Personal obesity risk prediction system and method based on AI of big data

Abstract

The invention discloses a personal obesity risk prediction system and method based on AI of big data, and belongs to the field of medical health, and the system 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 credibility evaluation and calibration module and the like. The system collects multi-source heterogeneous data through wearable equipment, a biochemical interface and a health platform API (Application Program Interface), uniformly encodes the multi-source heterogeneous data into a standard time sequence and then constructs behavior-metabolism-environment coupling characteristics. And performing joint modeling on the dynamic features and the labels by adopting a graph neural network in combination with causal factorization, and outputting an individual obesity risk prediction result. The result credibility is improved through Monte Carlo Dropout and a confidence interval calibration mechanism, and calibration information is fed back to a feature construction link to optimize a modeling strategy. Finally, the key risk factors are presented in the form of a visual thermodynamic diagram and a causal path diagram, and an individualized intervention suggestion vector is generated. The method has the beneficial effects of improving prediction accuracy and enhancing individual intervention pertinence.

Description

Technical Field

[0001] The present invention relates to the field of medical health, and more specifically, to a personal obesity risk prediction system and method based on big data and AI. Background Art

[0002] In recent years, with the widespread use of wearable devices, health apps, and vital sign monitoring systems, the collection of personal health data has become increasingly rich, and big data-driven health risk prediction 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. Although these methods can achieve basic risk stratification, they have significant limitations in handling high-dimensional heterogeneous data, multi-factor interactions, and individual differences. In recent years, the introduction of graph neural networks and causal inference techniques has driven the development of health prediction models to a deeper level, enabling initial attempts at complex association modeling and causal factor identification.

[0003] However, existing systems generally suffer from the following shortcomings: First, their modeling structure is sparse, making it difficult to integrate graph-structured information from multi-source data; second, they ignore the effects of time series offset and population heterogeneity, leading to biased predictions; third, they lack a causal factor screening mechanism, making predictions difficult to interpret; and fourth, the credibility of risk predictions is not clearly calibrated, making it impossible to form a stable feedback mechanism. Therefore, an AI system that integrates multi-graph modeling, causal separation analysis, and risk calibration mechanisms is urgently needed to comprehensively improve the accuracy, interpretability, and adaptability of obesity risk prediction. Summary of the Invention

[0004] The purpose of the present invention is to provide a personal obesity risk prediction system and method based on big data AI to solve the problems raised in the above background technology: first, the modeling structure is single, and it is difficult to integrate graph structure information in multi-source data; second, the influence of time series offset and group heterogeneity is ignored, resulting in biased prediction results; third, there is a lack of causal factor screening mechanism, and the prediction results are difficult to interpret; fourth, the credibility of risk prediction is not clearly calibrated, and a stable feedback mechanism cannot be formed.

[0005] Technical Solution: A big data-based AI-based personal obesity risk prediction system 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 assessment and calibration module, and a visual explanation and intervention recommendation module. The data collection module adopts a multi-source heterogeneous data acquisition mechanism, and synchronously collects gait change data, physiological time series parameters, dietary intake frequency logs, indoor and outdoor environmental parameters and social interaction logs through wearable devices, biochemical detection interfaces and third-party health platform API interfaces, and uniformly encodes them into five categories of standard structured time series sequences and then passes them to the multidimensional feature construction module; the multidimensional feature construction module takes the encoded standard time series sequence as input, constructs a multi-channel nested attention structure based on residual connection, fuses the physiological change frequency curve with the behavior trigger node through a graph nesting strategy, and outputs the behavior-metabolism-environment coupling feature set to the risk label dynamic generation module; the risk label dynamic generation module uses continuous feature input of three months or more as a window, generates a continuous risk state label set with a recursive clustering-adaptive boundary optimization strategy, and synchronously annotates the feature evolution path to form a semi-supervised training data set; The model training and risk prediction module jointly models coupling features and dynamic labels based on a graph neural network-causal factor decomposition structure, extracts interactive causal factors between long-term behavioral paths and short-term physiological mutations through a structured regularized enhanced graph convolutional layer, and outputs an obesity risk prediction score. The credibility assessment and calibration module performs a posteriori calibration on each model output based on Monte Carlo Dropout and confidence interval adjustment mechanisms, and passes the calibration results to the feedback closed-loop optimization module; the feedback closed-loop optimization module uses cross-stage calibration deviations and model overfitting flags as core inputs, automatically selects mismatched features, and reversely adjusts the feature construction strategy; The visualization explanation and intervention recommendation module constructs an interactive risk path map based on the structural visualization heat map and high-weight intervention factors, which is used to generate individual-oriented intervention recommendation vectors.

[0006] Preferably, the multidimensional feature construction module includes a three-layer nested structure, the first layer is the input reconstruction layer, which performs periodic smoothing on each type of time series data, and uses a 3×1 convolution kernel plus a bidirectional LSTM to perform temporal reconstruction on the behavioral frequency data; the second layer is a multi-head nested attention structure, including a gait-heart rate joint attention head, a physiological-environmental fluctuation collaborative attention head, and a social-feeding cross attention head. Each attention head extracts short-term interdependencies based on the self-attention mechanism and window convolution and dynamically weights them; the third layer is a feature alignment layer, which uses the maximum information coefficient rearrangement technology to reorder the multi-head outputs according to the causal coupling strength, and outputs the three types of coupled feature groups of behavior-metabolism-environment through jump residual connections and outputs them to the next module according to the label window structure.

[0007] Preferably, the maximum information coefficient rearrangement technology is based on the nonlinear mapping relationship between multiple feature groups, uses a weighted dynamic time warping algorithm under a sliding window to align different feature sequences, and calculates multi-order mutual information gain. The strongest interconnected node path is extracted from the mapping sequence graph constructed based on spectral clustering, and finally the length of the coupled feature sequence is reorganized so that the output feature is completely consistent with the label generation window in time.

[0008] Preferably, the risk label dynamic generation module generates individual dynamic risk labels through the evolution clustering graph convolution structure under a dense time window. First, the multi-dimensional coupling feature sequence is subjected to three-fold dimensionality reduction processing, principal component analysis, maximum variance preserving projection, and embedded graph kernel dimensionality reduction. Then, an evolution graph is constructed based on the density change rate of the graph structure, and TopK density peaks are marked for each time slice. The node aggregation of the peak mark is stratified by risk status through an adaptive boundary shrinkage function, and the final output label matrix has point-by-point continuity on the time axis.

[0009] Preferably, the adaptive boundary shrinkage function is composed of a dynamic boundary sliding factor and a local label density function; The dynamic boundary slip factor is composed of the product of the time step and the characteristic volatility, dynamically compressing and expanding the cluster boundary; the local label density function adjusts the demarcation threshold of each type of risk label based on the compactness of the sample distribution in the current time window, ultimately achieving continuous expression of the individual's risk level in multiple time periods.

[0010] Preferably, the model training and risk prediction module comprises 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: behavioral graphs, physiological fluctuation graphs, social influence graphs, and environmental fluctuation graphs. Each type of graph is input in the form of a joint adjacency matrix and a node embedding matrix. The graph neural network uses a structured attention mechanism to hierarchically update node representations and outputs a global embedding by aggregating the weighted average of the central nodes. 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.

[0011] 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 and applies edge update weights to nodes that are close in time. The group bias compensation factor uses the mean deviation of behavioral characteristics between clusters as an adjustment amount to standardize the structural deviation across groups.

[0012] Preferably, the minimum mutual information causal separation process includes the following steps: First, a feature causal graph is constructed using the Bayesian structural learning mechanism, and the Markov equivalence clustering technology is used to screen out the uncaused structural paths; then the inverse probability weighting mechanism is introduced to resample the sample space so that the potential confounding factors are normalized and distributed similarly in the sample; finally, the reparameterization technology is used to perform second-order supervised reinforcement training on the stripped feature subset to generate high-resolution causal factors.

[0013] Preferably, the credibility assessment and calibration module constructs a confidence interval based on a multiple forward prediction mechanism, uses Monte Carlo Dropout to sample the model prediction results 100 times and calculates the range, mean and skewness coefficient of each group of outputs, then calibrates the distribution shape of the prediction score through Beta distribution fitting, and outputs the risk level adjustment factor with the confidence upper bound as the decision limit, and finally forms a feedback signal with the calibration risk vector as input.

[0014] Preferably, the personal obesity risk prediction method based on big data AI includes the personal obesity risk prediction system based on big data AI described in any one of claims 1-9.

[0015] Compared with the prior art, the advantages of the present invention are: (1) The introduction of four types of heterogeneous graphs, namely behavioral graph, physiological fluctuation graph, social influence graph and environmental fluctuation graph, enhances the ability to express the interactive influence of complex factors and has stronger structural characterization capabilities compared to traditional single feature models.

[0016] (2) A hierarchical structured attention mechanism is used to optimize the information transmission and aggregation process of graph neural networks, improving the ability to identify key nodes and edges, thereby improving the accuracy and interpretability of risk prediction.

[0017] (3) Laplace feature filtering and minimum mutual information stripping operations are introduced to extract more causally relevant feature factors from mixed information, which improves the rigor of causal modeling and the generalization ability of the model compared with the existing empirical weight method.

[0018] (4) To address the problem of node timing offset in time series data, a position offset graph is constructed and edge updates are performed, which improves the modeling accuracy of dynamic correlations between temporally approximate events.

[0019] (5) By normalizing the mean deviation of cross-group behavior to deal with structural deviation, the model prediction bias caused by population heterogeneity is alleviated, and the model's adaptability to people of different body shapes, groups, and lifestyles is improved.

[0020] (6) The feature causal graph is optimized and trained using inverse probability weighting and reparameterization techniques, so that the output causal factors have stronger discriminative ability and fine-grained expression.

[0021] (7) Monte Carlo Dropout combined with Beta distribution fitting is used to construct the prediction confidence interval, and the upper confidence bound is set as the basis for risk classification, which solves the problem that the credibility of the prediction results cannot be quantified in the existing system.

[0022] (8) Using the calibrated risk vector as feedback input, a closed-loop dynamic correction mechanism is formed to enhance the system’s continuous learning and adaptive capabilities in real-world usage scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a schematic diagram of the overall system of an AI-based personal obesity risk prediction system based on big data of the present invention; DETAILED DESCRIPTION

[0024] For examples, see Figure 1 ,A personal obesity risk prediction system based on big data and AI 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 assessment and calibration module, and a visual explanation and intervention suggestion module; The data collection module adopts a multi-source heterogeneous data acquisition mechanism, and synchronously collects gait change data, physiological time series parameters, dietary intake frequency logs, indoor and outdoor environmental parameters, and social interaction logs through wearable devices, biochemical detection interfaces, and third-party health platform API interfaces, and uniformly encodes them into five categories of standard structured time series sequences before passing them to the multidimensional feature construction module; the multidimensional feature construction module takes the encoded standard time series sequences as input, constructs a multi-channel nested attention structure based on residual connection, fuses the physiological change frequency curve with the behavioral trigger node through a graph nesting strategy, and outputs the behavior-metabolism-environment coupling feature set to pass to the risk label dynamic generation module; the risk label dynamic generation module uses continuous feature input of three months or more as a window, and uses a recursive clustering-adaptive boundary optimization strategy to generate a continuous risk state label set, and simultaneously annotates the feature evolution path to form a semi-supervised training data set; The model training and risk prediction module jointly models coupling features and dynamic labels based on a graph neural network-causal factor decomposition structure. It extracts interactive causal factors between long-term behavioral pathways and short-term physiological mutations through a structured regularized enhanced graph convolutional layer, and outputs an obesity risk prediction score. The credibility assessment and calibration module performs a posteriori calibration of each model output based on Monte Carlo Dropout and a confidence interval adjustment mechanism. The calibration results are then passed to the feedback closed-loop optimization module. The feedback closed-loop optimization module uses cross-stage calibration deviations and model overfitting flags as core inputs to automatically select mismatched features and adjust the feature construction strategy in reverse. The visualization explanation and intervention recommendation module constructs an interactive risk path map based on structural visualization heat maps and high-weight intervention factors, which is used to generate individual-oriented intervention recommendation vectors.

[0025] Specifically, the data collection module adopts the following configuration: The wearable device is Fitbit Charge 5 (a fitness tracker under Google), which collects data every 5 minutes; The collected physiological parameters include: heart rate (HeartRate), skin temperature (SkinTemperature), blood oxygen saturation (SpO2). All signals are synchronously sampled and unified into a 60Hz frequency time series signal; Behavior logs are obtained through the API interface.

[0026] Data preprocessing includes: Outlier elimination: Detecting outliers based on the 3σ principle; Linear interpolation filling: linear fitting is performed using the 5-minute interval values ​​before and after. Sliding normalization uses the following formula: ; in, :time The normalized value of :time The original observation value of :by The center of the window is The sliding window mean of ; : Standard deviation of the corresponding sliding window; : Sliding window radius ( ) :Prevent extremely small positive numbers with denominators equal to zero ( ).

[0027] The multi-dimensional feature construction module takes the encoded standard sequence as input and has the following structure: The nested attention structure is three layers, where: The number of multiple heads is 8, and each attention head structure is: ; in, : query matrix (obtained from the input through Dense(64)); : key matrix (obtained from the input through Dense(64)); : value matrix (obtained from the input through Dense(64)); : The dimension of the key vector (when it is consistent with your Dense(64) setting ); : An exponential normalization function performed row-wise to generate attention weights; : Similarity matrix between query and key (dot product); : Scaling factor, used to stabilize gradient and probability distribution; Positional encoding uses the standard sine-cosine function: ; ; in, 、 :Location Positional encoding values ​​in even / odd dimensions; : discrete position index in the sequence; : Dimension "pair" index starting from 0 (each pair contains both odd and even dimensions); : The total embedding dimension of the model (the dimension of the vector on which the positional encoding acts); constant : wavelength scale base of position encoding, used to cover multi-scale frequencies; Feature residual connection: Each layer of LayerNorm is connected to the input of the next layer, and the outputs of the first two layers are superimposed in a jump manner in the third layer.

[0028] The feature alignment method uses the maximum information coefficient rearrangement; Dynamic MIC calculations were performed using a sliding window of 3 days, and correlation estimation was achieved using the R package minerva; The output features are sorted and reorganized according to the mutual information gain to form behavior-metabolism-environment coupling features.

[0029] Specifically, model training and risk prediction modules: A graph structure is established for each type of data: each node in the behavior graph is a daily behavior summary; Edges are co-occurrence frequencies (e.g., diet and exercise appear together); Graph neural network configuration: Graph Convolutional Network (GCNConv) + GRU structure, 3 layers stacked; hidden layer dimensions are 128-64-32 respectively.

[0030] Causal factor modeling structure: The NOTEARS algorithm is used to learn the structural causal graph; the minimum mutual information estimation is based on the MINE neural estimator: the network has 128 hidden layers and samples 128 pairs of positive and negative samples.

[0031] Credibility evaluation and calibration module: Monte Carlo Dropout: Dropout rate is 0.2, and sampling is 100 times during inference; Fit a confidence interval (95% confidence) using the Beta distribution: Count the number of correct predictions k for each sample, the total number of samples n=100, ; in, 、 : Two shape parameters of Beta distribution; :For the same sample The number of “correct” counts in random dropout sampling; : The total number of Monte Carlo sampling; Feedback mechanism and closed-loop optimization: If the model error is greater than 0.2: The corresponding features are down-weighted (the weight is reduced to the original value × 0.6); In the next round of training, the system automatically removes unstable features and updates the feature construction layer configuration.

[0032] Visual explanation and intervention suggestion module: Use the Plotly library to build a risk heat map: The X-axis is time (day level), and the Y-axis is risk level (0–4); The intervention suggestion structure adopts a nested dictionary format; Technical effects and experimental data: Sample size: N=1000, 90 consecutive days of tracking; prediction accuracy of this system: 92.3%; accuracy of existing control methods: 79.4%.

[0033] Sampling interval stability experiment: After changing the sampling frequency, the prediction error fluctuation was controlled within the range of ±0.05. Comparison of the credibility interval calibration effect: The coverage rate was 87% before calibration and increased to 95.2% after calibration.

[0034] Changes in interpretability: After stripping away the causal factors, the overall average interpretability of SHAP decreased from 0.72 to 0.41, indicating that the causal structure enhances interpretability.

[0035] Specifically, the feedback closed-loop optimization module is used for performance backtracking and automatic feature combination adjustment after model training. The core mechanism is as follows: Error tracking mechanism: Each time the model outputs an error calculation between its predicted result and the actual risk label; The error threshold is set to 0.2. If the prediction error under a certain feature combination exceeds this threshold continuously, the full readjustment process will be automatically entered.

[0036] Feature downgrading mechanism: On feature paths where the error exceeds the threshold, the impact contribution is evaluated in combination with the SHAP value; For features with high errors and large contributions, the weights are automatically reduced by 20%-50% in the next round of training.

[0037] Feature combination reconstruction: After triggering feature demotion, the system calls historical efficient combination templates to perform new feature cross-combinations; The new combination is then fed into a data augmentation module (including SMOTE and temporal perturbations) to expand sample diversity.

[0038] Automatic iterative optimization: After each round of training, residual analysis is performed on the calibration confidence interval. If the overall deviation tends to increase, structural fine-tuning (adjusting the edge weights of the graph structure) is initiated. At the same time, based on the causal structure diagram output by NOTEARS, the nodes with reverse causal paths are downgraded.

[0039] Feedback parameter configuration interface: Users can set custom parameters such as feedback adjustment strength (between 0 and 1), feature sensitivity threshold, and optimized cooling cycle (such as readjustment every 10 days).

[0040] Example results: After 30 days of continuous training on 1,000 users, the feedback module improved prediction accuracy by an average of 3.2% and reduced overfitting risk by 10.7%.

[0041] The multidimensional feature construction module includes a three-layer nested structure. The first layer is the input reconstruction layer, which performs periodic smoothing on each type of time series data and uses a 3×1 convolution kernel plus a bidirectional LSTM to reconstruct the behavioral frequency data in time. The second layer is a multi-head nested attention structure, including a gait-heart rate joint attention head, a physiological-environmental fluctuation collaborative attention head, and a social-feeding cross attention head. Each attention head extracts short-term interdependencies based on the self-attention mechanism and window convolution and dynamically weights them; the third layer is the feature alignment layer, which uses the maximum information coefficient rearrangement technology to reorder the multi-head outputs according to the causal coupling strength, and outputs the three types of coupled feature groups of behavior-metabolism-environment through jump residual connections and outputs them to the next module according to the label window structure.

[0042] Specifically, the three-layer nested structure is as follows: Input reconstruction layer: Use 3×1 one-dimensional convolution to perform periodic filtering on behavioral frequency data (such as step number); It is followed by a bidirectional LSTM (hidden dimension 128) to reconstruct the temporal dynamic pattern of the behavior.

[0043] Nested attention structure: Three joint attention head configurations: gait-heart rate attention head; physiological-environmental attention head; social-feeding attention head; Each attention head implements short-term local dependency modeling based on self-attention mechanism and 1×3 convolution; Output dynamic weighted fusion features.

[0044] Feature alignment layer: Apply the maximum information coefficient (MIC) to realign multiple source sequences; Use residual connections to skip-fuse attention layer results; Output three feature groups: [behavior, metabolism, environment], and send them to the labeling module after dimension unification.

[0045] The maximum information coefficient rearrangement technology is based on the nonlinear mapping relationship between multiple feature groups. It uses the weighted dynamic time warping algorithm under the sliding window to align different feature sequences and calculate the multi-order mutual information gain. The strongest interconnected node path is extracted from the mapping sequence graph constructed based on spectral clustering, and finally the length of the coupled feature sequence is reorganized so that the output feature is completely consistent with the label generation window in time.

[0046] Specifically, the maximum information coefficient rearrangement technology details are as follows: Perform nonlinear mapping between each feature group (e.g., heart rate-exercise-diet); The alignment method is sliding window dynamic time warping (DTW) with a window width of 3 days; The multi-order mutual information gain is calculated using the formula: ; in, : In the sliding window The maximum information coefficient of the group pairing characteristics; :For several pairs in the window The comprehensive correlation index is obtained by summing and taking the maximum value; : Sliding window set, window width 3 days; : The number of pairs participating in accumulation / screening in the window; : Two types of characteristic sequences or their sub-fragments to be aligned / rearranged; In the spectral clustering map, the strongest interconnected path nodes are used to reorganize the coupled sequences to ensure alignment with the label boundaries within the sliding window.

[0047] The risk label dynamic generation module generates individual dynamic risk labels through the convolutional structure of the evolutionary clustering graph under a dense time window. First, the multi-dimensional coupling feature sequence is subjected to a three-fold dimensionality reduction process: principal component analysis, maximum variance preserving projection, and embedded graph kernel dimensionality reduction. Then, an evolution graph is constructed based on the density change rate of the graph structure. The TopK density peak is marked for each time slice. The nodes of the peak mark are aggregated and the risk status is stratified through an adaptive boundary shrinkage function. The final output label matrix has point-by-point continuity on the time axis.

[0048] Specifically, the structure of the risk label dynamic generation module is as follows: A triple dimensionality reduction process: principal component analysis (PCA) to maintain 85% variance; maximum variance projection (LPP) to reduce to 256 dimensions; and embedding kernel dimensionality reduction (graph structure Laplacian feature embedding) to a final dimension of 128. Evolution graph construction: Build a graph based on the density changes of each time slice; Perform TopK density peak marking on each fragment, kernel function: ; in, : Sample points The kernel density estimate of ; : No. Nearest neighbor samples; : Euclidean distance; : Gaussian kernel bandwidth; Aggregation method: Adaptive boundary shrinkage function automatically stratifies to ensure continuous time risk label matrix generation.

[0049] The adaptive boundary shrinkage function is composed of a dynamic boundary sliding factor and a local label density function; The dynamic boundary slip factor is composed of the product of the time step and the characteristic volatility, which dynamically compresses and expands the clustering boundary; the local label density function adjusts the demarcation threshold of each type of risk label based on the compactness of the sample distribution in the current time window, and ultimately achieves the continuous expression of the individual's risk level in multiple time periods.

[0050] Specifically, the adaptive boundary shrinkage function is composed of: Dynamic boundary slip factor:

[0051] in, : Boundary slip (positive value expansion, negative value contraction); : Proportional coefficient of boundary slip (0.1); : At the moment the characteristic volatility of (given by the sliding window standard deviation or coefficient of variation); Local label density function: ; in, : No. The compactness index of the class within the local window; :The local window belongs to The number of samples in the class; : No. The feature vector of each sample; : No. The feature mean of the class sample in the window; : Euclidean norm; : A very small positive number added to avoid the denominator being zero; Controls the compression or expansion of the boundaries of each type of label in the local window.

[0052] The model training and risk prediction module includes a multi-graph fusion neural network structure and a causal factor decomposition unit; The multi-graph fusion graph neural network structure includes four types of graphs: behavioral graphs, physiological fluctuation graphs, social influence graphs, and environmental fluctuation graphs. Each type of graph uses a joint input form of an adjacency matrix and a node embedding matrix. The graph neural network uses a structured attention mechanism to hierarchically update node representations and outputs a global embedding by aggregating the weighted average of the central nodes. The causal factor decomposition unit performs Laplace feature filtering and minimum mutual information causal stripping on all graph embedding results, and finally outputs a causal weighted prediction score vector.

[0053] Specifically, the model training and risk prediction modules use a "multi-graph fusion neural network structure" and a "causal factor decomposition unit" to work together to improve the personalization and interpretability of predictions: Graph structure construction: Activity Graph: Nodes represent different types of behaviors (such as walking, sitting, and binge drinking), and edge weights represent the probability of behavior transitions; PhysiologyVariationGraph: nodes are dynamic physiological indicators such as heart rate, blood sugar, and weight, and edges are temporal correlations; Social Influence Graph: Nodes are users’ social interaction objects, and edge weights are interaction frequency × sentiment similarity. Environmental Change Graph: Nodes represent environmental factors (temperature, humidity, PM2.5, noise, etc.), and edges represent spatial or temporal adjacency.

[0054] Graph embedding input: Each type of graph is represented by a two-dimensional adjacency matrix and the initial feature matrix Formal input graph neural network; Among them, R indicates that the elements in the matrix come from the real number domain, N represents the number of nodes in the graph, and d represents the feature dimension of each node; All node dimensions are aligned to d=64 and normalized.

[0055] Structured attention mechanism: Based on graph convolution, a structured attention module is introduced, and the attention weight of node pair (i, j) is defined as: ; in , fusing the semantic similarity of node pairs.

[0056] in, :node With node The attention weight of :node 、 The eigenvector of :node The set of neighbor nodes of : attention score function; : weight matrix; Global embedding aggregation: all node representations output by each layer of the graph neural network Take the weighted average of the set of central nodes to obtain graph-level embedding Causal Factorization Unit: Embedding all graphs Laplacian feature filtering is performed to preserve low-frequency structural information; Use the minimum mutual information objective function: ; Force graphs to be decoupled and strip away collinear uncaused dimensions; The final output prediction vector is: ; in, : predicted output vector; : No. Global embedding of graphs (behavioral graph, physiological graph, social graph, environmental graph); : causal distribution weight; : Multilayer perceptron mapping function.

[0057] 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 timestamp differences and imposes enhanced edge update weights on nodes with close time. The group bias compensation factor uses the mean deviation of behavioral characteristics between clusters as an adjustment amount to standardize the structural deviation across groups.

[0058] Specifically, based on the graph neural network, we further introduce position-nested attention factors and group bias compensation factors to optimize time sensitivity and group fairness: Position-AwareNestedAttention: Constructing a position offset graph: Calculating the timestamp difference of all time series nodes ; Update adjacent edges using weighted positional convolution: ; in, : Nodes in the original adjacency matrix With node The edge right; : Edge weight after time offset adjustment; :node 、 The timestamp difference; : Time decay coefficient (set to 0.01, indicating a weighted increase in temporally adjacent nodes); Group Bias Correction Factor: Cluster all users (e.g., by BMI, gender, age) and calculate the mean behavioral characteristics of each group ; The compensation factor is defined as: ; in, : Bias compensation value between groups; :respectively nodes ,node The mean of the characteristics of the group to which it belongs; All cross-group edge use Do standardization to achieve structural alignment, : The adjacent edge weight after bias compensation.

[0059] Actual effect: By introducing this mechanism, the AUC value in cross-group prediction is improved by an average of 4.1%, and the problem of imbalanced training samples is significantly alleviated.

[0060] The minimum mutual information causal separation process includes the following steps: First, a feature causal graph is constructed using the Bayesian structural learning mechanism, and the Markov equivalence clustering technology is used to screen out the uncaused structural paths; then the inverse probability weighting mechanism is introduced to resample the sample space so that the potential confounding factors are normalized and distributed similarly in the sample; finally, the reparameterization technology is used to perform second-order supervised reinforcement training on the stripped feature subset to generate high-resolution causal factors.

[0061] Specifically, the specific steps of minimum mutual information causal separation processing are as follows: Constructing causal graphs: Use structural learning algorithms (such as NOTERS or GES) to construct directed graphs between variables; At the same time, Markov equivalence class technology (such as CPDAG) is used to merge undirected edge structures and eliminate redundant paths; Only causal paths that satisfy the directed acyclic graph (DAG) structure are retained.

[0062] Inverse Propensity Weighting: For potential confounders C, calculate the propensity score of treatment variable T , used for weighting: ; in, :sample The inverse probability weighted value of ; :sample treatment variables (e.g., whether a certain behavior pattern occurs); :sample Confounding factors (such as age, environment, etc.); : The probability of the treatment variable appearing under the condition of confounding factors; The weights are used to resample the training samples to make the confounding factors uniformly distributed across groups and reduce pseudo-causal bias.

[0063] Reparameterized causal subset training to extract causally significant subsets from the stripped variable set; For each sub-factor Construct an independent supervision objective and use second-order gradient optimization: ; in, : Regularized loss function for causal subsets; : The causal factors removed; : prediction loss function; : second-order gradient operator; Finally, the high-resolution causal factor is used for model backbone fusion.

[0064] The credibility assessment and calibration module constructs confidence intervals based on a multiple forward prediction mechanism, uses Monte Carlo Dropout to sample the model prediction results 100 times and calculates the range, mean and skewness coefficient of each set of outputs. It then calibrates the distribution shape of the prediction scores through Beta distribution fitting, and outputs the risk level adjustment factor with the confidence upper bound as the decision limit, ultimately forming a feedback signal with the calibrated risk vector as input.

[0065] Specifically, the credibility assessment and calibration module is designed based on a multiple forward prediction mechanism. The specific steps are as follows: Monte Carlo Dropout Sampling: Keep Dropout activated during the prediction phase and perform 100 forward propagations; Output prediction score each time , constitute the distribution .

[0066] Confidence Statistics Extraction: Calculate the range of each set of predictions , mean μ, skewness coefficient γ; If the range > 0.4 or skewness , it is judged that the forecast volatility is high and the confidence is low.

[0067] Distribution Fitting and Beta Calibration: Fitting to Beta distribution , and obtain the parameters through maximum likelihood estimation; Use a calibrated upper confidence bound (e.g., 95%) as the decision limit: Risk-S = (0.95; ); Among them, Risk-S :sample Calibrated risk scores; : The quantile of Beta distribution at the confidence level of 0.95; : Fitting parameters of Beta distribution; Risk Level Adjustment Factor: Set multiple levels of risk based on calibrated risk scores: Level I: <0.3, low risk Level II: 0.3–0.6, medium risk Level III: >0.6, high risk The adjustment factors are fed back to the feature fusion and causal units to improve the prediction robustness.

[0068] The above shows and describes the basic principles, main features and advantages of the present invention; those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements all fall within the scope of the present invention to be protected; the scope of protection claimed in the present invention is defined by the attached claims and their equivalents.

Claims

1. A personal obesity risk prediction system based on big data AI, characterized by: The AI-based personal obesity risk prediction system based on big data includes a data collection module, a multidimensional feature construction module, a risk label dynamic generation module, a model training and risk prediction module, a feedback closed-loop optimization module, a credibility assessment and calibration module, and a visual explanation and intervention suggestion module. The data collection module adopts a multi-source heterogeneous data acquisition mechanism to synchronously collect gait change data, physiological time series parameters, dietary intake frequency logs, indoor and outdoor environmental parameters, and social interaction logs through wearable devices, biochemical detection interfaces, and third-party health platform API interfaces. The data is uniformly encoded into five types of standard structured time series sequences and then transmitted to the multidimensional feature construction module; The multidimensional feature construction module takes the encoded standard time series as input, constructs a multi-channel nested attention structure based on residual connection, fuses the physiological change frequency curve with the behavioral trigger node through a graph nesting strategy, and outputs a behavior-metabolism-environment coupling feature set to the risk label dynamic generation module; the risk label dynamic generation module uses continuous feature input of three months or more as a window, generates a continuous risk state label set using a recursive clustering-adaptive boundary optimization strategy, and simultaneously annotates the feature evolution path to form a semi-supervised training dataset; The model training and risk prediction module jointly models coupling features and dynamic labels based on a graph neural network-causal factor decomposition structure, extracts interactive causal factors between long-term behavioral paths and short-term physiological mutations through a structured regularized enhanced graph convolutional layer, and outputs an obesity risk prediction score. The credibility assessment and calibration module performs a posteriori calibration on each model output based on Monte Carlo Dropout and confidence interval adjustment mechanisms, and passes the calibration results to the feedback closed-loop optimization module; the feedback closed-loop optimization module uses cross-stage calibration deviations and model overfitting flags as core inputs, automatically selects mismatched features, and reversely adjusts the feature construction strategy; The visualization explanation and intervention recommendation module constructs an interactive risk path map based on the structural visualization heat map and high-weight intervention factors, which is used to generate individual-oriented intervention recommendation vectors.

2. The personal obesity risk prediction system based on big data AI according to claim 1, characterized in that: The multidimensional feature construction module includes a three-layer nested structure. The first layer is the input reconstruction layer, which performs periodic smoothing on each type of time series data and uses a 3×1 convolution kernel plus a bidirectional LSTM to reconstruct the behavioral frequency data in time. The second layer is a multi-head nested attention structure, including a gait-heart rate joint attention head, a physiological-environmental fluctuation collaborative attention head, and a social-feeding cross attention head. Each attention head extracts short-term interdependencies based on the self-attention mechanism and window convolution and dynamically weights them; the third layer is a feature alignment layer, which uses the maximum information coefficient rearrangement technology to reorder the multi-head outputs according to the causal coupling strength, and outputs the three types of coupled feature groups of behavior-metabolism-environment through jump residual connections and outputs them to the next module according to the label window structure.

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

4. The personal obesity risk prediction system based on big data AI according to claim 1, characterized in that: The risk label dynamic generation module generates individual dynamic risk labels through the evolution clustering graph convolution structure under a dense time window. First, the multi-dimensional coupling feature sequence is subjected to a three-fold dimensionality reduction process, including principal component analysis, maximum variance preserving projection, and embedded graph kernel dimensionality reduction. Then, an evolution graph is constructed based on the density change rate of the graph structure. The TopK density peak is marked for each time slice, and the node aggregation of the peak mark is stratified by risk status through an adaptive boundary shrinkage function. The final output label matrix has point-by-point continuity on the time axis.

5. The personal obesity risk prediction system based on big data AI according to claim 4, characterized in that: The adaptive boundary shrinkage function is composed of a dynamic boundary sliding factor and a local label density function; The dynamic boundary slip factor is composed of the product of the time step and the characteristic volatility, dynamically compressing and expanding the cluster boundary; the local label density function adjusts the demarcation threshold of each type of risk label based on the compactness of the sample distribution in the current time window, ultimately achieving continuous expression of the individual's risk level in multiple time periods.

6. The personal obesity risk prediction system based on big data AI according to claim 1, characterized in that: The model training and risk prediction module includes a multi-graph fusion neural network structure and a causal factor decomposition unit; The multi-graph fusion graph neural network structure includes four types of graphs: behavioral graphs, physiological fluctuation graphs, social influence graphs, and environmental fluctuation graphs. Each type of graph is input in the form of a joint adjacency matrix and a node embedding matrix. The graph neural network uses a structured attention mechanism to hierarchically update node representations and outputs a global embedding by aggregating the weighted average of the central nodes. 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 personal obesity risk prediction system based on big data AI according to claim 6, characterized in that: 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 timestamp differences and applies edge update weights to nodes that are close in time. The group bias compensation factor uses the mean deviation of behavioral characteristics between clusters as an adjustment amount to standardize the structural deviation across groups.

8. The personal obesity risk prediction system based on big data AI according to claim 7, characterized in that: The minimum mutual information causal separation process comprises the following steps: First, a feature causal graph is constructed using the Bayesian structural learning mechanism, and the Markov equivalence clustering technology is used to screen out the uncaused structural paths; then the inverse probability weighting mechanism is introduced to resample the sample space so that the potential confounding factors are normalized and distributed similarly in the sample; finally, the reparameterization technology is used to perform second-order supervised reinforcement training on the stripped feature subset to generate high-resolution causal factors.

9. The AI-based personal obesity risk prediction system based on big data according to claim 1, characterized in that: The credibility assessment and calibration module constructs confidence intervals based on a multiple forward prediction mechanism, uses Monte Carlo Dropout to sample the model prediction results 100 times and calculates the range, mean and skewness coefficient of each set of outputs, then calibrates the distribution shape of the prediction scores through Beta distribution fitting, and outputs the risk level adjustment factor with the confidence upper bound as the decision limit, ultimately forming a feedback signal with the calibrated risk vector as input.

10. A method for predicting personal obesity risk based on AI using big data, characterized in that: The method for predicting personal obesity risk based on AI using big data uses the system for predicting personal obesity risk based on AI using big data as described in any one of claims 1 to 9.

Images