An ai-based personal obesity risk prediction system and method based on big data

Abstract

The application discloses a kind of AI's personal obesity risk prediction system and method based on big data, belong to medical health field, including data collection module, multi-dimensional feature construction module, risk label dynamic generation module, model training and risk prediction module, credibility evaluation and calibration module etc..System is collected multi-source heterogeneous data by wearable device, biochemical interface and health platform API, and is constructed behavior-metabolism-environment coupling feature after uniform coding as standard time sequence sequence.Adopt graph neural network to combine causal factor decomposition, and the joint modeling of dynamic feature and label is carried out, and the output individual obesity risk prediction result is output individual obesity risk prediction result.Through Monte Carlo Dropout and confidence interval calibration mechanism, the result credibility is improved, and calibration information is fed back to feature construction link to optimize modeling strategy.Finally, in the form of visualized heat map and causal path diagram, the key risk factor is presented, and individualized intervention suggestion vector is generated.Profitable effect: improve prediction accuracy, enhance individual intervention pertinence.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical health, more particularly, to an AI-based personal obesity risk prediction system and method based on big data. BACKGROUND

[0002] In recent years, with the popularity of wearable devices, health apps and vital sign monitoring systems, personal health data collection has become increasingly abundant, and health risk prediction driven by big data has become a research hotspot. Traditional methods are mainly based on linear regression, decision trees or shallow neural networks, combined with a small number of behavioral or physiological characteristics for modeling, which can achieve basic risk classification, but have significant limitations in processing high-dimensional heterogeneous data, multi-factor interaction and individual differences. In recent years, the introduction of graph neural networks and causal reasoning technology has promoted the development of health prediction models to a deeper level, enabling preliminary attempts at complex correlation modeling and causal factor identification.

[0003] However, existing systems generally have the following shortcomings: first, the modeling structure is single and difficult to integrate graph structure information in multi-source data; second, the influence of time series shift and group heterogeneity is ignored, leading to biased prediction results; third, there is a lack of causal factor screening mechanism, making the prediction results difficult to explain; fourth, the credibility of risk prediction is not explicitly calibrated, and a stable feedback mechanism cannot be formed. Therefore, there is an urgent need for an AI system that integrates multi-graph modeling, causal analysis and risk calibration mechanism to improve the accuracy, interpretability and adaptability of obesity risk prediction. SUMMARY

[0004] The present application aims to provide an AI-based personal obesity risk prediction system and method based on big data to solve the problems raised in the background art: first, the modeling structure is single and difficult to integrate graph structure information in multi-source data; second, the influence of time series shift and group heterogeneity is ignored, leading to biased prediction results; third, there is a lack of causal factor screening mechanism, making the prediction results difficult to explain; fourth, the credibility of risk prediction is not explicitly calibrated, and a stable feedback mechanism cannot be formed.

[0005] Technical solution: An AI-based personal obesity risk prediction system based on big data includes a data collection module, a multi-dimensional feature construction module, a risk label dynamic generation module, a model training and risk prediction module, a feedback closed-loop optimization module, a credibility evaluation and calibration module, a visual explanation and intervention suggestion module;

[0006] The data collection module adopts a multi-source heterogeneous data collection mechanism, synchronously collects gait change data, physiological time sequence parameters, dietary intake frequency logs, indoor and outdoor environment parameters, and social interaction logs through wearable devices, biochemical detection interfaces, and third-party health platform API interfaces, and uniformly encodes them into five types of standard structured time sequence sequences and transmits them to the multi-dimensional feature construction module; the multi-dimensional feature construction module takes the encoded standard time sequence sequence as input, constructs a multi-channel nested attention structure based on residual connection, fuses the physiological change frequency curve and the behavior trigger node through the atlas nesting strategy, and outputs the behavior-metabolism-environment coupling feature set to the risk label dynamic generation module; the risk label dynamic generation module takes the continuous feature input of more than three months as a window, generates a continuous risk state label set with a recursive clustering-adaptive boundary optimization strategy, and synchronously labels the feature evolution path to form a semi-supervised training data set;

[0007] The model training and risk prediction module jointly models the coupling features and dynamic labels based on a graph neural network-causal factor decomposition structure, extracts the interaction causal factors between long-term behavior paths and short-term physiological mutations through a structured regular enhanced graph convolution layer, and outputs an obesity risk prediction score;

[0008] The credibility evaluation and calibration module performs posterior calibration on each model output based on the Monte Carlo Dropout and confidence interval adjustment mechanism, and the calibrated results are transmitted to the feedback closed-loop optimization module; the feedback closed-loop optimization module takes the cross-stage calibration bias and model overfitting label as the core input, automatically selects mismatched features and adjusts the feature construction strategy in reverse;

[0009] The visual explanation and intervention suggestion module constructs an interactive risk path atlas based on a structural visualization heat map and high-weight intervention factors to generate individual-oriented intervention suggestion vectors.

[0010] Preferably, the multi-dimensional feature construction module includes a three-layer nested structure, the first layer is an input reconstruction layer, which performs periodic smoothing on each type of time sequence data, uses a 3x1 convolution kernel and a bidirectional LSTM to reconstruct the behavior frequency data in time, the second layer is a multi-head nested attention structure, including a gait-heart rate joint attention head, a physiological-environment fluctuation collaborative attention head, and a social-eating cross attention head, each attention head extracts the mutual dependence relationship in the short term based on the self-attention mechanism and window convolution and dynamically weights; the third layer is a feature alignment layer, which uses the maximum information coefficient rearrangement technique to reorder the multi-head output according to the causal coupling strength, and outputs the behavior-metabolism-environment three-type coupling feature group through a skip residual connection and labels the window structure to the next module.

[0011] Preferably, the maximum information coefficient rearrangement technology is based on a nonlinear mapping relationship between multiple categories of feature groups, uses a weighted dynamic time warping algorithm under a sliding window to align different feature sequences, calculates a multi-order mutual information gain, extracts the strongest interconnected node path in a mapping sequence graph constructed based on spectral clustering, and finally rearranges the coupled feature sequence length to make the output features completely consistent with the label generation window in time.

[0012] Preferably, the risk label dynamic generation module generates individual dynamic risk labels through an evolution clustering graph convolution structure under a dense time window. First, the multi-dimensional coupled feature sequence is subjected to 3-dimensional dimension reduction processing, principal component analysis, maximum variance preserving projection, and embedded graph kernel dimension reduction. Then, an evolution graph is constructed based on the graph structure density change rate, TopK density peak marking is performed on each time slice, and the node aggregation of the peak marking is subjected to risk state layering through an adaptive boundary shrinkage function. Finally, the output label matrix has point-by-point continuity on the time axis.

[0013] Preferably, the adaptive boundary shrinkage function is composed of a dynamic boundary sliding factor and a local label density function.

[0014] The dynamic boundary sliding factor is composed of the product of the time step and the feature volatility rate, and dynamically compresses and expands the clustering boundary. The local label density function adjusts the boundary threshold of each risk label based on the sample distribution tightness in the current time window, and finally realizes the continuous expression of the individual risk level in multiple time periods.

[0015] Preferably, the model training and risk prediction module includes a multi-graph fusion graph neural network structure and a causal factor decomposition unit.

[0016] The multi-graph fusion graph neural network structure includes four types of graphs, namely, behavior graph, physiological fluctuation graph, social influence graph, and environmental fluctuation graph. Each type of graph adopts a joint input form of an adjacency matrix and a node embedding matrix. The graph neural network adopts a structured attention mechanism to update the node representation in layers and outputs the global embedding in an aggregated center node weight mean manner.

[0017] The causal factor decomposition unit performs Laplace feature filtering and minimum mutual information causal stripping processing on all graph embedding results, and finally outputs a causal weighted prediction score vector.

[0018] Preferably, the model training and risk prediction module includes a position nested attention factor and a group bias compensation factor. The position nested attention factor constructs a position offset graph based on the timestamp difference, applies a strong edge update weight to nodes close in time. The group bias compensation factor adjusts the inter-group behavior feature mean deviation as an adjustment amount to standardize the inter-group structure deviation.

[0019] Preferably, the minimum mutual information causal disentanglement processing comprises the following steps:

[0020] Firstly, a Bayesian structure learning mechanism is used to construct a feature causal graph, and a Markov equivalence class clustering technique is used to filter out non-causal structure paths; then, an inverse probability weighting mechanism is introduced to resample the sample space, so that the latent confounding factors are distributed similarly after normalization in the sample; finally, a reparameterization technique is used to perform second-order supervised reinforcement training on the disentangled feature subset, and a high-resolution causal factor is generated.

[0021] Preferably, the credibility evaluation and calibration module is based on a multiple forward prediction mechanism to construct a confidence interval, uses Monte Carlo Dropout to sample 100 model prediction results and calculates the range, mean and skewness coefficient of each output, then fits the distribution shape of the calibrated prediction score through a beta distribution, and outputs a risk level adjustment factor with the upper confidence limit as the decision limit, and finally forms a feedback signal with the calibrated risk vector as the input.

[0022] Preferably, the AI-based personal obesity risk prediction method based on big data comprises the aforementioned AI-based personal obesity risk prediction system based on big data.

[0023] Compared with the prior art, the application has the following advantages:

[0024] (1) Four types of heterogeneous graphs, including behavior graph, physiological fluctuation graph, social influence graph and environmental fluctuation graph, are introduced to enhance the expression ability of complex factor interaction, and the structural description ability is stronger than that of traditional single feature model.

[0025] (2) A hierarchical structured attention mechanism is used to optimize the information transmission and aggregation process of the graph neural network, improve the recognition ability of key nodes and edges, and thus improve the accuracy and interpretability of risk prediction.

[0026] (3) Laplace feature filtering and minimum mutual information disentanglement operation are introduced to extract feature factors with stronger causal correlation from mixed information, and the rigor and model generalization ability of causal modeling are improved compared with existing empirical weight methods.

[0027] (4) For the node time sequence offset problem in time series data, a position offset graph is constructed and a strong edge update is applied to improve the modeling accuracy of dynamic correlation between time approximate events.

[0028] (5) The structural bias is alleviated by normalizing the mean deviation of cross-population behavior, which alleviates the model prediction deviation caused by population heterogeneity and improves the adaptability of the model to different body types, populations and lifestyles.

[0029] (6) Using inverse probability weighting and reparameterization techniques to optimize the training of feature causal graph, so that the output causal factor has stronger discriminant ability and fine-grained expression.

[0030] (7) The prediction confidence interval is constructed by adopting Monte Carlo Dropout combined with Beta distribution fitting, and the upper confidence limit is set as the risk grading basis, solving the problem that the prediction result credibility cannot be quantified in the existing system.

[0031] (8) The calibrated risk vector is used as feedback input to form a closed-loop dynamic correction mechanism, improving the continuous learning and adaptive ability of the system in real use scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a whole system schematic diagram of the AI-based personal obesity risk prediction system based on big data; DETAILED DESCRIPTION

[0033] Embodiment, please refer to Figure 1 An AI-based personal obesity risk prediction system based on big data includes a data collection module, a multi-dimensional feature construction module, a risk label dynamic generation module, a model training and risk prediction module, a feedback closed-loop optimization module, a credibility evaluation and calibration module, a visual explanation and intervention suggestion module;

[0034] The data collection module adopts a multi-source heterogeneous data acquisition mechanism, synchronously acquires gait change data, physiological time sequence parameters, dietary intake frequency logs, indoor and outdoor environment parameters and social interaction logs through wearable devices, biochemical detection interfaces and third-party health platform API interfaces, and uniformly encodes the five types of standard structured time sequence to pass to the multi-dimensional feature construction module; the multi-dimensional feature construction module takes the encoded standard time sequence as input, constructs a multi-channel nested attention structure based on residual connection, fuses the physiological change frequency curve and the behavior trigger node through a graph embedding strategy, and outputs a behavior-metabolism-environment coupled feature set to the risk label dynamic generation module; the risk label dynamic generation module takes three months or more continuous features as a window, generates a continuous risk state label set using a recursive clustering-adaptive boundary optimization strategy, and synchronously labels the feature evolution path to form a semi-supervised training data set;

[0035] The model training and risk prediction module jointly models the coupled features and dynamic labels based on a graph neural network-causal factor decomposition structure, extracts the interaction causal factors between long-term behavior paths and short-term physiological mutations through a structured regular enhanced graph convolution layer, and outputs an obesity risk prediction score;

[0036] The credibility evaluation and calibration module performs posterior calibration on each model output based on Monte Carlo Dropout and confidence interval adjustment mechanism, and the calibrated results are transmitted to the feedback closed-loop optimization module; the feedback closed-loop optimization module takes the cross-stage calibration bias and model overfitting mark as the core input, automatically selects the mismatched features and reversely adjusts the feature construction strategy;

[0037] The visual explanation and intervention suggestion module constructs an interactive risk path atlas based on the structural visual heat map and high-weight intervention factors, and generates individual-oriented intervention suggestion vectors.

[0038] Specifically, the data collection module adopts the following configuration:

[0039] The wearable device is Fitbit Charge5 (a fitness tracker under the Google brand), and data is collected every 5 minutes;

[0040] The collected physiological parameters include heart rate (HeartRate), skin temperature (SkinTemperature), and blood oxygen saturation (SpO2). All signals are synchronized and converted into 60Hz frequency time series signals;

[0041] The behavior log is obtained through an API interface.

[0042] Data preprocessing includes:

[0043] Outlier rejection: based on the 3σ principle to detect threshold abnormalities;

[0044] Linear interpolation filling: linear fitting of 5-minute interval values before and after;

[0045] The sliding normalization adopts the following formula:

[0046] ;

[0047] Where, : the normalized value at time ;

[0048] : the original observed value at time ;

[0049] : the sliding window mean with as the center and window radius ;

[0050] : the standard deviation of the corresponding sliding window;

[0051] : sliding window radius ( )

[0052] : Prevents extremely small positive numbers (close to zero) in the denominator ).

[0053] The multi-dimensional feature construction module takes the encoded standard sequence as input, and the structure is as follows:

[0054] The nested attention structure is three layers, and the structure is as follows:

[0055] The number of multi-head is 8, and the structure of each attention head is as follows:

[0056] ;

[0057] Where, : query matrix (obtained by inputting Dense(64));

[0058] : key matrix (obtained by inputting Dense(64));

[0059] : value matrix (obtained by inputting Dense(64));

[0060] : The dimension of the key vector (consistent with your Dense(64) setting );

[0061] : Exponential normalization function by row, used to produce attention weights;

[0062] : Similarity matrix (dot product) of query and key;

[0063] : Scaling factor, used to stabilize the gradient and probability distribution;

[0064] Position encoding uses standard sine-cosine functions:

[0065] ;

[0066] ;

[0067] Where, , : Position Encoding values in even / odd dimensions;

[0068] : Discrete position index in the sequence;

[0069] : Dimension "pair" index (each pair contains two dimensions of odd and even) starting from 0;

[0070] : Total embedding dimension of the model (vector dimension that positional encoding acts on);

[0071] Constant : Wavelength scale base for positional encoding to cover multi-scale frequencies;

[0072] Feature residual connection: connected to the input of the next layer after each layer LayerNorm, and the outputs of the first two layers are superimposed in the third layer in a jump way.

[0073] Feature alignment method uses maximum information coefficient rearrangement;

[0074] Using a sliding window of 3 days for dynamic MIC calculation, using R package minerva to realize correlation estimation;

[0075] Output features are sorted and reorganized according to mutual information gain to form behavior-metabolism-environment coupling features.

[0076] Specifically, the model training and risk prediction module:

[0077] Each type of data establishes a graph structure: the behavior graph is a daily behavior summary for each node;

[0078] The edge is the co-occurrence frequency (such as diet and exercise appearing at the same time);

[0079] Graph neural network configuration: graph convolution network (GCNConv) + GRU structure, 3 layers stacked; The hidden layer dimensions are 128-64-32 in turn.

[0080] Causal factor modeling structure: use NOTEARS algorithm to learn the structure causal graph; The minimum mutual information estimation is based on the MINE neural estimator: network hidden layer 128, sampling 128 pairs of positive and negative sample pairs.

[0081] Credibility evaluation and calibration module: Monte Carlo Dropout: Dropout rate is 0.2, and sample 100 times during inference;

[0082] Use Beta distribution to fit the confidence interval (95% confidence):

[0083] Count the number of correct predictions k for each sample, and the total number of samples n=100,

[0084] ;

[0085] Where, , : Two shape parameters of the Beta distribution;

[0086] : the same sample is performed the count of "correct" in the second random inactivation sampling;

[0087] : the total number of Monte Carlo sampling;

[0088] Feedback mechanism and closed-loop optimization: if the model error is greater than 0.2:

[0089] The corresponding feature is de-weighted (the weight is reduced to 0.6 times the original value);

[0090] In the next round of training, the system automatically removes unstable features and updates the feature construction layer configuration.

[0091] Visual interpretation and intervention suggestion module:

[0092] Use the Plotly library to build a risk heat map:

[0093] The X-axis is time (daily), and the Y-axis is the risk level (0-4);

[0094] The intervention suggestion structure adopts a nested dictionary format;

[0095] Technical effects and experimental data:

[0096] Sample size: N = 1000, followed up for 90 consecutive days; The prediction accuracy of the system is 92.3%; The accuracy of the existing control method is 79.4%.

[0097] Sampling interval stability experiment: after changing the sampling frequency, the prediction error fluctuation is controlled within ±0.05. Comparison of the calibration effect of the confidence interval: the coverage rate before calibration is 87%, and after calibration, it is improved to 95.2%.

[0098] Interpretability changes: after stripping the causal factors, the total average explanation of SHAP decreased from 0.72 to 0.41, indicating that the enhanced causal structure improves interpretability.

[0099] Specifically, the feedback closed-loop optimization module is used for performance backtracking and automatic adjustment of feature combinations after model training, and the core mechanism is as follows:

[0100] Error tracking mechanism: each model output calculates the error between its prediction result and the actual risk label;

[0101] The error threshold is set to 0.2, and if the prediction error of a feature combination continuously exceeds the threshold, it will automatically enter the full re-adjustment process.

[0102] Feature de-weighting mechanism: on the feature path with error exceeding the threshold, the SHAP value is evaluated for impact contribution;

[0103] Features with high error and large contribution are automatically reduced in weight by 20-50% in the next round of training.

[0104] Feature combination reconstruction: after triggering feature weight reduction, the system calls historical efficient combination templates to perform new feature cross combination;

[0105] And send the new combination to the data augmentation module (including SMOTE and time disturbance) to expand sample diversity.

[0106] Automatic iterative optimization: after each round of training, residual analysis is performed on the calibration confidence interval. If the overall bias tends to rise, structural fine-tuning (adjusting the graph structure edge weight) is started;

[0107] At the same time, based on the causal structure graph output by NOTEARS, the nodes with reverse causal paths are reduced in weight.

[0108] Feedback parameter configuration interface: users can set feedback adjustment strength (between 0 and 1), feature sensitivity threshold, optimization cooling period (such as adjusting once every 10 days), and other custom parameters.

[0109] Example effect: after 1000 users are continuously trained for 30 days, the feedback module improves the prediction accuracy by an average of 3.2% and reduces the overfitting risk by 10.7%.

[0110] The multi-dimensional feature construction module includes a three-layer nested structure. The first layer is the input reconstruction layer, which performs periodic smoothing processing on each type of time series data. A 3x1 convolution kernel and a bidirectional LSTM are used to reconstruct the behavior frequency data in time. The second layer is a multi-head nested attention structure, including gait-heart rate joint attention head, physiological-environmental fluctuation collaborative attention head, and social-feeding cross attention head. Each attention head is based on self-attention mechanism and window convolution to extract short-term mutual dependence and dynamically weight. The third layer is the feature alignment layer, which uses the maximum information coefficient rearrangement technology to reorder the multi-head output according to the causal coupling strength. Through a jump residual connection, it outputs the behavior-metabolism-environment three-coupling feature groups and outputs them to the next module according to the label window structure.

[0111] Specifically, the three-layer nested structure is detailed as follows:

[0112] Input reconstruction layer: 3x1 one-dimensional convolution is used for periodic filtering of behavior frequency data (such as step count);

[0113] Followed by bidirectional LSTM (hidden dimension 128) to reconstruct the behavior time series dynamic pattern.

[0114] Nested attention structure:

[0115] Three joint attention head configurations: gait-heart rate attention head; physiological-environmental attention head; social-feeding attention head;

[0116] Each attention head is based on self-attention mechanism and 1x3 convolution to model short-term local dependency;

[0117] Output dynamic weighted fusion features.

[0118] Feature alignment layer: apply maximum information coefficient (MIC) to rearrange multi-source sequences;

[0119] Use residual connection to jump fusion attention layer results;

[0120] Output three feature groups: [behavior, metabolism, environment], after dimension unification, send to label module.

[0121] The maximum information coefficient rearrangement technology is based on the nonlinear mapping relationship between multi-class feature groups. It uses a weighted dynamic time warping algorithm under a sliding window to align different feature sequences, calculates multi-order mutual information gain, extracts the strongest interconnected node path in the mapping sequence graph constructed based on spectral clustering, and finally reorganizes the coupled feature sequence length. The output features are completely consistent with the label generation window in time.

[0122] Specifically, the details of the maximum information coefficient rearrangement technology are as follows:

[0123] Nonlinear mapping is performed between each class of feature groups (such as heart rate-movement-diet);

[0124] The alignment method is sliding window dynamic time warping (DTW), and the window width is 3 days;

[0125] The multi-order mutual information gain calculation uses the formula:

[0126] ;

[0127] Where, : The maximum information coefficient of the paired features in the sliding window

[0128] : The comprehensive correlation index obtained by summing the maximum information coefficients of a number of paired in the window and then taking the maximum value;

[0129] : Sliding window set, window width 3 days;

[0130] : The number of paired features participating in accumulation / filtering in the window;

[0131] : Two types of feature sequences to be aligned / rearranged or their sub-segments;

[0132] ​In the spectral clustering map, the strongest interconnected path nodes are used to reorganize the coupling sequence, ensuring alignment with the label boundary within the sliding window.

[0133] The risk label dynamic generation module generates individual dynamic risk labels through the evolution clustering graph convolution structure under the dense time window. First, the multi-dimensional coupling feature sequence is processed by three-dimensional dimension reduction, principal component analysis, maximum variance preserving projection, and embedded graph kernel dimension reduction. Then, the evolution graph is constructed based on the graph structure density change rate, the TopK density peak value is marked for each time slice, the peak marked node aggregation is processed by the adaptive boundary shrinkage function for risk state layering, and finally the label matrix is output with point-by-point continuity on the time axis.

[0134] Specifically, the structure of the risk label dynamic generation module is as follows:

[0135] The three-dimensional combination of dimension reduction processing: principal component analysis (PCA) maintains 85% variance; maximum variance projection (LPP) is reduced to 256 dimensions; and embedded graph kernel dimension reduction (graph structure Laplacian eigenmap) has a final dimension of 128;

[0136] Evolution graph construction: construct a graph based on the density change of each time slice;

[0137] Perform TopK density peak marking on each segment, and the kernel function is:

[0138] ;

[0139] Wherein, : the kernel density estimation value of the sample point ;

[0140] : The th nearest neighbor sample of ;

[0141] : Euclidean distance;

[0142] : Gaussian kernel bandwidth;

[0143] Aggregation method: adaptive boundary shrinkage function automatically stratifies to ensure continuous time risk label matrix generation.

[0144] The adaptive boundary shrinkage function is composed of a dynamic boundary sliding factor and a local label density function;

[0145] The dynamic boundary sliding factor is composed of the product of the time step and the feature volatility rate, which dynamically compresses and expands the clustering boundary; the local label density function adjusts the threshold of each risk label based on the compactness of the sample distribution within the current time window, finally realizing the continuous expression of individual risk level in multiple periods.

[0146] Specifically, the adaptive boundary contraction function is composed of:

[0147] Dynamic boundary slip factor:

[0148]

[0149] wherein, : boundary slip amount (positive value expansion, negative value contraction);

[0150] : Proportion coefficient of boundary slip (0.1);

[0151] : Characteristic volatility at time (the sliding window standard deviation or coefficient of variation is given);

[0152] Local label density function:

[0153] ;

[0154] wherein, : The compactness index of the th class in the local window;

[0155] : The number of samples belonging to the th class in the local window;

[0156] : The feature vector of the th sample;

[0157] : The feature mean of the th class sample in the window;

[0158] : Euclidean norm;

[0159] : A small positive number is added to avoid zero denominator;

[0160] Control the boundary compression or expansion of each class label in the local window.

[0161] The model training and risk prediction module includes a multi-graph fusion graph neural network structure and a causal factor decomposition unit;

[0162] The multi-graph fusion graph neural network structure includes four types of graphs: behavior graph, physiological fluctuation graph, social influence graph and environmental fluctuation graph. Each type of graph adopts the joint input form of adjacency matrix and node embedding matrix. The graph neural network adopts a structured attention mechanism to update the node representation layer by layer, and outputs the global embedding in the form of aggregated center node weight mean;

[0163] The causal factor decomposition unit performs Laplacian Eigenmap filtering and minimum mutual information causal disentanglement processing on all graph embedding results, and finally outputs a causal weighted prediction score vector.

[0164] Specifically, the model training and risk prediction module cooperates with the "multi-graph fusion graph neural network structure" and the "causal factor decomposition unit" to work together to improve the personalized and explainability of the prediction:

[0165] Graph structure construction: Activity Graph: nodes are different types of behaviors (such as walking, sitting, and binge drinking), and edge weights are behavior transition probabilities;

[0166] Physiology Variation Graph: nodes are dynamic physiological indicators such as heart rate, blood sugar, and body weight, and edges are temporal correlations;

[0167] Social Influence Graph: nodes are user social interaction objects, and edge weights are interaction frequency x emotional similarity;

[0168] Environmental Change Graph: nodes are environmental factors (temperature and humidity, PM2.5, noise, etc.), and edges are spatial or temporal adjacency relationships.

[0169] Graph embedding input: each type of graph is input in the form of a two-dimensional adjacency matrix and an initial feature matrix to a graph neural network;

[0170] where R represents that the elements in the matrix come from the real number field, N represents the number of nodes in the graph, and d represents the feature dimension of each node.

[0171] All node dimensions are aligned to d=64 and standardized.

[0172] Structured attention mechanism: a structured attention module is introduced based on graph convolution, and the attention weight of node pair (i, j) is defined as:

[0173] ;

[0174] where fuses the semantic similarity of node pairs.

[0175] where, : the attention weight of node and node ;

[0176] : the attention weight of node , eigenvectors;

[0177] :node The set of neighboring nodes;

[0178] Attention scoring function;

[0179] Weight matrix;

[0180] Global embedding aggregation: the representation of all nodes output by each layer of the graph neural network. The graph-level embedding is obtained by taking the weighted average of the set of center nodes.

[0181] Causal factor decomposition unit: for all graph embeddings Laplacian feature filtering is performed to preserve low-frequency structural information;

[0182] Using the minimum mutual information objective function:

[0183] ;

[0184] Force decoupling between graphs and remove collinear, causeless dimensions;

[0185] The final output prediction vector is:

[0186] ;

[0187] in, : Predict the output vector;

[0188] : No. Global embedding of individual graphs (behavioral graphs, physiological graphs, social graphs, environmental graphs);

[0189] Causal allocation weights;

[0190] : Multilayer perceptron mapping function.

[0191] The model training and risk prediction module includes a location-nested attention factor and a group bias compensation factor. The location-nested attention factor constructs a location offset map based on timestamp differences and applies strong edges to update weights for nodes with similar times. The group bias compensation factor uses the mean deviation of inter-cluster behavioral characteristics as the adjustment amount to standardize the cross-group structural bias.

[0192] Specifically, based on graph neural networks, a position-nested attention factor and a group bias compensation factor are further introduced to optimize time sensitivity and group fairness:

[0193] Position-Aware Nested Attention:

[0194] Build Position Offset Graph: Calculate the timestamp difference for all time-series nodes ;

[0195] Update Adjacent Edges with Weighted Position Convolution:

[0196] ;

[0197] where, : edge weight of node and node in the original adjacent matrix;

[0198] : edge weight after time offset adjustment;

[0199] : timestamp difference of node , ;

[0200] : time decay coefficient (set to 0.01, representing the weight promotion of time adjacent nodes);

[0201] Group Bias Correction Factor:

[0202] Calculate the average behavior feature of each group for all users (e.g., by BMI, gender, age) ;

[0203] The compensation factor is defined as:

[0204] ;

[0205] where, : bias compensation value between groups;

[0206] : the average feature of the group to which node and node belongs, respectively;

[0207] All cross-group edges use for standardization to achieve structural alignment, : bias-compensated adjacent edge weight.

[0208] Actual effect:

[0209] By introducing this mechanism, the AUC value in cross-group prediction is improved by an average of 4.1%, and the problem of unbalanced training samples is significantly alleviated.

[0210] The minimum mutual information causal disentanglement process includes the following steps:

[0211] First, a Bayesian structure learning mechanism is used to construct a feature causal graph, and a Markov equivalence class clustering technique is used to filter out the structure path without cause; then an inverse probability weighting mechanism is introduced to resample the sample space, so that the latent confounding factors are distributed similarly after normalization in the sample, and finally a reparameterization technique is used to perform second-order supervised reinforcement training on the disentangled feature subset to generate high-resolution causal factors.

[0212] Specifically, the specific steps of the minimum mutual information causal disentanglement process are as follows:

[0213] Construct a causal graph: use a structure learning algorithm (such as NOTEARS or GES) to construct a directed graph between variables;

[0214] At the same time, merge the undirected edge structure through the Markov equivalence class technology (such as CPDAG), and eliminate redundant paths;

[0215] Only causal paths that meet the directed acyclic graph (DAG) structure are retained.

[0216] Inverse probability weighting (Inverse Propensity Weighting):

[0217] For the latent confounding factor C, the propensity score of the treatment variable T is calculated , which is used for weighting:

[0218] ;

[0219] Where, : the inverse probability weighting value of the sample ;

[0220] : the treatment variable (such as whether a certain behavior pattern occurs) of the sample ;

[0221] : the confounding factor (such as age, environment, etc.) of the sample ;

[0222] : the probability of the treatment variable occurring under the confounding factor condition;

[0223] The weight is used to resample the training sample, so that the confounding factors are uniformly distributed in each group, reducing the pseudo-causal bias.

[0224] Reparameterization causal subset training: extract a causal significant subset from the disentangled variable set;

[0225] For each sub-factor Construct independent supervision objectives and use second-order gradient optimization:

[0226] ;

[0227] where, : regularization loss function of causal subset;

[0228] : stripped causal factor;

[0229] : prediction loss function;

[0230] : second-order gradient operator;

[0231] Finally, high-resolution causal factors are used for model backbone fusion.

[0232] The credibility evaluation and calibration module is based on multiple forward prediction mechanisms to construct confidence intervals, uses Monte Carlo Dropout to sample 100 times of model prediction results, calculates the range, mean and skewness coefficient of each group of outputs, then fits the distribution shape of the calibrated prediction score through Beta distribution, and outputs the risk level adjustment factor with the upper confidence limit as the decision limit. Finally, a feedback signal is formed with the calibrated risk vector as input.

[0233] Specifically, the credibility evaluation and calibration module is designed based on multiple forward prediction mechanisms, and the specific steps are as follows:

[0234] Monte Carlo Dropout sampling:

[0235] Keep Dropout active during the prediction phase and perform 100 forward propagations;

[0236] Each output prediction score constitutes a distribution .

[0237] Confidence statistics extraction: calculate the range , mean μ, and skewness coefficient γ of each group of predictions;

[0238] If the range is >0.4 or the skewness is high, the prediction fluctuation is high and the confidence is low.

[0239] Distribution fitting and Beta calibration: fit to Beta distribution , and obtain the parameters through maximum likelihood estimation;

[0240] Use the calibrated upper confidence limit (such as 95%) as the decision limit:

[0241] Risk-S = (0.95; );

[0242] wherein Risk-S : calibrated risk score of a sample ;

[0243] : quantile of Beta distribution at confidence 0.95

[0244] : parameters of fitted Beta distribution

[0245] Risk level adjustment factor:

[0246] Set multiple levels of risk according to calibrated risk score:

[0247] Level I: <0.3, low risk

[0248] Level II: 0.3-0.6, medium risk

[0249] Level III: >0.6, high risk

[0250] Feedback adjustment factor to feature fusion and causality unit to improve prediction robustness.

[0251] The above shows and describes the basic principles, main features and advantages of the present application; those skilled in the art should understand that the present application is not limited to the above-mentioned embodiments, the above-mentioned embodiments and descriptions in the specification are only preferred examples of the present application, and are not intended to limit the present application, various changes and improvements can be made to the present application without departing from the spirit and scope of the present application, and these changes and improvements all fall within the scope of the claimed present application; the scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A big data based AI personal obesity risk prediction system, characterized by, The AI-based personal obesity risk prediction system based on big data comprises a data collection module, a multi-dimensional feature construction module, a risk label dynamic generation module, a model training and risk prediction module, a feedback closed-loop optimization module, a credibility evaluation and calibration module, a visual interpretation and intervention suggestion module; The data collection module adopts a multi-source heterogeneous data acquisition mechanism, synchronously collects gait change data, physiological time sequence parameters, dietary intake frequency logs, indoor and outdoor environment parameters, and social interaction logs through wearable devices, biochemical detection interfaces, and third-party health platform API interfaces, uniformly encodes the five types of standard structured time sequence, and transmits the encoded standard time sequence to the multi-dimensional feature construction module; The multi-dimensional feature construction module takes the encoded standard time sequence as input, constructs a multi-channel nested attention structure based on residual connection, fuses physiological change frequency curves and behavior trigger nodes through a graph embedding strategy, and outputs behavior-metabolism-environment coupled feature sets to the risk label dynamic generation module; the risk label dynamic generation module takes continuous features for more than three months as a window, generates a continuous risk state label set using a recursive clustering-adaptive boundary optimization strategy, and synchronously labels feature evolution paths to form a semi-supervised training data set; The model training and risk prediction module jointly models coupled features and dynamic labels based on a graph neural network-causal factor decomposition structure, extracts the interaction causal factors between long-term behavior paths and short-term physiological mutations through a structured regular enhanced graph convolution layer, and outputs an obesity risk prediction score; The credibility evaluation and calibration module performs posterior calibration on each model output based on a Monte Carlo Dropout and confidence interval adjustment mechanism, and transmits the calibrated results to the feedback closed-loop optimization module; the feedback closed-loop optimization module takes cross-stage calibration bias and model overfitting markers as core inputs, automatically selects mismatched features, and adjusts the feature construction strategy in reverse; The visual interpretation and intervention suggestion module constructs an interactive risk path graph based on a structural visual heat map and high-weight intervention factors, and generates an individual-oriented intervention suggestion vector.

2. The big data based AI personal obesity risk prediction system according to claim 1, characterized in that, The multi-dimensional feature construction module comprises a three-layer nested structure, the first layer is an input reconstruction layer, which performs periodic smoothing on each type of time sequence data, uses a 3x1 convolution kernel and a bidirectional LSTM to reconstruct the behavior frequency data in time, the second layer is a multi-head nested attention structure, which includes a gait-heart rate joint attention head, a physiological-environment fluctuation collaborative attention head, and a social-eating cross attention head, each attention head extracts the mutual dependence relationship in the short term based on the self-attention mechanism and window convolution and dynamically weights; the third layer is a feature alignment layer, which uses the maximum information coefficient rearrangement technology to reorder the multi-head output according to the causal coupling strength, and outputs the behavior-metabolism-environment three-type coupled feature groups through a skip residual connection and outputs them to the next module according to the label window structure.

3. The big data based AI personal obesity risk prediction system according to claim 2, characterized in that, The maximum information coefficient rearrangement technology is based on a nonlinear mapping relationship between multiple types of feature groups, uses a weighted dynamic time warping algorithm under a sliding window to align different feature sequences, calculates a multi-order mutual information gain, extracts a strongest interconnection node path in a mapping sequence graph constructed based on spectral clustering, and finally rearranges a coupled feature sequence length, so that the output feature is completely consistent with the label generation window in time.

4. The big data based AI personal obesity risk prediction system according to claim 1, wherein, The risk label dynamic generation module generates individual dynamic risk labels through an evolution clustering graph convolution structure under a dense time window. First, the multi-dimensional coupled feature sequence is subjected to three-dimensional dimension reduction processing, principal component analysis, maximum variance preserving projection, and embedded graph kernel dimension reduction. Then, an evolution graph is constructed based on the graph structure density change rate, and the TopK density peak value of each time slice is marked. The peak value marked node aggregation is subjected to risk state layering through an adaptive boundary shrinkage function, and finally a label matrix with point-by-point continuity on the time axis is output.

5. The big data based AI personal obesity risk prediction system according to claim 4, wherein, The adaptive boundary shrinkage function is composed of a dynamic boundary sliding factor and a local label density function. The dynamic boundary sliding factor is composed of the product of the time step and the feature volatility rate, and dynamically compresses and expands the clustering boundary. The local label density function adjusts the boundary threshold of each risk label based on the sample distribution tightness in the current time window, and finally realizes the continuous expression of the individual risk level in multiple time periods.

6. The big data based AI personal obesity risk prediction system according to claim 1, wherein, The model training and risk prediction module includes a multi-graph fusion graph neural network structure and a causal factor decomposition unit. The multi-graph fusion graph neural network structure includes four types of graphs, namely, a behavior graph, a physiological fluctuation graph, a social influence graph, and an environmental fluctuation graph. Each type of graph adopts an adjacent matrix and a node embedding matrix joint input form. The graph neural network adopts a structured attention mechanism to update the node representation layer by layer, and outputs the global embedding in an aggregated center node weight mean manner. The causal factor decomposition unit performs Laplace feature filtering and minimum mutual information causal stripping processing on all graph embedding results, and finally outputs a causal weighted prediction score vector.

7. The big data based AI personal obesity risk prediction system according to claim 6, wherein, The model training and risk prediction module includes a location nested attention factor and a group bias compensation factor. The location nested attention factor constructs a location offset graph based on the timestamp difference, and applies a strong edge update weight to nodes close in time. The group bias compensation factor uses the mean deviation of behavior characteristics between clusters as an adjustment amount to standardize the structural bias between groups.

8. The big data based AI personal obesity risk prediction system according to claim 7, characterized in that, The minimum mutual information causal stripping processing includes the following steps: First, a feature causal graph is constructed based on a Bayesian structure learning mechanism, and a Markov equivalence class clustering technique is used to exclude non-causal structure paths. Then, an inverse probability weighting mechanism is introduced to resample the sample space, so that the latent confounding factors are normalized and distributed similarly in the sample. Finally, a reparameterization technique is used to perform second-order supervised reinforcement training on the stripped feature subset, generating high-resolution causal factors. 9.The big data based AI personal obesity risk prediction system of claim 1, wherein, The credibility evaluation and calibration module is based on a multiple forward prediction mechanism to construct a confidence interval, uses Monte Carlo Dropout to sample 100 model prediction results, calculates the range, mean and skewness coefficient of each group of outputs, fits the distribution shape of the calibrated prediction score through a beta distribution, and outputs a risk level adjustment factor with a confidence upper limit as the decision limit, and finally forms a feedback signal with a calibrated risk vector as input. 10.A personal obesity risk prediction method based on big data AI, characterized in that, The AI-based personal obesity risk prediction method based on big data uses the AI-based personal obesity risk prediction system based on big data in any of claims 1-9.

Images