AI-fused multi-source data health risk prediction system

Abstract

The invention discloses an AI-fused multi-source data health risk prediction system, and relates to the field of health risk prediction.The AI-fused multi-source data health risk prediction system comprises the steps that firstly, wearable device data, historical medical record data and behavior log data are collected and preprocessed in a unified mode, and a standardized time sequence data set is generated; then, a self-supervised pre-training model is adopted to extract three types of potential risk features: a time sequence micro-variation anomaly index, a cross-modal interaction index and a high-dimensional space embedding deviation index; modeling the features by using an attention enhancement graph neural network, and constructing a health risk map of the target user; calculating a risk score through an atlas propagation path and a node embedding relation, and generating an interpretable causal chain for traceability analysis; when the score exceeds a set threshold value, early warning is triggered, and personalized intervention suggestions are output; according to the invention, early recognition and explanation of low-frequency high-risk health events can be realized, and the method has strong timeliness, interpretability and practical guidance value.

Description

Technical Field

[0001] This invention relates to the field of health risk prediction, and more specifically to a multi-source data health risk prediction system that integrates AI. Background Technology

[0002] With the advancement of medical informatization and individual health management, health risk prediction is becoming increasingly important in areas such as chronic disease prevention and control, critical illness screening, and public health emergency response. Currently, most health risk prediction systems rely primarily on single data sources (such as electronic health records and physical examination data) for modeling, making it difficult to effectively integrate multi-source heterogeneous data. This is especially true when facing low-incidence but high-risk health risk events (such as rhabdomyolysis, drug-induced arrhythmias, and sudden hyperkalemia), where prediction accuracy and response efficiency are severely inadequate.

[0003] These types of health events typically share the following characteristics: ① extremely low incidence and lack of adequately labeled training data; ② complex pathological mechanisms, related to subtle changes in multiple dimensions such as environment, behavior, physiology, and biochemistry; ③ rapid progression, requiring early warning to avoid serious consequences. Because traditional prediction models cannot handle sparse samples, high-dimensional unstructured data, and nonlinear latent variables, current technologies are insufficient to meet clinical needs. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-source data health risk prediction system that integrates AI to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-source data health risk prediction system integrating AI, comprising: Data acquisition module: Collects multi-source health data from several target users, including wearable device data, historical medical record data, and behavioral log data, to construct the original dataset D; Data preprocessing module: performs unified preprocessing on the original dataset D to generate a normalized time series dataset D′; Feature extraction module: Based on the historical labeled samples of the target health risk event, a self-supervised pre-trained model is used to extract features from the normalized time series dataset D′ to obtain the potential health risk associated feature set F, where F includes time series micro-variation features, cross-modal interaction features and high-dimensional space embedding features; Graph modeling module: Uses attention-enhanced graph neural network to model the potential health risk associated feature set F, and combines the structural correlation between multi-source data to construct the health risk graph G of each target user; Causal reasoning module: Based on the risk propagation path and intensity of different nodes in the health risk map G, calculate the risk score R of the target health event and generate an interpretable causal chain C, where C is used to indicate the possible triggering path and reasoning logic of the health risk. Early warning module: If the risk score R exceeds the set threshold, an early warning mechanism is triggered, and an interpretable causal chain C and recommended intervention suggestions are output.

[0006] Preferably, the step of using a self-supervised pre-trained model to extract features from the normalized time-series dataset D′ to obtain a potential health risk-related feature set F includes: We construct a self-supervised pre-training model based on a time encoder and mask prediction mechanism to learn feature representations on large-scale unlabeled time series data in an unsupervised manner. The normalized time series dataset D′ is input into the pre-trained model, and some data segments are hidden through a random time masking strategy to guide the model to learn the time dependence and change trend. The output of the intermediate hidden layer of the model is extracted as a potential health risk associated feature set F. The feature set F includes time-varying features, cross-modal interaction features, and high-dimensional spatial embedding features. The clustering consistency criterion is used to initially screen the feature set F, retaining key features whose information entropy is higher than a set threshold.

[0007] Preferably, the generation of a time-series micro-variation anomaly index after analyzing the time-series micro-variation characteristics includes: Perform sliding difference analysis on the time series micro-variation characteristics to calculate the incremental sequence between adjacent time steps; A dynamic fluctuation matrix is ​​constructed based on the incremental sequence, and the mean and variance of local changes are calculated. The incremental sequence is decomposed using the empirical mode decomposition algorithm to extract the intrinsic mode functions of different frequency components and identify high-frequency anomalous components; The energy proportion of high-frequency components is used as an indicator of anomaly contribution, and a time-series micro-variation anomaly index is generated through weighted integration.

[0008] Preferably, the cross-modal interaction index generated after analyzing cross-modal interaction features includes: Construct a modality correlation matrix for cross-modal interaction features and extract the weight coefficients between each modality; The modal correlation matrix is ​​quantified using a mutual information algorithm based on joint probability density estimation to quantify the degree of information sharing among modes; The mutual information values ​​are regularized, and a modal coupling degree vector is constructed to represent the relative contribution strength of different modalities to health risk prediction. The weighted index value is calculated based on the modal coupling degree vector to generate the cross-modal interaction index.

[0009] Preferably, the high-dimensional space embedding deviation index generated after analyzing the high-dimensional space embedding features includes: Construct a set of feature embedding vectors for the high-dimensional space embedding features, and calculate its center embedding vector during the training phase as a reference benchmark; The Mahalanobis distance algorithm is used to measure the deviation between the current sample embedding vector and the center embedding vector; Principal component analysis algorithm is introduced to reduce the dimensionality of the embedded vector set, retaining principal components with a cumulative contribution rate greater than 90%. A normalized deviation score is constructed based on the Mahalanobis distance result of the dimensionality-reduced vector, and a high-dimensional space embedding deviation index is generated.

[0010] Preferably, constructing the health risk profile G for each target user includes: Using the temporal micro-variation anomaly index, cross-modal interaction index, and high-dimensional space embedding deviation index as node features, an initial graph structure centered on the target user is constructed, where the edge weights represent the correlation coefficients between different features. The nodes in the graph are weighted and aggregated, and the edge weights between nodes are dynamically adjusted through adaptive learning. By using multi-layer graph convolution operations to propagate and update the graph structure layer by layer, high-order interaction features between nodes are extracted to obtain the embedded representation of each node in the health risk graph. The final graph G is generated based on node embedding representation and full graph aggregation features, which is used to characterize the comprehensive risk structure of an individual's health status in a multidimensional indicator space.

[0011] Preferably, the step of calculating the risk score R of the target health event based on the risk propagation path and intensity of different nodes in the health risk map G includes: Using the node embedding representation and edge weight relationship in the health risk graph G, a risk impact matrix based on the structure propagation model is constructed; A risk trigger threshold is set at the target health event node, and the maximum information flow algorithm is used to identify the set of upstream nodes that have a high-strength path connection with the node. A risk score R is generated based on the risk contribution of upstream nodes; Generate an interpretable causal chain C based on the propagation path between nodes.

[0012] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. This invention proposes an individual-oriented health risk prediction method by integrating multi-source heterogeneous health data with artificial intelligence modeling technology. It forms an end-to-end risk identification process, from raw data collection, feature extraction, and graph modeling to causal inference and intervention suggestion generation. The technical solution introduces three key feature indicators: the temporal micro-variation anomaly index, the cross-modal interaction index, and the high-dimensional space embedding deviation index, effectively enhancing the model's early perception capability of rare and sudden health risks. Simultaneously, the health risk graph constructed based on an attention-enhanced graph neural network enables structural modeling of complex relationships between features, providing a strong logical foundation for risk scoring and causal chain generation.

[0013] 2. This invention has significant advantages in feature representation, modeling accuracy, and result interpretability. First, by introducing a self-supervised learning mechanism and structured index design, it solves the feature learning challenge in scenarios with small samples and weak labels. Second, by combining causal path analysis and risk scoring mechanisms, it not only quantifies the degree of risk but also clearly presents the risk triggers and propagation paths. Third, the personalized intervention suggestions based on the risk structure output are highly targeted and responsive, providing users or medical personnel with highly operable early warning and response guidelines, significantly improving the intelligence and precision of health risk management. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0015] Figure 1 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] For examples, please refer to Figure 1 As shown in this embodiment, the multi-source data health risk prediction system integrating AI includes: Data acquisition module: Collects multi-source health data from several target users, including wearable device data, historical medical record data, and behavioral log data, to construct the original dataset D.

[0018] Step 1: Continuously sample physiological signal data collected by the user's smart health monitoring device (such as a wristband heart rate monitor, smartwatch, etc.) at a minute-level time granularity. The collected data includes, but is not limited to, heart rate, body temperature, blood oxygen saturation, steps, and sleep stage.

[0019] The second step involves processing biochemical test results and historical medical records from different medical institutions. First, a data parsing module designed based on health information interoperability standards (such as the HL7 FHIR protocol) performs format recognition and field mapping on the original medical records. The parsing process includes field renaming, unit conversion, field missing validation, and diagnostic code standardization (based on the International Classification of Diseases, ICD-10). Subsequently, a field-oriented nested structure generation strategy is used to organize the standardized medical data into structured health records. Fields include test time, biochemical indicator name, numerical result, unit, testing institution code, and diagnostic label.

[0020] Step 3: After obtaining user authorization, call the embedded sensors in the smart terminal to collect behavior log data, including screen usage time, application usage frequency, call records, movement trajectory and daily activity patterns.

[0021] Data preprocessing module: Performs unified preprocessing on the original dataset D to generate a normalized time series dataset D′.

[0022] Step 1: For the heterogeneous data contained in the original dataset D, first establish a unified timeline to coordinate the sampling frequency and acquisition time of different data sources. Set a fixed-length sliding window to slice the data of each type. The window length is set to 30 minutes and the step size is set to 10 minutes to ensure that data segments with consistent time sequence are formed within each sliding window.

[0023] Step 2: After completing data alignment and window slicing, perform missing field detection and completion operations for each data segment. When continuous variables are missing, a bidirectional linear interpolation method is used, that is, the nearest non-missing value is found on both sides of the missing position, and the value of the intermediate missing point is estimated with linear weights to ensure that the completed value is smooth and conforms to the data trend.

[0024] Step 3: Perform outlier identification on the numerical fields in each data window. Specifically, calculate the mean and standard deviation of each field in the training samples. For a new sampled value, if it exceeds the range of plus or minus three times the standard deviation of the field mean, it is considered an outlier. This strategy is based on the normal distribution assumption and exhibits good stability and interpretability in large sample scenarios. Detected outliers are not directly removed; instead, they are corrected using the median substitution method. That is, the median value of the field in the adjacent windows is used to replace the outlier, avoiding model training bias caused by extreme values.

[0025] Step 4: After completing missing value imputation and outlier correction, perform uniform normalization on all numeric fields. Use the min-max normalization method, which calculates the normalized value y for each value x based on its minimum and maximum values ​​in the field. The expression is: In the formula, x represents the original data value to be normalized, which comes from the numeric field before normalization. This indicates the minimum value of the field in the training dataset. This represents the maximum value of the field in the training dataset. This method maps all values ​​to between 0 and 1, which helps improve the convergence speed of model training and avoids gradient vanishing and feature weight bias problems. The normalization parameters are pre-calculated on the training set and remain unchanged during the testing phase to ensure the consistency of model inference.

[0026] Finally, after completing the above steps, a standardized time-series dataset D′ with uniform structure, continuous temporal sequence, and stable numerical values ​​is constructed, which serves as the standard input for subsequent feature extraction and model modeling.

[0027] Feature extraction module: Based on the historical labeled samples of the target health risk event, a self-supervised pre-trained model is used to extract features from the normalized time series dataset D′ to obtain the potential health risk associated feature set F, where F includes time series micro-variation features, cross-modal interaction features and high-dimensional space embedding features.

[0028] First, a self-supervised pre-trained model combining a temporal encoder and a mask prediction mechanism is constructed to automatically learn latent feature representations on large-scale unlabeled temporal series data. The model is based on an encoder-decoder architecture, where the encoder consists of stacked temporal attention layers to capture temporal dependencies across multiple time scales; the decoder reconstructs time segments hidden by the mask. The input to the temporal encoder is multi-source fields from a normalized temporal series dataset D′. During the encoding process, positional encoding is added to each time point to preserve temporal order information. The self-supervised training objective is to minimize the mean squared error between the original and reconstructed values, driving the model to learn latent distribution patterns from the data without manual labeling.

[0029] After inputting the normalized time-series dataset D′ into the model, a random time masking strategy is introduced to improve the model's sensitivity to time dependencies and subtle changes. This strategy randomly selects no more than 15% of consecutive time periods on the time axis of each training sample and replaces the corresponding input features with zero or average values, simulating missing data scenarios and guiding the model to learn long-term dependencies and local trend changes. In the encoding-decoding process, the model reconstructs the masked portion using contextual information, thereby capturing weak but crucial dynamic change features, suitable for modeling precursors of anomalies in health data.

[0030] After the model has completed training, the output of the intermediate hidden layers of the model is taken as the potential health risk associated feature set F, which includes the following three types of sub-features: (1) Time-varying characteristics, which represent the small fluctuations of continuous variables per unit time, are used to reflect the potential unstable trend of health status; (2) Cross-modal interaction features represent the dynamic correlation between different data modalities (such as physiological data and environmental data). The weights between each modality are calculated and combined by multi-head attention mechanism. (3) High-dimensional space embedding features refer to the dense vector representation of the input original data in the model embedding space, which captures the complex nonlinear interaction information between data. The above features are uniformly output as a three-dimensional tensor, with the dimension being time step × feature dimension × number of modes.

[0031] The feature set F is initially screened using a clustering consistency criterion. This method first uses the K-means algorithm to cluster the feature set F, setting the number of clusters K to 8, and calculates the distribution stability of each feature among the cluster centers. Then, the Shannon information entropy of each feature is calculated, where the entropy value is represented as the negative logarithmic weighted sum of the proportions of that feature across all clusters. The specific formula is: In the formula, H represents the information entropy value of the feature, which measures the dispersion of the feature in different cluster categories. The higher the entropy value, the greater the contribution of the feature to class distinction. K represents the number of clusters, which is fixed at 8 in the feature selection method. The frequency of a feature in the i-th category is expressed as the number of samples containing that feature in the i-th category divided by the total number of samples containing that feature across all categories. An information entropy threshold of 1.5 is set, retaining only features with entropy values ​​higher than this threshold as key features for subsequent steps. This strategy ensures that the selected features possess strong information discrimination and sample generalization capabilities.

[0032] To identify hidden short-term mutation trends in healthy states, the following steps are performed on time-series microvariation characteristics to generate a time-series microvariation anomaly index: Step 1: Based on the extracted temporal micro-variation features, a sliding window with a fixed step size of 1 is set, and the values ​​between adjacent time steps are differentially processed to obtain the increment sequence. Let the feature value of continuous time points be xt, then the increment value is Δxt = xt−x(t−1). This operation is used to reveal the rate and direction of change of temporal features over a short period of time.

[0033] The second step involves reorganizing the incremental sequences into a two-dimensional matrix based on time windows, where rows represent time segments and columns represent different feature fields. Within each local window, the mean and variance of the sequence are calculated to characterize the center and amplitude of fluctuations in the health status within that window, providing a distribution reference for subsequent mode decomposition.

[0034] The third step involves applying the Empirical Mode Decomposition (EMD) algorithm to the incremental sequence, decomposing the original nonlinear nonstationary sequence into several Intrinsic Mode Functions (IMFs) with different frequency components. This algorithm extracts local extrema through an iterative envelope averaging process, forming a set of components from high to low frequencies, while retaining the high-frequency modal components representing sudden changes.

[0035] Step 4: Calculate the energy percentage of each high-frequency component, assuming the energy of each high-frequency component is... Then the total energy is all The sum of these components. The ratio of high-frequency energy to total energy is taken as the anomaly contribution, and different components are assigned weights. Finally, an anomaly index is constructed. The expression is: ;in It is a temporal microvariation anomaly index used to quantify the potential risk of sudden changes in an individual's health status within a short period of time.

[0036] To measure the dynamic coupling relationships among multi-source health data, the following process is constructed based on cross-modal interaction features to generate a cross-modal interaction index: Step 1: Extract the intermodal weight relationships from the cross-modal interaction features and establish a modality correlation matrix. Each matrix element ai,j represents the interaction strength between modality i and modality j learned through a multi-head attention mechanism. The attention mechanism captures the correlation by normalizing the dot product between modality vectors.

[0037] Step 2: Based on the modality correlation matrix, the mutual information I(i,j) between mode i and mode j is calculated using a joint probability density estimation algorithm. This mutual information metric reflects the amount of information shared between the two modes in the joint distribution, and is defined as: ;in Let i be the joint probability density of mode i and mode j. , Each is distributed along its own edge.

[0038] Step 3: Perform min-max normalization on each mutual information value, map it to the interval between 0 and 1, and then construct the modal coupling vector C=[c1,c2,...,cn], where each element represents the intensity ratio of the corresponding mode's contribution to health risk prediction.

[0039] Step 4: Calculate the weighted average value as the interaction index based on the modal coupling degree vector. The expression is: ;in The contribution weights for each modality are preset (which can be obtained through cross-validation). IC is the cross-modal interaction index, which reflects the degree of synergy between multi-source data and its comprehensive impact on health status.

[0040] To identify anomalous deviations of individual samples in the latent embedding space, the following steps are performed based on the high-dimensional embedding features to generate a high-dimensional embedding deviation index: Step 1: Collect the high-dimensional spatial embedding feature vectors corresponding to all healthy samples during the training phase, denoted as set V={v1,v2,...,vn}, and calculate its central embedding vector vc, which is the mean of all embedding vectors, representing the spatial center of the normal sample.

[0041] Step 2: Calculate the deviation of the current predicted sample embedding vector vt from the center vector vc using the Mahalanobis distance algorithm. The expression is: In the formula, S is the covariance matrix of the embedded vector set. It represents a measure of the anomaly of a sample in its spatial distribution.

[0042] The third step is to introduce principal component analysis to reduce the dimensionality of the embedded vector set. Select the top principal components with a cumulative contribution rate greater than 90% to obtain the dimensionality-reduced vectors vt′ and vc′, so as to reduce the impact of high-dimensional data on the stability of covariance matrix inversion in Mahalanobis distance calculation.

[0043] Step 4: Perform max-min normalization on the dimensionality-reduced Mahalanobis distance values. The normalization expression is: Where Dmin and Dmax are the minimum and maximum values ​​of Mahalanobis distance in the training set, respectively, and the final result ID is the high-dimensional embedding deviation index, which is used to represent the degree of deviation of the sample from the normal state in the health feature distribution space.

[0044] The graph modeling module uses an attention-enhanced graph neural network to model the potential health risk associated feature set F, and combines the structural correlations between multi-source data to construct the health risk graph G for each target user.

[0045] Step 1: Centered on the target user, establish a heterogeneous attribute graph structure, where each node in the graph represents a core health risk characteristic, including the temporal micro-variation anomaly index, the cross-modal interaction index, and the high-dimensional space embedding deviation index. The initial feature vector of each node consists of time series segments of the corresponding indicator, representing its dynamic trend.

[0046] In the graph, the edges between nodes represent the correlation between indicators, and the edge weights are constructed by calculating the Pearson correlation coefficient between their historical sequences. Specifically, for any two feature sequences A and B, the edge weight wAB is calculated using the following formula: Where cov(A,B) is the covariance, and σA and σB are the standard deviations of the feature sequences, respectively. This method can effectively quantify the linear correlation between different features and is used to initialize the connection strength of graph structures.

[0047] Based on the initial graph structure, a structural attention mechanism is introduced to perform weighted aggregation of features of neighboring nodes. The single-head attention method from Graph Attention Networks is used to calculate attention scores for the feature vectors of each node's neighbors.

[0048] Attention scores are calculated as follows: Let the feature vectors of node i and node j be... and Then its attention weight Defined as: Where W is the shared linear transformation matrix, and a is the attention weight vector. This represents a vector concatenation operation. Let be the set of neighbors of node i. By adaptively learning this attention weight, the edge weights between nodes are dynamically adjusted, so that key nodes contribute more to the learning of the graph structure.

[0049] On the updated graph structure, multi-layer graph convolution operations are applied for feature propagation. Each layer of graph convolution updates the representation of the current node using the aggregated information of its neighbors, forming a high-order interaction structure of cross-dimensional features. The update expression for each layer of graph convolution is: ;in, Let represent the representation vector of node j in the l-th layer. Let be the learnable weight matrix of the l-th layer, and σ be the activation function (such as ReLU). By stacking multiple layers of graph convolution, deep interactions between metrics can be captured, forming stable embedded representations.

[0050] Step 4: After the propagation process of the graph neural network is completed, the embedding vector of each node in the final layer is extracted as the high-order feature representation of the indicator. The embedding vectors of all nodes are aggregated at the graph level, and the global graph vector is obtained by using the full graph average pooling method, which represents the overall health risk of the user in the current time window.

[0051] Finally, the embedded representations of each node in the graph are fused with the aggregated features of the entire graph to construct a health risk graph G. Graph G not only preserves the degree of abnormality of an individual's health status in different dimensions, but also expresses the structural relationships between these features, providing a high-dimensional structural foundation for subsequent risk scoring and causal analysis.

[0052] Causal reasoning module: Based on the risk propagation path and intensity of different nodes in the health risk map G, calculate the risk score R of the target health event and generate an interpretable causal chain C, where C is used to indicate the possible triggering paths and reasoning logic of the health risk.

[0053] Based on the node embedding representation and edge weight parameters in the health risk graph G, a risk impact matrix is ​​constructed to characterize the risk propagation capability between nodes in the graph. Each node represents a health risk-related index (including the temporal micro-variation anomaly index, the cross-modal interaction index, and the high-dimensional space embedding deviation index), and the weight of each edge represents the strength of the risk impact between nodes, which is derived from the adaptive edge weights obtained during the graph structure learning stage.

[0054] The risk impact matrix M is a symmetric matrix, and its i,j-th element is... The risk propagation capability of node i to node j is represented by the following formula: ;in, Let be the edge weights from node i to node j (learned through an attention mechanism). This matrix comprehensively considers node feature similarity and structural connection strength to model potential causal propagation relationships.

[0055] In the graph G, select the target node corresponding to the target health event as the risk receiver. Set a risk trigger threshold θ at this node. When the cumulative risk impact value exceeds the threshold, the risk is considered to have met the triggering conditions.

[0056] The Maximum Information Flow Algorithm is employed to perform path search in the graph to identify the set of upstream nodes connected to the target node by high-intensity propagation paths. The information flow strength is determined by the minimum edge weight on the path. The algorithm aims to find the set of paths with the maximum information flow from any node to the target node throughout the entire graph, where the information flow of each path is not less than a preset threshold θ = 0.7. This process is used to identify causal node groups with direct or indirect risk-driving capabilities.

[0057] The identified upstream nodes are used as risk contribution sources, and the overall risk score R is calculated based on their relative influence in the propagation path. The risk contribution ci of each upstream node is determined by the following weighted factors: the average edge weight of the path between node i and the target node; the similarity between the embedding vector of node i in the graph and that of the target node; and the reciprocal of the path propagation depth (i.e., the number of hops). The formula for calculating the overall risk score R is: ;in, Let represent the set of path edge weights from node i to the target node. Let represent the vector cosine similarity between node i and the target node. Indicates the number of hops in the propagation path. These are weighting coefficients, with a weight sum of 1. The default settings are 0.4, 0.4, and 0.2. The score R is a continuous value between 0 and 1, representing the overall risk probability of a health event occurring.

[0058] Based on the high-impact paths identified in the risk impact matrix, an interpretable causal chain C is generated. Causal chain C represents a sequence that propagates step by step from one or more upstream high-risk nodes through a structural path to the target health event node. Each node in the chain is a high-impact node with an edge weight higher than a set threshold (e.g., 0.5).

[0059] The causal chain C is expressed as a directed weighted path, with each node's specific risk value and its position in the propagation chain, forming the following structure: ;in The target health event node is represented by several preceding nodes, which are potential risk triggers. This structure allows for the tracking of the potential causes and evolutionary paths of health risks, assisting doctors or users in interpretation and intervention.

[0060] Early warning module: If the risk score R exceeds the set threshold, an early warning mechanism is triggered, and an interpretable causal chain C and recommended intervention suggestions are output.

[0061] Step 1: Set a risk score trigger threshold θR to determine whether an individual is at potential health risk. The threshold θR can be an empirical value or set based on the distribution of historical case data. A recommended initial value is 0.65, meaning that an alert is triggered when the risk score R is greater than or equal to this threshold.

[0062] The judgment condition expression is: if R≥θR, then it is considered that there is a high probability of a health event occurring, and the early warning response process is initiated; otherwise, only the risk score is recorded for subsequent trend tracking.

[0063] Step 2: Once the triggering conditions are met, extract the generated interpretable causal chain C from the graph analysis results and output it in a structured format. Causal chain C consists of a set of ordered risk nodes, representing a high-risk path from a potential trigger to the target health event. Each node contains the following fields: Node name (corresponding to specific health indicators, such as "sleep disorder index"); Node risk value (risk contribution of node embedding vector and propagation path); The weight of the edge connecting to the previous node (representing the propagation strength); Timestamp (indicating the point in time when the abnormal index first exceeded the normal fluctuation range).

[0064] This causal chain is presented in a graphical and visual form, and can be directly used in the risk tracing and explanation module of clinical or individual health management platforms.

[0065] Step 3: Based on the key node types and risk patterns involved in the causal chain C, construct the intervention recommendation vector VI, where each vector element corresponds to a weight value for a specific intervention type. Intervention types include, but are not limited to: behavioral recommendations, physiological monitoring recommendations, medical treatment recommendations, and environmental adjustment recommendations. The vector is constructed as follows: If a causal node is a behavior-related feature (such as "circadian rhythm disorder"), then the weight of behavior-related suggestions will be increased. If high-frequency abnormalities are concentrated in physiological signals (such as sudden changes in heart rate), physiological monitoring and clinical consultation suggestions will be prioritized. If environmental modal factors contribute significantly, environmental exposure interventions (such as ventilation or travel adjustments) are recommended.

[0066] Once the system is built, a similarity matching algorithm (such as cosine similarity) is used to search for the standard suggestion entry that is most similar to the intervention vector in the preset intervention suggestion library, and outputs it as a personalized suggestion list.

[0067] Step 4: Combine the risk score (R), causal chain (C), and intervention recommendation list to form a complete health risk warning report. The report format includes: warning level (based on risk score classification, 0.65–0.8 is medium risk, ≥0.8 is high risk); risk score (R value); interpretable causal chain diagram; summary of personalized intervention recommendations and recommended action path.

[0068] The report is automatically pushed to the target user's health terminal or synchronized to the medical service platform for professionals' reference. Simultaneously, the risk score trend is tracked in subsequent time windows, and the warning status and intervention measures are dynamically adjusted.

[0069] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A multi-source data health risk prediction system integrating AI, characterized by: include: Data acquisition module: Collects multi-source health data from several target users, including wearable device data, historical medical record data, and behavioral log data, to construct the original dataset D; Data preprocessing module: performs unified preprocessing on the original dataset D to generate a normalized time series dataset D′; Feature extraction module: Based on the historical labeled samples of the target health risk event, a self-supervised pre-trained model is used to extract features from the normalized time series dataset D′ to obtain the potential health risk associated feature set F, where F includes time series micro-variation features, cross-modal interaction features and high-dimensional space embedding features; Graph modeling module: Uses attention-enhanced graph neural network to model the potential health risk associated feature set F, and combines the structural correlation between multi-source data to construct the health risk graph G of each target user; Causal reasoning module: Based on the risk propagation path and intensity of different nodes in the health risk map G, calculate the risk score R of the target health event and generate an interpretable causal chain C, where C is used to indicate the possible triggering path and reasoning logic of the health risk. Early warning module: If the risk score R exceeds the set threshold, an early warning mechanism is triggered, and an interpretable causal chain C and recommended intervention suggestions are output.

2. The multi-source data health risk prediction system integrating AI according to claim 1, characterized in that: The method employs a self-supervised pre-trained model to extract features from the normalized time-series dataset D′, obtaining a potential health risk-related feature set F, including: We construct a self-supervised pre-training model based on a time encoder and mask prediction mechanism to learn feature representations on large-scale unlabeled time series data in an unsupervised manner. The normalized time series dataset D′ is input into the pre-trained model, and some data segments are hidden through a random time masking strategy to guide the model to learn the time dependence and change trend. The output of the intermediate hidden layer of the model is extracted as a potential health risk associated feature set F. The feature set F includes time-varying features, cross-modal interaction features, and high-dimensional spatial embedding features. The clustering consistency criterion is used to initially screen the feature set F, retaining key features whose information entropy is higher than a set threshold.

3. The multi-source data health risk prediction system integrating AI according to claim 2, characterized in that: The time-series micro-variation anomaly index generated after analyzing the characteristics of time-series micro-variation includes: Perform sliding difference analysis on the time series micro-variation characteristics to calculate the incremental sequence between adjacent time steps; A dynamic fluctuation matrix is ​​constructed based on the incremental sequence, and the mean and variance of local changes are calculated. The incremental sequence is decomposed using the empirical mode decomposition algorithm to extract the intrinsic mode functions of different frequency components and identify high-frequency anomalous components; The energy proportion of high-frequency components is used as an indicator of anomaly contribution, and a time-series micro-variation anomaly index is generated through weighted integration.

4. The multi-source data health risk prediction system integrating AI according to claim 2, characterized in that: The cross-modal interaction indices generated after analyzing the cross-modal interaction characteristics include: Construct a modality correlation matrix for cross-modal interaction features and extract the weight coefficients between each modality; The modal correlation matrix is ​​quantified using a mutual information algorithm based on joint probability density estimation to quantify the degree of information sharing among modes; The mutual information values ​​are regularized, and a modal coupling degree vector is constructed to represent the relative contribution strength of different modalities to health risk prediction. The weighted index value is calculated based on the modal coupling degree vector to generate the cross-modal interaction index.

5. The multi-source data health risk prediction system integrating AI according to claim 2, characterized in that: After analyzing the high-dimensional spatial embedding features, a high-dimensional spatial embedding deviation index is generated, including: Construct a set of feature embedding vectors for the high-dimensional space embedding features, and calculate its center embedding vector during the training phase as a reference benchmark; The Mahalanobis distance algorithm is used to measure the deviation between the current sample embedding vector and the center embedding vector; Principal component analysis algorithm is introduced to reduce the dimensionality of the embedded vector set, retaining principal components with a cumulative contribution rate greater than 90%. A normalized deviation score is constructed based on the Mahalanobis distance result of the dimensionality-reduced vector, and a high-dimensional space embedding deviation index is generated.

6. The multi-source data health risk prediction system integrating AI according to claim 1, characterized in that: The construction of the health risk map G for each target user includes: Using the temporal micro-variation anomaly index, cross-modal interaction index, and high-dimensional space embedding deviation index as node features, an initial graph structure centered on the target user is constructed, where the edge weights represent the correlation coefficients between different features. The nodes in the graph are weighted and aggregated, and the edge weights between nodes are dynamically adjusted through adaptive learning. By using multi-layer graph convolution operations to propagate and update the graph structure layer by layer, high-order interaction features between nodes are extracted to obtain the embedded representation of each node in the health risk graph. The final graph G is generated based on node embedding representation and full graph aggregation features, which is used to characterize the comprehensive risk structure of an individual's health status in a multidimensional indicator space.

7. The multi-source data health risk prediction system integrating AI according to claim 1, characterized in that: The step of calculating the risk score R of the target health event based on the risk propagation path and intensity of different nodes in the health risk map G includes: Using the node embedding representation and edge weight relationship in the health risk graph G, a risk impact matrix based on the structure propagation model is constructed; A risk trigger threshold is set at the target health event node, and the maximum information flow algorithm is used to identify the set of upstream nodes that have a high-strength path connection with the node. A risk score R is generated based on the risk contribution of upstream nodes; Generate an interpretable causal chain C based on the propagation path between nodes.

Images