for weight and body fat percentage changes can be obtained. This helps individuals and healthcare professionals anticipate the effectiveness of weight management and adjust intervention strategies accordingly. The intervention report generated from the interpretability analysis results must clearly identify key influencing factors, intervention priorities, and specific measures to ensure the intervention plan's relevance and operability. Intervention priorities clarify the order of interventions, while specific measures provide individuals with actionable guidelines. Furthermore, the report needs to be dynamically updated based on individual implementation, as an individual's physiological state and behavioral habits may change during the intervention process (e.g., improved blood sugar fluctuations, increased exercise). Dynamic updates ensure the intervention plan remains adapted to the individual's current state, enhancing the long-term effectiveness of weight management.

[0058] In one embodiment of this application, the continuous glucose monitoring data in step S1 includes continuous blood glucose time series data, and the calculated blood glucose fluctuation parameters include one or more of blood glucose variability, area under the blood glucose curve, percentage of time when blood glucose reaches the target, and nighttime blood glucose trend. Blood glucose variability is quantified by standard deviation, coefficient of variation, and maximum blood glucose fluctuation amplitude, and a blood glucose target time percentage ≥70% is considered normal.

[0059] It should be noted that blood glucose variability is assessed using a combination of three indicators: standard deviation, coefficient of variation, and maximum blood glucose variability. Standard deviation reflects the absolute degree of dispersion, the coefficient of variation eliminates individual differences in average blood glucose levels, and maximum blood glucose variability captures extreme fluctuation events.

[0060] Area under the blood glucose curve: The area enclosed by the blood glucose curve and the baseline over 24 hours is calculated. It reflects the total glucose load exposure and has a non-linear correlation with glycated hemoglobin, but is more sensitive to postprandial blood glucose peaks.

[0061] Percentage of time with blood glucose targets: This is usually defined as the percentage of time that blood glucose levels are within a certain range, with ≥70% as the normal threshold. It is significantly negatively correlated with the risk of microvascular complications.

[0062] Nocturnal blood glucose trends: The focus is on analyzing the slope, lowest point, and duration of blood glucose changes during the period from 0:00 to 06:00. Nocturnal hypoglycemia is an independent risk factor for sudden cardiac death, while nocturnal hyperglycemia suggests the dawn phenomenon or worsening insulin resistance.

[0063] In one embodiment of this application, the gut microbiome metabolomics data in step S1 includes the abundance of microbial genes related to short-chain fatty acid synthesis, and the concentration of short-chain fatty acids in serum and feces; the short-chain fatty acids include at least propionic acid and butyric acid.

[0064] It should be noted that short-chain fatty acids, as key products of gut microbial metabolism, are closely related to the structure of the gut microbiota. Propionate can regulate blood glucose homeostasis by participating in gluconeogenesis, while butyrate can provide energy for intestinal epithelial cells and maintain the integrity of the intestinal barrier. Both participate in physiological processes such as energy metabolism and inflammatory response, and are important intermediate mediators linking gut microbiota and weight changes. Therefore, they are used as core detection indicators for gut microbial metabolomics data.

[0065] In one embodiment of this application, the behavioral data in step S1 includes physical activity data monitored by a smart wearable device and dietary data obtained through a diet record APP. The physical activity data includes at least daily steps, duration of moderate and vigorous intensity exercise, and resting metabolic rate. The dietary data includes at least daily total calories and macronutrient intake. The basic clinical indicators include age, gender, body mass index, body fat percentage, and insulin resistance index, wherein an insulin resistance index ≥2.5 is defined as insulin resistance.

[0066] It should be noted that physical activity data should at least include daily steps, duration of moderate and vigorous intensity exercise, and resting metabolic rate; dietary data should at least include daily total calorie intake and macronutrient intake. Meanwhile, the basic clinical indicators collected in step S1 cover indicators closely related to weight changes and metabolic status. Among these, age and sex are fundamental physiological factors affecting metabolic rate, body mass index and body fat percentage are key indicators reflecting weight status and body composition, and the insulin resistance index is a core indicator for assessing abnormal glucose metabolism.

[0067] In one embodiment of this application, step S2 includes: interpolating and imputing outliers in continuous glucose monitoring data and calculating blood glucose fluctuation characteristics; imputing missing values ​​in gut microbiota metabolomics and behavioral data using the mean or median and removing outliers using the Z-score method; applying Min-Max standardization to continuous features and one-hot encoding or label encoding to categorical features; and screening significant features through correlation analysis to ensure the effectiveness of the feature set.

[0068] Specifically, for continuous glucose monitoring data, since data may be missing or outliers may occur during the monitoring process due to factors such as equipment failure or improper wearing, the interpolation imputation method is used to supplement the missing data, ensure the continuity of the data, avoid interference from abnormal data on the calculation of blood glucose fluctuation characteristics, and thus accurately calculate blood glucose fluctuation characteristics.

[0069] For gut microbiome metabolomics and behavioral data, considering the different missing mechanisms of different data types, continuous missing data were imputed using the mean or median to ensure the rationality of the imputed data. Then, to eliminate the dimensional differences between different features and ensure the fairness of model training, the Min-Max normalization method was used for continuous features (such as short-chain fatty acid concentration, daily steps, total calorie intake, etc.) to map feature values ​​to the [0,1] interval. For categorical features (such as gender, insulin resistance status, etc.), one-hot encoding or label encoding was used. One-hot encoding is suitable for categorical features with no order relationship, while label encoding is suitable for categorical features with an order relationship (such as insulin resistance level) to ensure that categorical features can be effectively identified by the model. Then, significant features were screened through correlation analysis, and the correlation coefficient between each feature and weight change was calculated. Features with insignificant correlation were removed to avoid redundant features increasing model complexity and affecting model performance.

[0070] In one embodiment of this application, the machine learning algorithm in step S3 is a gradient boosting tree, random forest, or neural network, and the ratio of the training set, validation set, and test set is 7:2:1; the model parameters are optimized by grid search and five-fold cross-validation until the average absolute error of the model on the validation set is ≤5%.

[0071] It should be noted that gradient boosting trees can accurately capture the nonlinear relationship between features and weight changes by iteratively training weak classifiers and gradually correcting errors; random forests can effectively reduce the risk of overfitting and improve model stability by integrating multiple decision trees; and neural networks can mine deep correlations between features through nonlinear transformations of multiple layers of neurons, making them suitable for multi-dimensional and highly complex data modeling.

[0072] The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for initial training of model parameters, the validation set is used for optimizing and adjusting model parameters, and the test set is used for final evaluation of the model's generalization ability. This division ratio conforms to the conventional norms for machine learning model training, and can effectively verify model performance while ensuring sufficient training data.

[0073] Grid search filters out the initial optimal parameters by traversing the preset parameter combinations. Five-fold cross-validation randomly divides the training set into 5 subsets, using 4 subsets as the training set and 1 subset as the validation set, repeating the training and validation process 5 times. The average of the 5 validation results is taken as the model performance evaluation index. The parameters are further optimized in this way until the model's mean absolute error on the validation set is ≤5%, ensuring that the model has high prediction accuracy and stability and can meet the needs of accurate prediction of weight change.

[0074] In one embodiment of this application, the global interpretability in step S4 specifically involves: calculating the SHAP value of each feature, drawing a SHAP summary map and a feature importance ranking map, quantifying the contribution of glucose-related features to weight change, and identifying the top three core features in terms of contribution; the local interpretability specifically involves: generating a SHAP force map for a single individual, showing the positive and negative contribution direction and absolute value of each feature to the prediction result of that individual, and outputting a decision path map.

[0075] Specifically, in global interpretability, the SHAP value quantifies the contribution of each feature to the prediction result. A positive value indicates that the feature promotes weight gain, while a negative value indicates that the feature inhibits weight gain. The larger the absolute value of the SHAP value, the stronger the influence of the feature. Based on the calculated SHAP values, a SHAP summary map (showing the distribution of SHAP values ​​for each feature across all samples, intuitively reflecting the overall influence trend of the features) and a feature importance ranking map (sorted by the average absolute value of SHAP values, clarifying the influence priority of each feature) are drawn, thereby quantifying the specific contribution of glucose-related features (such as blood glucose fluctuation range, mean blood glucose, etc.) to weight change. Local interpretability focuses on individual individuals. For each individual to be predicted, a SHAP force map is generated, and a decision path map is output, making the prediction results more convincing and providing precise guidance for individualized intervention plans.

[0076] In one embodiment of this application, the short-term prediction in step S5 is 1-2 weeks, the medium-term prediction is 1-3 months, and the long-term prediction is 3-6 months; the key influencing factors include at least the amplitude of blood glucose fluctuation, short-chain fatty acid concentration, daily steps, duration of moderate-intensity exercise, insulin resistance index, and total calorie intake; the intervention priority is ranked according to the average size of the absolute value of SHAP, with core influencing factors receiving priority intervention and secondary factors receiving auxiliary intervention.

[0077] It should be noted that the short-term prediction is for 1-2 weeks, during which weight changes are mainly affected by recent behavioral factors such as diet and exercise, and the prediction results can be used to adjust short-term lifestyle in a timely manner; the medium-term prediction is for 1-3 months, during which weight changes gradually show a stable trend, which can reflect the synergistic effect of multiple factors such as diet, exercise, and gut microbiota, and can be used to evaluate the effectiveness of phased intervention programs; the long-term prediction is for 3-6 months, during which weight changes can reflect the long-term effects of intervention strategies and can be used to formulate long-term weight management goals.

[0078] Factors with larger average absolute values ​​of SHAP have a more significant impact on weight change and are identified as core influencing factors, which are given priority for intervention (e.g., for individuals with high insulin resistance index, interventions to improve insulin resistance are given priority). Factors with smaller average absolute values ​​of SHAP are considered secondary factors and are used for auxiliary intervention.

[0079] It should be noted that, in this document, the terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0080] The present application and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present application. The actual structure is not limited to this. In conclusion, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present application, such design should fall within the protection scope of the present application.

Claims

1. A multidimensional glucose indicator driven interpretable Al weight management decision and intervention method, characterized in that, Includes the following steps: S1. Obtain multimodal data related to weight management for the target individual, including at least continuous glucose monitoring data, gut microbiota metabolomics data, behavioral data and basic clinical indicators. The collection period shall be no less than 7 days, and the collection frequency of continuous glucose monitoring data shall be no less than once every 5 minutes. All data shall be associated with a unique individual identifier to establish an individual-specific data file. S2. Preprocess the multimodal data to remove invalid, redundant and interference data. Extract core features and construct a standardized feature set according to the characteristics of different types of data. Continuous features are uniformly standardized and classification features are encoded. The feature set is validated for effectiveness. S3. Based on the feature set, a multimodal fusion weight management effect prediction model is constructed and trained using machine learning algorithms. The core prediction targets are changes in weight and body fat percentage. The training set, validation set and test set are divided by stratified sampling, and the model parameters are optimized by cross-validation. S4. Use interpretability tools to quantify the contribution of each feature to the prediction of weight management effect, taking into account both global and local interpretability. Global interpretability is achieved by ranking feature importance and global analysis of SHAP values ​​to quantify the contribution of each feature to the prediction of weight management effect. Local interpretability is achieved by analyzing the contribution of individual features to clarify the correlation between each feature of a specific individual and the prediction result. S5. Input the preprocessed individual features into the trained interpretable AI prediction model to obtain short-term to long-term predicted values ​​of weight and body fat percentage changes. Combine the interpretability analysis results to generate an actionable intervention report containing key influencing factors, intervention priorities, and specific measures. The report can be dynamically updated based on individual implementation.

2. The multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method according to claim 1, characterized in that, The continuous glucose monitoring data in step S1 includes continuous blood glucose time series data. The calculated blood glucose fluctuation parameters include one or more of the following: blood glucose variability, area under the blood glucose curve, percentage of time when blood glucose reaches the target, and nighttime blood glucose trend. Blood glucose variability is quantified by standard deviation, coefficient of variation, and maximum blood glucose fluctuation amplitude. A percentage of time when blood glucose reaches the target is ≥70% is considered normal.

3. The multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method according to claim 1, characterized in that, The gut microbiome data in step S1 include the abundance of genes related to short-chain fatty acid synthesis in microorganisms, and the concentration of short-chain fatty acids in serum and feces; the short-chain fatty acids include at least propionic acid and butyric acid.

4. The multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method according to claim 1, characterized in that, In step S1, the behavioral data includes physical activity data monitored by smart wearable devices and dietary data obtained through a diet record APP. The physical activity data includes at least daily steps, duration of moderate and vigorous intensity exercise, and resting metabolic rate. The dietary data includes at least daily total calories and macronutrient intake. The basic clinical indicators include age, gender, body mass index, body fat percentage, and insulin resistance index, wherein an insulin resistance index ≥2.5 is defined as insulin resistance.

5. The multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method according to claim 1, characterized in that, Step S2 includes: interpolating and imputing outliers in continuous glucose monitoring data and calculating blood glucose fluctuation characteristics; imputing missing values ​​in gut microbiota metabolomics and behavioral data using the mean or median and removing outliers using the Z-score method; using Min-Max standardization for continuous features and one-hot encoding or label encoding for categorical features; and screening significant features through correlation analysis to ensure the effectiveness of the feature set.

6. The multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method according to claim 1, characterized in that, In step S3, the machine learning algorithm is gradient boosting tree, random forest or neural network, and the ratio of training set, validation set and test set is 7:2:1; the model parameters are optimized by grid search and five-fold cross-validation until the average absolute error of the model on the validation set is ≤5%.

7. The multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method according to claim 1, characterized in that, The global interpretability in step S4 specifically involves: calculating the SHAP value of each feature, drawing a SHAP summary diagram and a feature importance ranking diagram, quantifying the contribution of glucose-related features to weight change, and identifying the top three core features in terms of contribution. The local interpretability specifically refers to: generating a SHAP force map for a single individual, showing the positive and negative contribution direction and absolute value of each feature to the prediction result of that individual, and outputting a decision path map.

8. The multidimensional glucose index-driven interpretable AI weight management decision-making and intervention method according to claim 1, characterized in that, In step S5, the short-term prediction is 1-2 weeks, the medium-term prediction is 1-3 months, and the long-term prediction is 3-6 months. The key influencing factors include at least the amplitude of blood glucose fluctuation, short-chain fatty acid concentration, daily steps, duration of moderate-intensity exercise, insulin resistance index, and total calorie intake. The intervention priority is ranked according to the average size of the absolute value of SHAP, with core influencing factors receiving priority intervention and secondary factors receiving auxiliary intervention.

Images