Machine learning-based intelligent assessment method and system for diabetes risk

Abstract

This invention relates to a machine learning-based intelligent assessment method and system for diabetes risk, belonging to the field of smart healthcare. The method involves each participant performing causal learning based on local sample data to construct a local sample causal relationship graph. Cross-party alignment and comparison of the local causal relationship graphs of each participant establishes an interaction network among them. Based on this interaction network, federated learning is performed on the local prediction model parameters of each participant, and reinforcement regions are labeled. Based on these reinforcement regions, each participant performs targeted reinforcement learning to obtain enhanced model parameters. These enhanced model parameters are used as the initial parameters for the next round of federated learning, and training is repeated until a preset convergence condition is met to obtain global model parameters. The resulting global model parameters are then used to assess the health status of the target individual and obtain the risk assessment result. This method addresses the problems of insufficient sample size and knowledge conflicts in single-center models, achieving improved performance in diabetes risk assessment while maintaining privacy compliance.

Description

Technical Field

[0001] This invention relates to the field of smart healthcare technology, and in particular to a machine learning-based intelligent assessment method and system for diabetes risk. Background Technology

[0002] Type 2 diabetes is a common chronic metabolic disease, and among its microvascular complications, diabetic nephropathy is one of the leading causes of end-stage renal failure. Early identification of the risk of developing diabetic nephropathy and targeted intervention are of great significance for delaying disease progression and improving patient prognosis. In recent years, with the development of medical informatization, diabetes risk assessment technology based on machine learning methods has gradually become a research hotspot.

[0003] Currently, machine learning in diabetes risk assessment mainly employs two technical approaches. One is a prediction model based on single-center data, utilizing electronic health records of patients accumulated in a single hospital to train a risk classifier using logistic regression, random forest, support vector machine, or deep learning models. However, single-center models face the inherent problem of limited sample size: the amount of labeled data accumulated by any single hospital is usually insufficient to support the fitting requirements of deep learning models for complex nonlinear relationships. Especially when the proportion of positive diabetic nephropathy samples is low, the model is prone to overfitting, leading to a significant decrease in accuracy in actual validation.

[0004] Another approach is multi-center data fusion training, which attempts to integrate samples from different hospitals to expand the training set. However, due to the privacy concerns involved in medical data and restrictions imposed by laws and regulations such as the Personal Information Protection Law and the Regulations on the Management of Human Genetic Resources, raw medical record data cannot be directly aggregated to a central server for centralized training. Traditional methods of sharing data after anonymization also face legal risks and the potential for data leakage. Therefore, most hospitals prefer independent modeling to the compliance risks of data leakage. Summary of the Invention

[0005] This invention addresses the technical problems in existing technologies, such as insufficient samples and heterogeneous knowledge leading to low model accuracy and lack of interpretability under multi-center data privacy protection, by providing a machine learning-based intelligent assessment method and system for diabetes risk.

[0006] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a machine learning-based intelligent assessment method for diabetes risk, comprising: Step S1: Each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, obtains local causal model parameters, and initializes local prediction model parameters.

[0007] Step S2: Perform cross-party alignment comparison on the local causal relationship graphs of each participant, and establish an interaction relationship network among the participants based on the comparison results.

[0008] Step S3: Based on the interaction relationship network, perform federated learning on the local prediction model parameters of each participant to obtain federated learning model parameters, and perform multi-dimensional evaluation and interpretation of the federated learning model parameters, and label the reinforcement regions.

[0009] Step S4: Return the reinforcement region to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain the enhanced model parameters.

[0010] Step S5: Use the enhanced model parameters as the initial parameters for the next round of federated learning, and repeat steps S3 to S4 until the preset convergence condition is met to obtain the global model parameters.

[0011] Step S6: Deploy the risk assessment scenario model using the global model parameters, conduct a risk assessment on the health status of the target individual, and obtain the risk assessment results.

[0012] Secondly, the present invention provides a machine learning-based intelligent assessment system for diabetes risk, comprising: The local training module is used to perform step S1: Each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, obtains local causal model parameters, and initializes local prediction model parameters.

[0013] The participant causal comparison module is used to perform step S2: perform cross-party alignment comparison on the local causal relationship graphs of each participant, and establish an interaction relationship network between participants based on the comparison results.

[0014] The federated learning module is used to perform step S3: based on the interaction network, perform federated learning on the local prediction model parameters of each participant to obtain federated learning model parameters, and perform multi-dimensional evaluation and interpretation of the federated learning model parameters, and label reinforcement regions.

[0015] The reinforcement learning module is used to execute step S4: return the reinforcement region to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain the enhanced model parameters.

[0016] The federated iteration module is used to execute step S5: using the enhanced model parameters as the initial parameters for the next round of federated learning, repeating steps S3 to S4 until the preset convergence condition is met, and obtaining the global model parameters.

[0017] The risk assessment module is used to perform step S6: deploy a risk assessment scenario model using the global model parameters, assess the health status of the target individual, and obtain the risk assessment result.

[0018] The beneficial effects of this invention are as follows: By having each participant first transform the raw data into a privacy-free causal relationship graph locally, compliance obstacles to data sharing are resolved at the source. Then, cross-party comparisons are used to construct an interactive relationship network to quantify knowledge heterogeneity. The causal relationship graph is then used to guide federated learning to achieve reinforcement of beneficial knowledge and complementation of missing knowledge. Next, the aggregated model is interpreted in multiple dimensions and reinforcement regions are labeled. Subsequently, the reinforcement regions are returned to each party for targeted enhancement, distinguishing between maintaining strong regions and supplementing weak regions. Finally, the enhanced parameters are looped back to federated learning for repeated iterations until convergence. This series of methods is interconnected, ultimately integrating causal knowledge from multi-center heterogeneous data without exposing the original data. This effectively solves the problems of low model accuracy and poor generalization ability caused by insufficient sample size and knowledge conflicts in single-center models. Simultaneously, through interpretable labeling and targeted enhancement, it overcomes the shortcomings of traditional black-box models that lack interpretability and continuous optimization mechanisms. This achieves a significant breakthrough in diabetes risk assessment performance across heterogeneous multi-center data under the premise of privacy compliance, effectively improving the technical effect of machine learning in diabetes risk assessment accuracy. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the intelligent diabetes risk assessment method based on machine learning provided by the present invention.

[0020] Figure 2 A schematic diagram of the structure of the intelligent diabetes risk assessment system based on machine learning provided by the present invention.

[0021] Figure labeling: Local training module 1, participant causal comparison module 2, federated learning module 3, reinforcement learning module 4, federated iteration module 5, risk assessment module 6. Detailed Implementation

[0022] 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, and 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.

[0023] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0024] In the description of this invention, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this invention is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed herein.

[0025] Example 1: like Figure 1 As shown, embodiments of the present invention provide a machine learning-based intelligent assessment method for diabetes risk, including: Step S1: Each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, obtains local causal model parameters, and initializes local prediction model parameters.

[0026] Furthermore, each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, and obtains local causal model parameters, including: expanding the local diabetes sample data by time slices, constructing cross-layer directed edges for variables between adjacent time slices to represent time-lag causal relationships, and constructing same-layer directed edges for variables within the same time slice to represent instantaneous causal relationships; simultaneously analyzing the direction and strength of causal edges for time-lag causal relationships and instantaneous causal relationships by minimizing a loss function with a cycle constraint penalty term, and outputting a local sample causal relationship graph containing the time dimension, wherein the local causal model parameters include the direction identifier, confidence score, and causal strength coefficient of each directed edge in the local sample causal relationship graph.

[0027] Specifically, each participating party trains its local assessment model based on local type 2 diabetes sample data. This is achieved by acquiring longitudinal follow-up data of type 2 diabetes patients stored in the local electronic health record system. Each record includes the patient's baseline static information, such as age, gender, disease duration, smoking history, and family history, as well as time-series observation indicators recorded at each follow-up time point, such as glycated hemoglobin (HbA1c), systolic blood pressure, diastolic blood pressure, fasting blood glucose, postprandial blood glucose, total cholesterol, triglycerides, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, serum uric acid, serum creatinine, estimated glomerular filtration rate (eGFR), and urinary microalbumin to creatinine ratio (UACR). Because the time points of consultation vary among patients, time alignment is required. A three-month time window is used to assign each patient's test values ​​to the corresponding time window. If multiple test values ​​are recorded within the same window, the median is used as the representative value for that time point. After processing, each patient is represented on the timeline as a sequence composed of multiple time slices, with each slice containing a vector of simultaneously measured indicators. Next, we need to identify the causal relationships between these indicators over time. Here, we employ a causal discovery algorithm specifically designed for time-series data. Causal relationships are divided into two categories: one is lagged causality across time, such as how glycated hemoglobin levels three months ago affect current kidney function indicators; the other is immediate causality within the same point in time, such as which influences which at the same moment—glycated hemoglobin or blood pressure. The algorithm expands all time slices and uses the current and historical values ​​of each variable as candidate explanatory variables. A loss function is used to measure the error in predicting the current variable using these variables. Two penalty terms are added to the loss function: a sparsity penalty, which forces the algorithm to retain only truly important causal connections, avoiding the generation of a large number of spurious relationships; and an acyclicity penalty, ensuring that the final causal graph is a directed acyclic graph, avoiding logically illogical situations such as circular dependencies.

[0028] During training, an adjacency matrix is ​​first randomly initialized. Then, gradient descent is used iteratively to gradually adjust the weights in the matrix, causing the loss function to decrease continuously. Each iteration calculates the prediction error and penalty term of the current model and updates the weights of each connection in the matrix accordingly. After multiple iterations, the loss function converges to its minimum, resulting in a weighted adjacency matrix. Connections with very small absolute values ​​(usually less than 0.03) are considered numerical noise and removed. The remaining connections form the basic framework of the causal graph. Finally, information about each directed edge is extracted from this skeleton, including the names of the start and end nodes, the direction of the edge (e.g., glycated hemoglobin-eGFR), and two values: a confidence score reflecting the robustness of the edge in the data, and a causal strength coefficient representing the relative strength of the causal relationship. All this information is compiled into a structured causal graph description file. The file only contains variable names and edge relationships, without any original patient test values, consultation dates, or personal identification information. The causal graph is all that local participants can upload to the central server. The original data always remains within the hospital and does not leave the local system. Local causal learning is thus completed. Through this process, the hospital transforms its clinical data into a causal knowledge map, preserving the core logic of dynamic dependencies between variables while completely avoiding the risk of patient privacy leaks.

[0029] Furthermore, the local sample causal relationship graph is a directed acyclic graph structure, where nodes are used to identify the types of variables involved in causal inference, and directed edges are used to characterize the data dependencies and causal directions between variables. The nodes include: static feature nodes, time-series observation nodes, and fluctuation-derived nodes. Static feature nodes characterize attribute variables that do not change or change slowly over time during the patient's baseline period, including age, gender, duration of diabetes, smoking history, and family history. Time-series observation nodes characterize longitudinal follow-up vital signs variables collected along a uniform time axis. Fluctuation-derived nodes characterize secondary derived variables obtained by calculating the coefficient of variation, mean true fluctuation range, and continuous slope changes of selected indicators in the time-series observation nodes, and are included as independent nodes in the causal relationship graph.

[0030] Specifically, the resulting causal relationship diagram of the local samples adopts a directed acyclic graph structure. Each node represents a variable involved in causal inference, and each directed edge points from the cause node to the result node, indicating the direction of the former's direct influence on the latter. Based on the data type and clinical significance of the variables, the nodes are divided into three categories: The first category is static feature nodes. These nodes correspond to attributes collected at the patient's baseline that do not change over time or change slowly over time, including the patient's age, gender, duration of diabetes, smoking history, and family history. These are typically recorded once upon initial enrollment and do not change or change very little in subsequent follow-ups. In the causal relationship diagram, these are considered fixed background conditions and may serve as cause nodes pointing to various indicators observed during the follow-up periods.

[0031] The second category is time-series observation nodes, which correspond to longitudinal follow-up vital signs variables collected according to a unified time axis. When processing them, it is necessary to first align the irregular medical visit time points of different patients to a unified discrete time window, such as a window every three months. Within each window, the median of multiple measurements of the same variable is taken as the representative value of the corresponding time point. Each patient has a set of indicators measured simultaneously at each time point.

[0032] The third type is volatility-derived nodes. These nodes are not the original measured values, but rather secondary derived variables calculated based on selected indicators from time-series observation nodes. The reason for constructing these nodes separately is that clinical studies have found that long-term volatility in blood glucose and blood pressure is an independent risk factor for diabetic nephropathy, and its predictive value is often no less than, and even more important than, the average level. If only the original values ​​of each measurement are used as input to the time-series nodes, the causal discovery algorithm can capture volatility patterns, but it will be highly dependent on the division of the time window and the sequence length. Moreover, volatility as a holistic feature is difficult to directly separate from the causal relationship between each measurement value.

[0033] This embodiment incorporates volatility as an independent node into the causal relationship diagram. Specifically, for each patient, three indicators—glycated hemoglobin, systolic blood pressure, and pulse pressure—are extracted from the complete follow-up sequence. Other indicators can also be added as needed clinically. Three volatility measures are calculated: the first is the coefficient of variation, which is equal to the standard deviation of the indicator divided by its mean, reflecting the overall volatility relative to the average level; the second is the mean true range of volatility, which is the average of the absolute values ​​of the difference between two adjacent measurements, reflecting the average magnitude of the indicator change between each follow-up; and the third is the cumulative change of the continuous slope. First, a linear trend line is fitted to the time-series data of the indicator to obtain the overall slope, and then the sum of the squares of the segmented slopes between each adjacent point is calculated. The combination of the two reflects whether the trend of the indicator change over time is stable, unidirectionally rising, or violently oscillating. Each variability metric is calculated to generate a numerical value, which serves as a static attribute for the patient. This attribute node is added to the causal graph, alongside the static feature nodes, and can establish directed edges with other nodes, such as eGFR, urinary microalbumin, and whether kidney disease eventually occurs. In this way, the causal discovery algorithm can directly determine whether blood glucose variability independently affects the decline in kidney function, rather than inferring it indirectly or implicitly. These three types of nodes together constitute all the vertices of the causal graph. The graph does not contain any patient's original test values ​​or timestamps, but only retains the direction and strength of the dependencies between variables, thus allowing for safe transmission and comparison between different participants.

[0034] Furthermore, the parameters of the local prediction model are initialized, including: using the local sample causal relationship graph as prior knowledge constraints, constructing a time series prediction model based on a long short-term memory network; using the causal edges determined in the causal relationship graph as the mandatory information transmission paths between time steps in the time series prediction model structure, and applying L1 sparsity penalty to the variable connections that do not exist in the causal relationship graph in the model loss function, so that the information flow direction of the time series prediction model is strictly aligned with the causal relationship graph; wherein, the initialization uses a Xavier uniform distribution to randomly initialize the model weights, and the causal strength coefficient is used as the initial attention weight of the corresponding connection path.

[0035] Specifically, after constructing the local causal graph, each participant needs to build a time-series model for predicting the risk of diabetic nephropathy based on the local causal graph. This embodiment uses a Long Short-Term Memory (LSTM) network as the basic model and utilizes the dependencies between variables already discovered in the local causal graph to constrain the information flow direction of the LSTM, ensuring that the model learns in a direction consistent with medical logic from the beginning, rather than blindly searching for parameters from scratch. Specifically, causal edges are mapped to information transmission paths in the LSTM. The core function of the LSTM is to process the input sequence sequentially according to time steps. At each time step, the model receives the current detection indicators such as glycated hemoglobin and blood pressure, as well as the hidden state passed from the previous time step, and then outputs the prediction result for the current time step and updates the hidden state. In a standard LSTM, each dimension of the input vector is connected to each neuron in the hidden layer through a fully connected manner. This means that any two variables can be linked, regardless of whether there are edges in the local causal graph. While this approach is flexible, it is also prone to learning a large number of pseudo-patterns that are statistically correlated rather than causally related.

[0036] This embodiment employs a more restrictive structure, using a causal graph as a template. Causal edges in the graph are directly mapped to information connections allowed within the corresponding time step in the LSTM. If the causal graph contains a directed edge from node A to node B, then in the LSTM model, when processing data at the current time step, the hidden representation of variable A should directly contribute to the predicted value of variable B. This is achieved by treating the connection weights corresponding to the causal edges as free parameters in the weight matrices of the LSTM's input, forget, and output gates, while forcing the connection weights between variable pairs not present in the causal graph to zero, or by applying significant compression in the penalty term described later. This ensures that the information propagation path within the model does not violate the dependencies revealed by the causal graph; that is, only dependent variables identified in the causal graph can influence effect variables.

[0037] Apply L1 sparsity penalty to non-causal edges: When causal discovery algorithms process limited sample data, the resulting causal graph may contain missing or redundant edges. If connections outside the causal graph are completely prohibited, the model will lose its ability to learn this important path once the causal graph misses a real causal edge. Conversely, if no restrictions are imposed, the model may overfit to noise through a large number of irrelevant connections.

[0038] This embodiment employs a compromise: the weights corresponding to edges present in the causal graph can be freely updated in the model without additional penalties; the weights corresponding to edges not present in the causal graph are not completely deleted from the model, but rather an L1 regularization penalty term is added to the loss function. The L1 penalty causes the absolute values ​​of these connection weights to tend towards zero, but does not force them to instantly become zero. If a missing edge in the causal graph actually has strong predictive power in the data, the model can retain it at a certain regularization cost; if it is merely noise, regularization will gradually compress it to near zero. The specific loss function is as follows: Total Loss = Prediction Error + λ × Σ (Absolute value of weights corresponding to non-causal edges), where λ is a hyperparameter, ranging from 0.001 to 0.01. The prediction error typically uses cross-entropy loss. Through the penalty term, the information flow direction of the model is guided to align with the causal graph, while preserving the possibility of discovering new causal relationships from the data.

[0039] Xavier Uniform Distribution Initialization of Model Weights: The initialization of deep neural networks significantly impacts training convergence speed and solution quality. This embodiment employs the Xavier uniform distribution initialization method. For each weight matrix in the LSTM, the parameters are uniformly sampled within an interval jointly determined by the input and output dimensions. Specifically, the lower bound of the interval is negative (the reciprocal of the square root of the sum of the input and output dimensions multiplied by a scaling factor), and the upper bound is the corresponding positive value. This initialization ensures that the input and output variances of each layer are similar, avoiding gradient vanishing or exploding in deep networks. In a standard LSTM, different input features contribute equally to the hidden state update. The causal strength coefficients provided by the causal graph are used to initially set the attention level for each input feature. At each time step, when calculating the impact of the current input on the forget gate, input gate, and output gate, the model multiplies each feature in the input vector by a coefficient. The initial values ​​of these coefficients are the corresponding causal strength coefficients. The strength coefficient of each causal edge has already been obtained in the preceding causal learning, and its value is typically between 0 and 1. For example, if the causal graph shows that the causal strength coefficient of glycated hemoglobin on eGFR is 0.8, then the model will assign an initial attention weight of 0.8 to the input feature glycated hemoglobin during initialization, while assigning smaller initial weights to features with lower causal strength. From the first round of training, the model focuses more on features that play a core role in the causal graph, reducing the large amount of exploratory computation required in the early stages. Furthermore, the attention weights can still be adjusted through backpropagation during training and are not fixed due to the initial settings. The causal relationship diagram transforms the analysis results into prior constraints and initial parameters for the model, achieving a balance between medical knowledge and data-driven approaches, and avoiding counterintuitive associations that might arise from pure black-box learning. The entire process is completed locally, without involving the uploading of patient data, thus complying with privacy compliance requirements.

[0040] Step S2: Perform cross-party alignment comparison on the local causal relationship graphs of each participant, and establish an interaction relationship network among the participants based on the comparison results.

[0041] Furthermore, cross-party alignment and comparison are performed on the local causal relationship graphs of each participant. Based on the comparison results, an interaction network among the participants is established. This includes: each participant converting its local causal relationship graph into a standardized directed acyclic graph description file. The description file contains only a set of node identifiers, a set of directed edges, and a causal strength coefficient and confidence score attached to each directed edge. It does not contain any original sample data or patient-level records. Differential privacy noise is added to the edge attributes before uploading to the central server. The central server maps each graph node to a unified ontology dictionary and calculates the interaction relationship between any two participants. The structural similarity and directional consistency among the causal relationship graphs are weighted and fused to obtain a total causal consistency score. Based on the distribution of all consistency scores, an adaptive threshold is used to divide the participating parties into three levels: strongly trustworthy, weakly trustworthy, and untrustworthy. Each participating party is a node, and the total causal consistency score is used as the edge weight. Edges corresponding to strongly trustworthy and weakly trustworthy pairs are retained, while edges corresponding to untrustworthy pairs are deleted. A sparse weighted interaction relationship network is constructed. The interaction relationship network is used to identify the missing, overlapping, or contradictory causal edges between participating parties, providing adaptive aggregation weights for participating parties in subsequent federated learning.

[0042] Specifically, after each participant completes the construction of the local causal relationship diagram, these diagrams need to be collected on the central server for comparison to identify the similarities and differences in the causal knowledge of each party, providing a basis for the subsequent overlapping utilization of samples, knowledge complementarity, and data missing supplementation in federated learning. Each hospital first converts the locally generated causal relationship diagram into a description file in a standard format. The file only contains node identifiers, directed edges, and the causal strength coefficients and confidence scores attached to each edge, without including any original test values, visit dates, or population statistics of patients. To prevent third parties from reverse-inferring distribution characteristics from the causal strength coefficients, each hospital adds Laplace-distributed random noise to the causal strength coefficients and confidence scores of each edge before uploading. The noise scale is controlled by the privacy budget parameter ε, usually taking values from 0.1 to 1.0, so that the noisy values can still be used to compare the similarity of causal structures but cannot restore the original exact values. After adding noise, each participant uploads the description file to the central server. After receiving the causal relationship diagrams of all parties, the central server first performs node mapping and unified alignment: through a pre-stored federated shared ontology dictionary, different variable names used by different hospitals, such as HbA1c, glycated hemoglobin, and A1C, are unified and mapped to the same identifier, ensuring that all nodes representing the same clinical meaning obtain consistent identification. Then, the server calculates the total causal consistency score between any two hospitals A and B. The score is obtained by weighting two parts: structural similarity and direction consistency. The structural similarity uses the directed acyclic graph edit distance algorithm to normalize the minimum number of node additions and deletions, edge additions and deletions, and edge reversals required to convert graph A into graph B to a value between 0 and 1, representing the degree of overlap of the two graph skeletons; the direction consistency counts the proportion of edges with the same direction among the total number of common edges for the same variable pairs that coexist in the two graphs, also obtaining a score between 0 and 1. Adding the structural similarity and direction consistency according to the preset weights gives the total consistency score. The higher the score, the closer the causal knowledge of the two hospitals. After the server calculates the total consistency scores between all pairs of participants, it obtains the score distribution, automatically calculates the median M and the lower quartile Q1, and sets two adaptive thresholds based on this. Among them, the first threshold T1 = M + 0.5×(M - Q1), and the second threshold T2 = Q1 - 0.5×(M - Q1). Comparing the total consistency score of each pair of participants with these two thresholds, they are divided into three categories: those with a total score ≥ T1 are strong-trust peers, those between T2 and T1 are weak-trust peers, and those with a total score < T2 are untrustworthy peers. Then, taking each hospital as a node and using the total consistency score as the edge weight, initially fully connected, and then deleting all edges between untrustworthy peers, only retaining the edges corresponding to strong-trust and weak-trust, thus constructing a sparsified weighted interaction network. In the weighted interaction network, the weight of the edge is the total consistency score, and the weighted degree centrality of each node is equal to the sum of the weights of all edges connected to it, which is used to measure the trustworthiness of this participant in the alliance.

[0043] The interactive relationship network is subsequently used to guide federated learning. Participants corresponding to strongly trusted edges reinforce each other during aggregation, while participants corresponding to weakly trusted edges selectively merge consensus components and attenuate conflicts. Untrusted pairs are disconnected to avoid gradient interference. Simultaneously, the server can analyze the presence of each causal edge among the participants in the network to identify consensus edges recognized by most hospitals, potential supplementary knowledge edges unique to a few hospitals, and edges missing by some hospitals. This provides a basis for subsequent reinforcement of beneficial knowledge and cross-party supplementation of missing knowledge. Throughout the process, the central server only accesses the causal relationship graph structure without individual patient information and never obtains any original medical record data, fully complying with data privacy protection requirements. Furthermore, those skilled in the art can implement cross-party alignment and comparison and construct the interactive relationship network based on the description.

[0044] Step S3: Based on the interaction relationship network, perform federated learning on the local prediction model parameters of each participant to obtain federated learning model parameters, and perform multi-dimensional evaluation and interpretation of the federated learning model parameters, and label the reinforcement regions.

[0045] Furthermore, federated learning is performed on the local prediction model parameters of each participant based on the interaction network to obtain federated learning model parameters. These parameters are then evaluated and interpreted in multiple dimensions, and reinforcement regions are labeled. This includes: identifying the favorable and missing causal knowledge of each participant relative to the whole based on the total causal consistency score and edge weights among the participants in the interaction network; and adopting a hierarchical federated learning strategy based on the identification results of favorable and missing causal knowledge. Specifically, for participants with favorable knowledge, the gradient contribution weight is increased during aggregation to strengthen consensus; for participants with missing knowledge, a regularization term is added during training to shift the update direction of the model parameters corresponding to the missing edges towards the consensus direction of the majority of participants, for cross-party supplementation of missing knowledge. Untrusted participants with low causal consistency with most other participants in the network are temporarily excluded from the main aggregation process. The central server calculates the adaptive aggregation weight of each participant based on the weighted degree centrality of each participant in the interactive relationship network and the results of the favorable causal knowledge identification. It then performs a weighted average aggregation of the local model parameter updates uploaded by each participant and iteratively updates the parameters of the federated learning model. After obtaining the parameters of the federated learning model, the central server samples a representative set of samples without patient identification from each participant. Using a combination of feature attribution and counterfactual explanation, it calculates the contribution of each feature to the prediction result and the minimum feature modification path required to reverse the prediction. Features with both contribution and frequency exceeding the adaptive threshold and time windows are merged and marked as enhanced regions after confidence verification.

[0046] Specifically, after the interactive relationship network is constructed, the central server uses it to guide each participant in federated learning and interprets the aggregated model from multiple dimensions, identifying feature regions that need to be strengthened. The central server then creates a knowledge profile for each participant based on the total causal consistency score and edge weights between each pair of participants in the interactive relationship network. Specifically, for a given participant, the server compares it with the causal relationship graphs of all strongly trusted partners: if a causal edge exists in the participant's graph and also exists in the graphs of most strongly trusted partners with the same direction, this edge is considered favorable causal knowledge for the participant, indicating that its local data reflects general patterns well; conversely, if a causal edge is missing in the participant's graph but exists in the graphs of most strongly trusted partners with the same direction, this edge is considered missing causal knowledge for the participant, indicating that its local data may be incomplete or have insufficient sample size, requiring supplementation from other participants.

[0047] Based on the recognition results, the server adopts a hierarchical federated learning strategy to distinguish the roles of different participants. For participants with a large amount of beneficial knowledge, that is, whose causal relationship graph is highly consistent with their strong and trustworthy peers, they are given higher weights in gradient update during federated aggregation, so that their contributed knowledge is strengthened in the global model.

[0048] For participants with significant knowledge gaps, learning of missing edges is not prohibited during local training. Instead, a regularization term is added to the loss function. The design principle of the regularization term is to slightly shift the update direction of the model parameters corresponding to the missing edges towards the consensus direction of most participants. In practice, an additional squared error term can be added to the objective function of local training to penalize the deviation between the participant's model parameters on these missing edges and the global consensus value. The magnitude of the shift is determined by the participant's weighted degree centrality in the interaction network. The closer the connection with strong trustworthy peers, the more inclined the participant is to accept the guidance of the consensus direction. In this way, missing knowledge can be supplemented cross-party without directly accessing the original data of other participants.

[0049] For untrusted participants in the interaction network whose causal consistency with most participants is too low, they will not be included in the main aggregation process for the time being. The data of these participants may not be able to form an effective consensus with other institutions due to data entry errors, extremely special patient populations, or too small sample sizes. Excluding them from the main aggregation can avoid gradient conflicts from interfering with the global model. These participants can be used as an independent validation set to evaluate the generalization ability of the global model. They will be re-added after their causal relationship graph gradually moves closer to the mainstream consensus.

[0050] Then, the central server calculates the adaptive aggregation weight based on the weighted centrality and the proportion of advantageous knowledge of each participant in the interaction network. The participant with richer advantageous knowledge has a larger weight. The server performs a weighted average of the local model parameter updates uploaded by each participant and iteratively updates the federated learning model parameters.

[0051] After obtaining the model parameters, the server needs to identify which features and time windows the model relies on for judgment. A batch of representative samples (with patient identifiers removed) is sampled from each participant, and cross-validation is performed using two methods: feature attribution (e.g., SHAP value, calculating the average contribution of each feature to the prediction result) and counterfactual interpretation (identifying the minimum feature modification path required to reverse the prediction result, and counting the frequency of each feature in the path). The features with the top 20% average contribution and the features with the top 75% frequency in the counterfactual path are merged. The merged candidate regions are then subjected to confidence verification, testing the stability of the prediction probability under small perturbations. Regions that pass the verification are marked as reinforcement regions. These reinforcement regions reflect the knowledge points that the model currently relies on most and that most need reinforcement or adjustment, and are sent back to each participant for further targeted optimization. Throughout this process, the central server only accesses the causal relationship graph structure and privacy-free sample features, without involving any original medical record data.

[0052] Step S4: Return the reinforcement region to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain the enhanced model parameters.

[0053] Furthermore, the reinforcement regions are returned to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain enhanced model parameters. The reinforcement regions also include strong and weak regions. Strong regions are causal edges with high model prediction confidence, stable feature contributions, and consistent consensus across participants, along with their corresponding time windows. Weak regions include uncertain regions with prediction errors exceeding a preset threshold, contradictory regions with divergent causal edge directions, and incomplete regions with local data missing rates exceeding a preset threshold. Each participant executes a differentiated training strategy based on the reinforcement regions, wherein, based on the strong regions, the corresponding regions are maintained during local training. The model parameters are kept unchanged or subject to minor regularization constraints to prevent knowledge forgetting; differentiated data augmentation and parameter update strategies are adopted for weak regions; for uncertain regions, the robustness of the model is improved by synthesizing boundary samples or adversarial training; for contradictory regions, local parameter updates are guided by the consensus direction of strongly trusted peers in the interaction network; for incomplete regions, evidence is supplemented by missing value imputation or multiple imputation methods based on causal relationships, and the sampling weight of the corresponding region features is increased during model training; after completing local targeted training based on differentiated training strategies, the enhanced model parameters are obtained, and the corresponding parameters are uploaded to the central server for the next round of federated learning.

[0054] Specifically, the central server distributes the labeled reinforcement regions to each participant. In practical applications, these reinforcement regions are further subdivided into two types: strong regions and weak regions. Strong regions refer to causal edges with high model prediction confidence, stable feature contributions, and consensus across participants, along with their corresponding time windows—areas where the model has learned well. Weak regions include: uncertain areas where prediction errors exceed a set threshold (i.e., the model is unsure where to proceed); contradictory areas where the direction of causal edges differs among participants; and incomplete areas with excessively high local data missing rates (i.e., some key indicators are not fully recorded). After receiving the reinforcement regions, each participant executes differentiated training strategies based on their local data. For strong regions, where the model has already performed well, corresponding parameters are kept unchanged during local training, or a small regularization constraint is applied to prevent knowledge forgetting in subsequent updates.

[0055] For weak regions, different enhancement methods are adopted for different types. For uncertain regions, i.e., where the model hesitates in prediction and the probability value hovers around the threshold, data augmentation is used to improve the situation. Specifically, this involves synthesizing boundary samples, such as selecting samples near the positive and negative sample boundaries in the original feature space for interpolation to generate new samples, or using adversarial training, adding small perturbations to the input features to generate adversarial samples and adding them to the training set, forcing the model to become insensitive to these perturbations, thereby improving robustness. For contradictory regions, i.e., where the causal relationship graphs of different participants conflict, each participant refers to the consensus direction of strongly trusted peers in the interaction network and adds a guiding term during local training to fine-tune their model parameters in the direction accepted by the majority, rather than excessively adhering to contradictory local conclusions.

[0056] For incomplete areas, that is, records of certain key indicators that are missing locally, a missing value imputation method based on causal relationship is adopted. For example, the missing values ​​are estimated by using other relevant variables in the local data through multiple imputation chain equations. During imputation, only the data of the participating party is used, without borrowing the original information of other hospitals. After the imputation is completed, the sampling weight of these imputed features is increased in the model training, so that the model pays more attention to the originally missing information.

[0057] After each participant completes local targeted training according to the above differentiated strategies, they obtain enhanced model parameters and upload these parameters back to the central server as the starting point for the next round of federated learning. In this way, in each iteration, the model retains its learned strong knowledge while specifically addressing weaknesses such as inaccurate predictions, conflicting directions, and missing data. After multiple iterations, the overall performance of the model gradually improves, maintaining the stability of cross-participant knowledge fusion, leveraging strengths and mitigating weaknesses, compensating for individual shortcomings, and learning from the strengths of other participants. Furthermore, the entire targeted fine-tuning process is executed entirely locally on each participant's premises, without involving the cross-party transmission of original data.

[0058] Step S5: Use the enhanced model parameters as the initial parameters for the next round of federated learning, and repeat steps S3 to S4 until the preset convergence condition is met to obtain the global model parameters.

[0059] Specifically, after each participant completes local targeted fine-tuning and uploads the enhanced model parameters, the central server uses these parameters as the initial parameters for the next round of federated learning. Then, steps S3 (federated learning and interpretation annotation) and S4 (targeted fine-tuning) are repeated. In each iteration, the model first absorbs the consensus knowledge of each participant through federated aggregation, then identifies the reinforcement regions of the current version through interpretation annotation for subsequent training targeted feedback guidance, and finally repairs weaknesses and strengthens advantages through targeted fine-tuning. After multiple iterations, the model parameters gradually stabilize, and the prediction performance tends to saturate. When the preset convergence condition is met, the iteration terminates. The convergence condition can be that the change in global model parameters is less than a preset threshold for multiple consecutive rounds, or that the AUC, precision, or recall on the independent validation set no longer improves for multiple consecutive rounds, or that the average total consistency score among the participants in the interaction network tends to be constant, or that the preset maximum number of iterations is reached. Once the convergence condition is met, the central server outputs the current global model parameters as the final optimal model.

[0060] Step S6: Deploy the risk assessment scenario model using the global model parameters, conduct a risk assessment on the health status of the target individual, and obtain the risk assessment results.

[0061] Specifically, the final global model parameters are deployed to specific risk assessment scenarios. For example, a doctor-side application can be deployed, where longitudinal follow-up data of the target individual is input, such as time-series indicators like glycated hemoglobin, blood pressure, and eGFR. The risk assessment scenario model automatically outputs the probability of developing diabetic nephropathy within a certain time window in the future, such as within 3 years or 5 years, and can simultaneously display the main features and time windows on which the corresponding prediction is based, helping doctors understand the basis for their judgment. Alternatively, it can be deployed to a patient-side mini-program, where patients manually enter simple indicators such as blood glucose and blood pressure to obtain their own risk trend chart and personalized prevention suggestions. The model parameters are locked before deployment, but in actual operation, a new round of incremental closed-loop optimization can be triggered by periodically collecting new data, allowing the model to continuously adapt to changes in the population.

[0062] This embodiment begins with local causal learning by each participant, proceeding through causal relationship graph alignment, interaction network construction, hierarchical federated learning, multi-dimensional interpretive annotation, targeted fine-tuning, closed-loop iteration, and finally deployment and application. This constitutes a complete and implementable diabetes risk assessment solution, protecting patient data privacy throughout the process while improving model accuracy, generalization ability, and interpretability. It effectively solves the problems of low model accuracy and poor generalization ability caused by insufficient sample size and knowledge conflicts in single centers. Furthermore, by leveraging interpretable annotation and targeted reinforcement, it overcomes the shortcomings of traditional black-box models, which lack interpretability and continuous optimization mechanisms. This achieves a significant breakthrough in diabetes risk assessment performance across heterogeneous multi-center data while adhering to privacy compliance, effectively improving the technical effectiveness of machine learning in diabetes risk assessment.

[0063] Example 2: like Figure 2 As shown, based on the same inventive concept as the machine learning-based intelligent assessment method for diabetes risk provided in Embodiment 1, this embodiment of the invention also provides a machine learning-based intelligent assessment system for diabetes risk, the system comprising: Local training module 1 is used to perform step S1: Each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, obtains local causal model parameters, and initializes local prediction model parameters.

[0064] Participant causal comparison module 2 is used to perform step S2: perform cross-party alignment comparison on the local causal relationship graph of each participant, and establish an interaction relationship network between participants based on the comparison results.

[0065] Federated learning module 3 is used to perform step S3: based on the interaction relationship network, perform federated learning on the local prediction model parameters of each participant to obtain federated learning model parameters, and perform multi-dimensional evaluation and interpretation of the federated learning model parameters, and mark the reinforcement regions.

[0066] Reinforcement learning module 4 is used to execute step S4: return the reinforcement region to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain the enhanced model parameters.

[0067] The federated iteration module 5 is used to execute step S5: using the enhanced model parameters as the initial parameters for the next round of federated learning, repeating steps S3 to S4 until the preset convergence condition is met, and obtaining the global model parameters.

[0068] Risk assessment module 6 is used to perform step S6: deploy a risk assessment scenario model using the global model parameters, assess the health status of the target individual, and obtain the risk assessment result.

[0069] Furthermore, the local training module 1 is also used to perform the following steps: The local diabetes sample data is expanded by time slices. Cross-layer directed edges are constructed for variables between adjacent time slices to represent time-lag causal relationships, and same-layer directed edges are constructed for variables within the same time slice to represent instantaneous causal relationships.

[0070] By minimizing the loss function with a loop constraint penalty term, the causal edge direction and strength of both time-lag causal relationships and immediate causal relationships are analyzed simultaneously, and a local sample causal relationship graph containing the time dimension is output. The parameters of the local causal model include the direction identifier, confidence score and causal strength coefficient of each directed edge in the local sample causal relationship graph.

[0071] Furthermore, the local sample causal relationship graph is a directed acyclic graph structure, where nodes are used to identify the types of variables involved in causal inference, and directed edges are used to characterize the data dependencies and causal directions between variables. The nodes include: static feature nodes, time-series observation nodes, and fluctuation-derived nodes. Static feature nodes characterize attribute variables that do not change or change slowly over time during the patient's baseline period, including age, gender, duration of diabetes, smoking history, and family history. Time-series observation nodes characterize longitudinal follow-up vital signs variables collected along a uniform time axis. Fluctuation-derived nodes characterize secondary derived variables obtained by calculating the coefficient of variation, mean true fluctuation range, and continuous slope changes of selected indicators in the time-series observation nodes, and are included as independent nodes in the causal relationship graph.

[0072] Furthermore, the local training module 1 is also used to perform the following steps: Using the local sample causal relationship graph as prior knowledge constraint, a time-series prediction model based on a long short-term memory network is constructed.

[0073] The causal edges determined in the causal graph are used as the mandatory information transmission paths between time steps in the time series prediction model structure. L1 sparsity penalty is applied to the model loss function for variable connections that do not exist in the causal graph, so that the information flow direction of the time series prediction model is strictly aligned with the causal graph.

[0074] The initialization process uses a Xavier uniform distribution to randomly initialize the model weights, and the causal strength coefficients are used as the initial attention weights for the corresponding connection paths.

[0075] Furthermore, the participant causality comparison module 2 is also used to perform the following steps: Each participating party converts its local causal relationship graph into a standardized directed acyclic graph description file. The description file contains only a set of node identifiers, a set of directed edges, and a causal strength coefficient and confidence score attached to each directed edge. It does not contain any original sample data or patient-level records. Differential privacy noise is added to the edge attributes before the file is uploaded to the central server.

[0076] The central server maps each graph node to a unified ontology dictionary, calculates the structural similarity and directional consistency between the causal relationship graphs of any two participants, and then performs a weighted fusion to obtain the total causal consistency score.

[0077] Based on the distribution of all consistency scores, an adaptive threshold is used to divide the participants into three levels: strongly trustworthy, weakly trustworthy, and untrustworthy. Each participant is a node, and the total causal consistency score is used as the edge weight. Edges corresponding to strongly trustworthy and weakly trustworthy pairs are retained, while edges corresponding to untrustworthy pairs are deleted. A sparse weighted interaction relationship network is constructed. The interaction relationship network is used to identify the missing, overlapping, or contradictory causal edges between participants, providing adaptive aggregation weights for participants in subsequent federated learning.

[0078] Furthermore, Federated Learning Module 3 is also used to perform the following steps: Based on the total causal consistency score and edge weight among the participants in the interactive relationship network, identify the favorable causal knowledge and missing causal knowledge of each participant relative to the whole.

[0079] Based on the identification results of favorable and missing causal knowledge, a hierarchical federated learning strategy is adopted. For participants with favorable knowledge, the gradient contribution weight is increased during aggregation to strengthen consensus. For participants with missing knowledge, a regularization term is added during training to shift the update direction of the model parameters corresponding to the missing edge towards the consensus direction of the majority of participants, which is used to supplement the missing knowledge across parties. Untrusted participants with low causal consistency with the majority of participants in the interaction network are temporarily not included in the main aggregation process.

[0080] The central server calculates the adaptive aggregation weight of each participant based on the weighted degree centrality of each participant in the interaction network and the results of the identification of favorable causal knowledge. It then performs a weighted average aggregation of the local model parameter updates uploaded by each participant and iteratively updates the parameters of the federated learning model.

[0081] After obtaining the parameters of the federated learning model, the central server samples a representative set of unidentified samples from each participant. It uses a combination of feature attribution and counterfactual explanation to calculate the contribution of each feature to the prediction result and the minimum feature modification path required to reverse the prediction. Features with both contribution and frequency exceeding the adaptive threshold and time windows are merged and marked as enhanced regions after confidence verification.

[0082] Furthermore, reinforcement learning module 4 is also used to perform the following steps: The reinforcement region includes strong regions and weak regions. The strong regions are causal edges and corresponding time windows with high model prediction confidence, stable feature contributions, and consistent consensus among participants. The weak regions include uncertain regions where the prediction error exceeds a preset threshold, contradictory regions where the causal edge directions are divergent, and incomplete regions where the local data missing rate is higher than a preset threshold.

[0083] Each participant implements a differentiated training strategy based on the strengthened regions. Specifically, for strong regions, the model parameters corresponding to the region are kept unchanged or subject to minor regularization constraints during local training to prevent knowledge forgetting. For weak regions, differentiated data augmentation and parameter update strategies are adopted. For uncertain regions, the robustness of the model is improved by synthesizing boundary samples or adversarial training. For contradictory regions, local parameter updates are guided by the consensus direction of strongly trusted peers in the interaction network. For incomplete regions, evidence is supplemented by missing value imputation or multiple imputation methods based on causal relationships, and the sampling weight of the corresponding region features is increased during model training.

[0084] After completing local targeted training based on the differentiated training strategy, the enhanced model parameters are obtained, and the corresponding parameters are uploaded to the central server for the next round of federated learning.

[0085] Through the foregoing detailed description of a machine learning-based intelligent assessment method for diabetes risk, those skilled in the art can clearly understand the machine learning-based intelligent assessment system for diabetes risk in this embodiment. As the system disclosed in the embodiment corresponds to the method disclosed in the embodiment, the description is relatively simple, and relevant parts can be referred to the method section.

[0086] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A machine learning-based intelligent assessment method for diabetes risk, characterized in that, include: Step S1: Each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, obtains local causal model parameters, and initializes local prediction model parameters; Step S2: Perform cross-party alignment comparison on the local causal relationship graphs of each participant, and establish an interaction relationship network among the participants based on the comparison results; Step S3: Based on the interaction network, perform federated learning on the local prediction model parameters of each participant to obtain federated learning model parameters, and perform multi-dimensional evaluation and interpretation of the federated learning model parameters, and label the reinforcement regions. Step S4: Return the reinforcement region to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain the enhanced model parameters; Step S5: Use the enhanced model parameters as the initial parameters for the next round of federated learning, and repeat steps S3 to S4 until the preset convergence condition is met to obtain the global model parameters. Step S6: Deploy the risk assessment scenario model using the global model parameters, conduct a risk assessment on the health status of the target individual, and obtain the risk assessment results.

2. The intelligent diabetes risk assessment method based on machine learning as described in claim 1, characterized in that, Each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, and obtains local causal model parameters, including: The local diabetes sample data is expanded by time slices. Cross-layer directed edges are constructed for variables between adjacent time slices to represent time-lag causal relationships, and same-layer directed edges are constructed for variables within the same time slice to represent instantaneous causal relationships. By minimizing the loss function with a loop constraint penalty term, the causal edge direction and strength of both time-lag causal relationships and immediate causal relationships are analyzed simultaneously, and a local sample causal relationship graph containing the time dimension is output. The parameters of the local causal model include the direction identifier, confidence score and causal strength coefficient of each directed edge in the local sample causal relationship graph.

3. The intelligent diabetes risk assessment method based on machine learning as described in claim 2, characterized in that, The local sample causal relationship graph is a directed acyclic graph structure, where nodes are used to identify the types of variables involved in causal inference, and directed edges are used to represent the data dependencies and causal directions between variables. The nodes include: static feature nodes, time-series observation nodes, and fluctuation-derived nodes. Static feature nodes represent attribute variables that do not change or change slowly over time during the patient's baseline period, including age, gender, duration of diabetes, smoking history, and family history. Time-series observation nodes represent longitudinal follow-up vital signs variables collected along a uniform time axis. Fluctuation-derived nodes represent secondary derived variables obtained by calculating the coefficient of variation, mean true range of fluctuation, and continuous slope changes of selected indicators in the time-series observation nodes, and are included as independent nodes in the causal relationship graph.

4. The intelligent diabetes risk assessment method based on machine learning as described in claim 2, characterized in that, Initialize the local prediction model parameters, including: Using the local sample causal relationship graph as prior knowledge constraint, a time series prediction model based on a long short-term memory network is constructed. The causal edges determined in the causal graph are used as the mandatory information transmission paths between time steps in the time series prediction model structure. L1 sparsity penalty is applied to the model loss function for variable connections that do not exist in the causal graph, so that the information flow direction of the time series prediction model is strictly aligned with the causal graph. The initialization process uses a Xavier uniform distribution to randomly initialize the model weights, and the causal strength coefficients are used as the initial attention weights for the corresponding connection paths.

5. The intelligent diabetes risk assessment method based on machine learning as described in claim 1, characterized in that, Cross-party alignment and comparison are performed on the local causal relationship graphs of each participant, and an interaction network among the participants is established based on the comparison results, including: Each participating party converts its local causal relationship graph into a standardized directed acyclic graph description file. The description file contains only a set of node identifiers, a set of directed edges, and a causal strength coefficient and confidence score attached to each directed edge. It does not contain any original sample data or patient-level records. Differential privacy noise is added to the edge attributes before the file is uploaded to the central server. The central server maps each graph node to a unified ontology dictionary, calculates the structural similarity and directional consistency between the causal relationship graphs of any two participants, and weighted fusion to obtain the total causal consistency score. Based on the distribution of all consistency scores, an adaptive threshold is used to divide the participants into three levels: strongly trustworthy, weakly trustworthy, and untrustworthy. Each participant is a node, and the total causal consistency score is used as the edge weight. Edges corresponding to strongly trustworthy and weakly trustworthy pairs are retained, while edges corresponding to untrustworthy pairs are deleted. A sparse weighted interaction relationship network is constructed. The interaction relationship network is used to identify the missing, overlapping, or contradictory causal edges between participants, providing adaptive aggregation weights for participants in subsequent federated learning.

6. The intelligent diabetes risk assessment method based on machine learning as described in claim 5, characterized in that, Federated learning is performed on the local prediction model parameters of each participant based on the aforementioned interaction network to obtain federated learning model parameters. These parameters are then evaluated and interpreted in multiple dimensions, and reinforcement regions are labeled, including: Based on the total causal consistency score and edge weight among the participants in the interactive relationship network, identify the favorable causal knowledge and missing causal knowledge of each participant relative to the whole. Based on the identification results of favorable causal knowledge and missing causal knowledge, a hierarchical federated learning strategy is adopted. For participants with favorable knowledge, the gradient contribution weight is increased during aggregation to strengthen consensus. For participants with missing knowledge, a regularization term is added during training to shift the update direction of the model parameters corresponding to the missing edge towards the consensus direction of the majority of participants, which is used for cross-party supplementation of missing knowledge. Untrusted participants with low causal consistency with the majority of participants in the interaction relationship network are temporarily not included in the main aggregation process. The central server calculates the adaptive aggregation weight of each participant based on the weighted degree centrality of each participant in the interaction network and the result of the identification of favorable causal knowledge. It then performs weighted average aggregation of the local model parameter update amounts uploaded by each participant and iteratively updates the parameters of the federated learning model. After obtaining the parameters of the federated learning model, the central server samples a representative set of unidentified samples from each participant. It uses a combination of feature attribution and counterfactual explanation to calculate the contribution of each feature to the prediction result and the minimum feature modification path required to reverse the prediction. Features with both contribution and frequency exceeding the adaptive threshold and time windows are merged and marked as enhanced regions after confidence verification.

7. The intelligent diabetes risk assessment method based on machine learning as described in claim 1, characterized in that, The reinforcement region is returned to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain the enhanced model parameters, including: The reinforcement region includes strong regions and weak regions. The strong regions are causal edges with high model prediction confidence, stable feature contributions, and consistent consensus across participants, along with their corresponding time windows. The weak regions include uncertain regions where the prediction error exceeds a preset threshold, contradictory regions where the causal edge directions differ, and incomplete regions where the local data missing rate exceeds a preset threshold. Each participant implements a differentiated training strategy based on the strengthened regions. Specifically, for strong regions, the model parameters corresponding to the region are kept unchanged or subject to minor regularization constraints during local training to prevent knowledge forgetting. For weak regions, differentiated data augmentation and parameter update strategies are adopted. For uncertain regions, the robustness of the model is improved by synthesizing boundary samples or adversarial training. For contradictory regions, local parameter updates are guided by the consensus direction of strongly trusted peers in the interaction network. For incomplete regions, evidence is supplemented by missing value imputation based on causal relationships or multiple imputation methods, and the sampling weight of the corresponding region features is increased during model training. After completing local targeted training based on the differentiated training strategy, the enhanced model parameters are obtained, and the corresponding parameters are uploaded to the central server for the next round of federated learning.

8. A machine learning-based intelligent risk assessment system for diabetes, characterized in that, The system for implementing the machine learning-based intelligent assessment method for diabetes risk according to any one of claims 1-7, the system comprising: The local training module is used to perform step S1: Each participant performs causal learning based on local diabetes sample data, constructs a local sample causal relationship graph, obtains local causal model parameters, and initializes local prediction model parameters. The participant causal comparison module is used to perform step S2: perform cross-party alignment comparison on the local causal relationship graph of each participant, and establish an interaction relationship network between participants based on the comparison results; The federated learning module is used to perform step S3: based on the interaction network, perform federated learning on the local prediction model parameters of each participant to obtain federated learning model parameters, and perform multi-dimensional evaluation and interpretation of the federated learning model parameters, and label reinforcement regions. The reinforcement learning module is used to execute step S4: return the reinforcement region to each participant, and each participant performs targeted reinforcement learning on the reinforcement learning model based on local data to obtain the enhanced model parameters; The federated iteration module is used to execute step S5: using the enhanced model parameters as the initial parameters for the next round of federated learning, repeating steps S3 to S4 until the preset convergence condition is met, and obtaining the global model parameters. The risk assessment module is used to perform step S6: deploy a risk assessment scenario model using the global model parameters, assess the health status of the target individual, and obtain the risk assessment result.

Images