A diabetic nephropathy early risk stratification early warning system and method

Abstract

This invention relates to the field of medical and health information processing technology, specifically disclosing an early risk stratification and warning system and method for diabetic nephropathy. The system includes a data fusion preprocessing module, a multi-scale causal network construction module, a dynamic risk trajectory calculation module, and a stratified warning decision-making module, connected sequentially. By fusing multimodal data to construct an interpretable causal network and utilizing a network-constrained prediction model, it achieves dynamic trajectory prediction of future renal function indicators and comprehensive risk indices. Based on the prediction results, it triggers different levels of risk warnings and outputs decision support information, thereby achieving early, dynamic, and causally interpretable risk stratification and warning.

Description

Technical Field

[0001] This invention belongs to the field of medical and health information processing technology, specifically relating to an early risk stratification and early warning system and method for diabetic nephropathy. Background Technology

[0002] Diabetic nephropathy is one of the most serious microvascular complications of diabetes and a leading cause of end-stage renal disease. Early and accurate risk stratification and early warning are crucial issues in clinical diagnosis and public health management. The development of artificial intelligence and bioinformatics technologies has provided new technical pathways for integrating multi-dimensional data and mining early disease characteristics. Risk prediction models based on multi-omics data have become an important research direction in the field of medical and health information processing technology.

[0003] Constructing risk prediction models for diabetic nephropathy using biomarkers and clinical indicators is a current research hotspot. Existing technologies are typically based on single-dimensional data, such as clinical biochemical indicators or genomic data, and use machine learning algorithms to build classification or regression models, aiming to identify high-risk individuals and predict disease progression.

[0004] Current technologies primarily rely on clinical data after patients exhibit macroscopic imaging findings or significant decline in renal function. By this time, kidney damage is often irreversible, resulting in a delayed warning window. Existing models fail to establish a coherent causal chain from early biological aging events at the molecular level to subsequent changes in the macroscopic structure and function of organs. This leads to weak biological mechanisms for risk stratification and limited clinical interpretability and intervention guidance value of the predicted results.

[0005] Semantic gaps exist between data from different sources and at different scales. The lack of a unified analytical framework that can deeply integrate molecular aging characteristics with clinical phenotypic evolution information makes it difficult to achieve truly early, dynamic, and mechanism-driven risk warnings. Therefore, there is an urgent need to develop a system and method that can connect early molecular events with clinical endpoints and achieve high-precision dynamic risk stratification and early warning. Summary of the Invention

[0006] The purpose of this invention is to provide an early risk stratification and warning system and method for diabetic nephropathy, in order to solve the problems of existing technologies that rely on late clinical data, resulting in delayed warnings, and the lack of a coherent causal explanation chain connecting early events with macroscopic clinical phenotypic evolution.

[0007] This invention provides an early risk stratification and warning system for diabetic nephropathy, comprising: The data fusion preprocessing module is used to receive and process heterogeneous multimodal input data from different sources, including molecular omics data, longitudinal clinical time-series data, and static baseline data. A multi-scale causal network construction module, connected to the data fusion preprocessing module, is used to construct an interpretable causal network from the molecular scale to the organ function scale based on the multi-modal fusion dataset. The dynamic risk trajectory calculation module is connected to the multi-scale causal network construction module and is used to perform individualized dynamic risk calculation and trajectory prediction based on the constructed causal graph model. The hierarchical early warning and decision-making module is connected to the dynamic risk trajectory calculation module and is used to generate hierarchical early warning and decision support information based on the calculation results.

[0008] Preferably, the molecular omics data includes at least transcriptome sequencing data and epigenome methylation sequencing data; The longitudinal clinical time-series data include at least the urine albumin-to-creatinine ratio, estimated glomerular filtration rate, and glycated hemoglobin collected over time. The static baseline data includes at least the patient's age, duration of diabetes, and history of hypertension. The data fusion preprocessing module further performs data alignment and standardization operations, specifically by resampling and aligning the longitudinal clinical time-series data based on a unified time axis, and standardizing the molecular omics data and clinical time-series data to eliminate dimensional differences, thereby generating a time-aligned and scale-uniform multimodal fusion dataset.

[0009] Preferably, the multi-scale causal network construction module first extracts biological aging-related features from the molecular omics data, specifically by calculating the expression aging index of a specific gene set and the acceleration value of the whole-genome methylation aging clock; The multi-scale causal network construction module further employs a constraint-based causal discovery algorithm, using the biological aging-related characteristics as candidate dependent variables, the longitudinal clinical time-series data as candidate outcome variables, and introducing the static baseline data as conditional variables, to learn the causal structure under time-series constraints. During the causal structure learning process, a directional causal graph model is output. Nodes represent feature variables at different scales, and directed edges represent potential causal relationships between variables with temporal order. The weights of the edges are calculated in subsequent causal effect quantification steps.

[0010] Preferably, the dynamic risk trajectory calculation module includes a recurrent neural network prediction model with embedded causal priors; The input to the recurrent neural network prediction model is a multimodal fusion data slice arranged in chronological order, which has been processed by the data fusion preprocessing module. The network structure of the recurrent neural network prediction model is constrained by the causal graph model output by the multi-scale causal network construction module. Specifically, trainable parameterized connections are established only between neural network units corresponding to nodes with directed edge connections in the causal graph model. For connections that do not exist in the causal graph model, the corresponding neural network connection weights are fixed to 0. The recurrent neural network prediction model learns the specific parameterization and strength of the relationships represented by each directed edge in the causal graph model through training; During the inference phase, the dynamic risk trajectory calculation module takes the observation data of the preceding time points as input, performs multi-step forward extrapolation through the recurrent neural network prediction model, predicts the values ​​of key clinical indicators such as the urinary albumin-to-creatinine ratio and the estimated glomerular filtration rate at multiple future time points, and simultaneously calculates the comprehensive risk index. The comprehensive risk index is a weighted function of the deviation of the future predicted urinary albumin-to-creatinine ratio from the estimated glomerular filtration rate relative to its baseline value, and the uncertainty of the predicted trajectory.

[0011] Preferably, the hierarchical early warning decision module receives the future comprehensive risk index prediction trajectory and its uncertainty range output by the dynamic risk trajectory calculation module; The hierarchical early warning decision module has three preset risk thresholds, which correspond to low risk, medium risk and high risk states respectively. The hierarchical early warning decision module compares the predicted comprehensive risk index trajectory with these thresholds in real time. When the predicted trajectory enters the medium-risk threshold range at any future observation time, the hierarchical early warning decision module generates a first-level early warning signal and outputs an analysis report prompting attention to changes in molecular aging characteristics. When the predicted trajectory remains in the medium-risk range for more than two preset observation periods, or directly enters the high-risk threshold range, the hierarchical early warning decision module generates a secondary early warning signal and outputs a list of key driving factors derived from causal network backtracking analysis, as well as targeted intervention simulation suggestions.

[0012] Preferably, the constraint-based causal discovery algorithm employs a time-extended PC algorithm combined with a linear non-Gaussian acyclic model; the execution process of this causal discovery algorithm is as follows: Based on the time order constraint, variables at later time points are prohibited from being the cause of variables at earlier time points; By using the conditional independence test, given a partial set of conditional variables, we can gradually eliminate unconditional edges between variables. The retained edges are inferred for directionality, and the causal effect strength of the edges is quantitatively estimated using a linear non-Gaussian acyclic model, thus obtaining a weighted causal graph.

[0013] Preferably, the recurrent neural network prediction model with embedded causal priors adopts a gated recurrent unit network structure; In the gated recurrent unit network structure, each feature variable in the causal graph model corresponds to a specific hidden state unit; The connection matrix between network layers is forced to be a sparse trainable matrix consistent with the adjacency matrix structure of the causal graph model. In the causal graph model, the matrix elements corresponding to directed edge connections are trainable parameters, otherwise they are constant zero values. The loss function of the recurrent neural network prediction model consists of two parts: the first part is the mean square error between the predicted clinical indicator value and the true value, and the second part is the uncertainty estimate represented by the variance of the predicted trajectory.

[0014] Preferably, the process for generating intervention measure simulation suggestions in the hierarchical early warning decision-making module is as follows: When the Level II warning signal is triggered, the system identifies the key causal path pointing to the high-risk state based on the causal graph model obtained by the multi-scale causal network construction module. The system performs counterfactual reasoning on the causal graph model to simulate the impact on the predicted trajectory of downstream clinical indicators if the upstream molecular aging characteristic variables in the list of key driving factors undergo positive changes. The simulation results quantitatively predict the possible magnitude and time of decline in the risk index under different intervention intensities, thereby forming simulation recommendations for intervention measures with priority ranking.

[0015] Preferably, the system further includes a model continuous learning and updating module; The model continuous learning and updating module is connected to the data fusion preprocessing module and the multi-scale causal network construction module, and is used to continuously integrate newly generated patient data after system deployment; The model continuous learning and updating module is started periodically, using new data to incrementally learn the causal discovery algorithm to update the causal graph structure; and using new data to incrementally train the prediction model in the dynamic risk trajectory calculation module to optimize the model parameters.

[0016] This invention also provides a method for early risk stratification and early warning of diabetic nephropathy, executed by the aforementioned system, comprising the following steps: Data fusion preprocessing steps: Collect molecular omics data, longitudinal clinical time-series data, and static baseline data of the target individuals; align the longitudinal clinical time-series data based on a unified time axis, and standardize all data to generate a multimodal fusion dataset; The steps of the multi-scale causal network construction module are as follows: calculate biological aging characteristics from the molecular omics data; use the aging characteristics, clinical time-series indicators and baseline data as variables, apply a time-constrained causal discovery algorithm to learn and construct a weighted directed causal graph model from the molecular scale to the organ function scale; Dynamic risk trajectory calculation steps: The individual's historical multimodal fusion data is input into a causal prior recurrent neural network prediction model according to the time series; the network structure of the recurrent neural network prediction model is constrained by the causal graph model; through model deduction, the key clinical indicator values ​​and comprehensive risk index trajectory of the individual at multiple future time points are predicted; The tiered early warning decision-making process involves comparing the predicted trajectory of the future comprehensive risk index with preset low, medium, and high risk thresholds; generating risk warning signals of different levels based on the threshold range and duration reached by the predicted trajectory, and outputting corresponding causal analysis reports and intervention simulation suggestions.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention fundamentally alters the mechanistic basis of risk warning by constructing a multi-scale causal network connecting early molecular aging events with macroscopic clinical phenotypes. The system not only predicts risk but also reveals potential causal pathways from abnormalities in specific molecular characteristics to deterioration of renal function indicators, providing clear biological targets and mechanistic explanations for clinical intervention and enhancing the clinical operability and scientific value of the warning results.

[0018] 2. This invention employs a causal-constrained dynamic trajectory prediction model, achieving true early and dynamic warning. By fusing longitudinal multimodal data and utilizing a causal prior constraint model structure, the system can more stably and accurately predict the future trajectory of the disease based on early molecular aging signals before macroscopic kidney damage occurs, thus advancing the risk warning window and gaining valuable time for early intervention.

[0019] 3. The hierarchical early warning decision-making mechanism of this invention, based on predicted trajectories and uncertainty quantification, achieves refined and intelligent risk management. The system dynamically triggers different responses according to the risk level and performs counterfactual reasoning from the constructed causal network to generate personalized intervention simulation suggestions. This upgrades early warning from simple risk alerts to a decision support system that includes causal attribution and strategy simulation, thereby improving the quality and efficiency of overall medical decision-making. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a framework diagram of the execution process of the causal discovery algorithm in this invention; Figure 3This is a schematic diagram of the process of generating simulated intervention suggestions in the hierarchical early warning decision-making module of this invention; Figure 4 This is a schematic diagram of the early risk stratification and early warning method for diabetic nephropathy in this invention. Detailed Implementation

[0021] Example 1: The overall architecture of the early risk stratification and early warning system for diabetic nephropathy described in this invention is shown in the attached figure. Figure 1 As shown, the system consists of a data fusion and preprocessing module, a multi-scale causal network construction module, a dynamic risk trajectory calculation module, and a hierarchical early warning decision-making module, connected sequentially. Designed with individualization, forward-looking perspectives, and interpretability in mind, this system integrates three types of heterogeneous data—molecular omics, longitudinal clinical time-series data, and static baseline data—to construct a causal chain from molecular aging events to the evolution of kidney function within a unified timeline. Based on this, it enables dynamic projection and hierarchical early warning of future risk trajectories. The following will combine the attached... Figure 1 To be continued Figure 4 The functions, internal structure, data processing logic, and inter-module interaction mechanisms of each module are explained layer by layer.

[0022] The data fusion and preprocessing module, serving as the input port of the entire system, is responsible for receiving and processing raw data from various sources. The received data falls into three categories: the first is molecular omics data, including at least transcriptome sequencing data and epigenome methylation sequencing data; the second is longitudinal clinical time-series data, including at least urine albumin-to-creatinine ratios, estimated glomerular filtration rates, and glycated hemoglobin collected at multiple follow-up time points; and the third is static baseline data, including at least patient age, duration of diabetes, and history of hypertension. These data differ in acquisition time, sampling frequency, numerical scale, and data type, and cannot be directly used for subsequent modeling and analysis.

[0023] The data fusion preprocessing module performs two operations: time alignment and standardization. For time alignment, the module establishes a unified timeline with the patient's first visit or baseline examination time as the zero point. All longitudinal clinical time-series data are resampled according to this timeline, and missing time points are filled using linear interpolation or spline interpolation methods to ensure that each observation time point contains a complete clinical indicator vector. For molecular omics data, which are typically collected only at limited time points (e.g., baseline, 6 months, 12 months), the module treats them as sparse time series and maps them to corresponding positions on the unified timeline using timestamps. Time points not collected are marked as missing and processed later in the causal network construction stage. For standardization, the module performs Z-score standardization on all numerical variables, calculating the mean and standard deviation of each variable in the training set and converting the raw values ​​into dimensionless values ​​with a mean of 0 and a standard deviation of 1. For categorical variables (e.g., history of hypertension), a unique encoding is used to convert them into binary vectors. After the above processing, the data fusion preprocessing module outputs a time-aligned, scale-uniform, and structurally regular multimodal fusion dataset. This multimodal fusion dataset is organized in the form of a time series matrix, with each row representing a time point and each column representing a feature variable. Missing values ​​are marked with special tags for subsequent modules to identify and process.

[0024] The multi-scale causal network construction module receives the aforementioned multimodal fusion dataset and constructs an interpretable causal network from the molecular scale to the organ function scale. The module first extracts two key aging-related features from the molecular omics data: one is the expression aging index of a specific gene set, obtained by calculating the weighted average expression level of a set of genes known to be closely related to cellular aging (such as CDKN2A, TP53, IL6, etc.), with weights derived from publicly available literature or pre-trained models; the other is the acceleration value of the genome-wide methylation aging clock, calculated by inputting an individual's methylation profile into an epigenetic clock model trained on a large-scale population (such as the Horvath clock or PhenoAge clock), and then subtracting their actual physiological age to obtain a positive or negative acceleration bias.

[0025] After obtaining the aforementioned biological aging characteristics, the module uses them, along with longitudinal clinical time-series data (urinary albumin-to-creatinine ratio, estimated glomerular filtration rate, and glycated hemoglobin) and static baseline data, as a candidate variable set to initiate a constraint-based causal discovery algorithm. This algorithm strictly adheres to time-series constraints, prohibiting any variable from a later time point from being the dependent variable of a variable from an earlier time point. Under this premise, the algorithm employs a time-extended PC algorithm combined with a linear non-Gaussian acyclic model for causal structure learning. The specific steps are as follows: First, a completely undirected graph is constructed, with each node representing a variable (including the same variable at different time points); second, given a partial set of conditional variables, unconditional edges between variables are gradually eliminated through conditional independence tests (such as partial correlation tests or kernel-based independence tests); third, the directionality of the remaining edges is inferred using the time-series and non-Gaussianity assumptions, for example, if the upstream variable... Downstream variables There is a dependency relationship, and If the distribution exhibits non-Gaussian properties, then it is more likely that... Finally, for each directed edge, the strength of its causal effect is estimated using a linear non-Gaussian acyclic model, that is, the expected impact of a unit change in the upstream variable on the downstream variable is quantified.

[0026] The simulation process for generating intervention recommendations ultimately outputs a weighted directed causal graph model. Nodes encompass molecular aging characteristics, clinical indicators, and baseline variables. Directed edges represent connections with temporal sequence and potential causal relationships, with edge weights reflecting the strength and direction of the causal effect. This causal graph model not only reveals which molecular changes may lead to renal function deterioration but also quantifies the magnitude of the impact, providing structured prior knowledge for subsequent prediction and intervention.

[0027] The dynamic risk trajectory calculation module, based on the aforementioned causal graph model, constructs and runs a recurrent neural network prediction model embedding causal priors to achieve dynamic extrapolation of individualized risk. The input to the recurrent neural network prediction model is multimodal fusion data slices arranged in chronological order, with each slice containing all feature variable values ​​at a specific time point. The recurrent neural network prediction model employs a gated recurrent unit network, where the number of hidden state units corresponds one-to-one with the number of nodes in the causal graph; that is, each feature variable has its own dedicated hidden state unit in the neural network.

[0028] The key innovation lies in the fact that the connection matrix of this gated recurrent unit network is constrained to be a sparse matrix consistent with the adjacency matrix structure of the causal graph. Specifically, if a directed edge exists from node i to node j in the causal graph, the connection weights of the corresponding hidden state units i to j in the neural network are trainable parameters; if the connection does not exist in the causal graph, the corresponding weights are fixed at 0 and cannot be updated. This structural constraint ensures that the model's learning process strictly follows the underlying mechanisms revealed in the causal discovery phase, avoiding spurious correlations that traditional black-box models may learn.

[0029] The training objective of the model is to learn the specific parameterized form of causal relationships through historical data. Its loss function consists of two parts: the first part is the mean squared error between the predicted clinical indicators (urinary albumin-to-creatinine ratio and estimated glomerular filtration rate) and the actual observed values; the second part is the variance of the predicted trajectory, used to quantify the model's uncertainty regarding future states. During the inference phase, the dynamic risk trajectory calculation module uses observation data from several preceding time points as initial input, and performs multi-step forward inference through this causal-constrained recurrent neural network to predict key clinical indicator values ​​for multiple future time points (e.g., 6 months, 12 months, and 18 months).

[0030] The dynamic risk trajectory calculation module calculates a comprehensive risk index, which is defined as the weighted sum of the percentage increase in the predicted urinary albumin-to-creatinine ratio relative to its baseline value and the percentage decrease in the estimated glomerular filtration rate relative to its baseline value. This is then combined with the standard deviation of the predicted trajectory for uncertainty correction. Weighting coefficients are preset based on clinical importance; for example, the weight of the urinary albumin-to-creatinine ratio is set to 0.6, and the weight of the estimated glomerular filtration rate is set to 0.4. The formula for calculating the comprehensive risk index is as follows: ; For time points The comprehensive risk index, and These represent the model-predicted urinary albumin-to-creatinine ratio and the estimated glomerular filtration rate, respectively. and For individual baseline values, , To preset weights, To predict the trajectory at time points standard deviation The uncertainty penalty coefficient is set to 0.2. The formula above shows that the risk index not only reflects the degree of deterioration of clinical indicators but also imposes an additional penalty for situations with high predictive uncertainty, thereby improving the robustness of the early warning.

[0031] Finally, the hierarchical early warning decision-making module receives the predicted trajectory of the future comprehensive risk index and its uncertainty range output by the dynamic risk trajectory calculation module, and generates hierarchical early warning and decision support information accordingly. The interaction logic of this hierarchical early warning decision-making module is as follows: The hierarchical early warning decision-making module internally presets three risk thresholds: a low-risk threshold of 0.15, a medium-risk threshold range of [0.15, 0.35], and a high-risk threshold of 0.35. These thresholds are determined based on the statistical distribution of the incidence and risk index of diabetic nephropathy progression events in large-scale cohort studies.

[0032] The hierarchical early warning decision-making module compares the predicted comprehensive risk index trajectory with the aforementioned thresholds in real time. When the predicted trajectory enters the medium-risk threshold range at any future observation point, the hierarchical early warning decision-making module triggers a first-level early warning signal. At this time, the system automatically generates an analysis report, highlighting the recent trends in molecular aging characteristics (such as methylation clock acceleration values ​​or expression aging indices) and listing the causal paths most relevant to the current risk status. The hierarchical early warning decision-making module triggers a second-level early warning signal when the predicted trajectory meets either of the following conditions: first, the predicted trajectory is in the medium-risk range for two consecutive preset observation periods (such as two consecutive 6-month follow-up points); second, the predicted trajectory directly crosses the medium-risk range and enters the high-risk threshold. After the second-level early warning is triggered, the system not only issues a high-risk alarm but also initiates causal backtracking analysis and intervention simulation.

[0033] Causal backtracking analysis, based on the causal graph model output by the multi-scale causal network construction module, identifies key driving factors pointing to high-risk states. Specifically, starting from the clinical indicator nodes corresponding to high-risk states, a breadth-first search is performed along the reverse edges of the causal graph to trace upstream molecular aging characteristic nodes. The cumulative causal effect weight of each path is calculated, and the top three paths with the highest effect weights are selected as key driving paths. Intervention simulation suggestions are implemented through counterfactual reasoning: the system assumes a positive change in upstream molecular aging characteristic variables (e.g., a 10% reduction in methylation clock acceleration) on the causal graph model, and uses the dynamic risk trajectory calculation module to re-predict the trajectory, comparing the changes in the comprehensive risk index before and after the intervention. Simulation results are presented in quantitative form, for example, "If the methylation clock acceleration value decreases by 10%, the comprehensive risk index is expected to decrease by 0.08 after 12 months, moving out of the high-risk range." The system prioritizes different interventions based on the magnitude of risk reduction and intervention feasibility, forming a structured list of intervention recommendations.

[0034] This system also includes a continuous model learning and updating module. This module periodically extracts samples from newly included patient data to incrementally learn the causal discovery algorithm in the multi-scale causal network construction module, detecting whether the causal structure drifts with changes in population characteristics or clinical practice. Simultaneously, it uses new data to fine-tune the prediction model in the dynamic risk trajectory calculation module, updating its trainable parameters to ensure performance does not degrade over time. This mechanism guarantees the system's adaptability and accuracy during long-term deployment.

[0035] In summary, this embodiment achieves accurate, interpretable, and dynamic early warning of diabetic nephropathy risks by strictly following a four-stage process of data fusion, causal modeling, dynamic extrapolation, and hierarchical decision-making. The system not only provides early warning time windows but also offers a complete decision-making chain, from identifying the risk to understanding why it occurs and how to intervene, thus enhancing its clinical applicability and scientific value.

[0036] Example 2: Building upon Example 1, this example optimizes the causal discovery algorithm in the multi-scale causal network construction module to improve robustness in scenarios with small samples or high noise. Specifically, this example introduces a Bayesian structure learning framework to replace the original frequentist PC algorithm, thereby better handling data uncertainty and incorporating prior knowledge.

[0037] In this embodiment, the multi-scale causal network construction module still uses biological aging characteristics, longitudinal clinical time-series data, and static baseline data as the input variable set. However, its causal discovery process no longer relies on the hard elimination of conditional independence tests, but instead constructs a posterior probability distribution of the causal graph structure. This posterior probability distribution is jointly determined by the likelihood function and the structural prior. The likelihood function is based on a linear Gaussian structural equation model, assuming that each variable is a linear combination of its parent nodes plus Gaussian noise. The structural prior encodes domain knowledge, such as requiring that molecular aging characteristics can only be used as dependent variables and clinical indicators can only be used as effect variables, and that the time sequence constraint remains strictly in effect.

[0038] The multi-scale causal network construction module extracts a large number of causal graph samples from the posterior distribution using the Markov chain Monte Carlo sampling method. The final output causal graph is not a single structure, but a weighted graph. The weight of each edge is its frequency of occurrence in all sampled graphs, i.e., the posterior probability of learning a causal relationship from the historical data. For example, if the edge "methylation clock acceleration value → urine albumin-creatinine ratio" appears 8,500 times out of 10,000 samples, then the weight of this edge is 0.85. This weight not only reflects the strength of the causal relationship but also directly quantifies its reliability.

[0039] In the dynamic risk trajectory calculation module, this weighted causal graph is used to construct a soft-constrained recurrent neural network. Specifically, the elements in the neural network connection matrix are no longer simply set to 0 or trainable parameters, but are initialized to the posterior probability values ​​of the corresponding causal edges, and are allowed to be slightly adjusted during training. A regularization term is added to the loss function to penalize the degree to which the connection weights deviate from the initial posterior probability. Its regularization coefficient is dynamically adjusted according to the sample size—the smaller the sample size, the stronger the regularization, and the more dependent it is on the prior; the larger the sample size, the weaker the regularization, and the more dependent it is on the data. This mechanism balances prior knowledge and data-driven approaches, preventing overfitting in small-sample scenarios and maintaining flexibility in large-sample scenarios.

[0040] The tiered early warning decision-making module has also been adjusted accordingly. When calculating the comprehensive risk index, not only predicted values ​​and uncertainties are considered, but also the causal marginal posterior probability is introduced as a confidence factor. For example, if the causal marginal posterior probability on a key driving path is less than 0.7, the intervention simulation recommendation will indicate "The evidence level of this key driving path is moderate; it is recommended to combine it with other biomarkers for verification." The triggering conditions for Level 1 early warning have been augmented with a requirement for causal confidence: an early warning signal is only generated when the average posterior probability of the causal path upon which the medium-risk prediction depends is greater than 0.75, thus avoiding false alarms caused by low-confidence causal relationships.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus.

[0042] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A risk stratification and early warning system for early diabetic nephropathy, characterized in that, include: The data fusion preprocessing module is used to receive and process heterogeneous multimodal input data from different sources, including molecular omics data, longitudinal clinical time-series data, and static baseline data. A multi-scale causal network construction module, connected to the data fusion preprocessing module, is used to construct an interpretable causal network from the molecular scale to the organ function scale based on the multi-modal fusion dataset. The dynamic risk trajectory calculation module is connected to the multi-scale causal network construction module and is used to perform individualized dynamic risk calculation and trajectory prediction based on the constructed causal graph model. The hierarchical early warning and decision-making module is connected to the dynamic risk trajectory calculation module and is used to generate hierarchical early warning and decision support information based on the calculation results.

2. The early risk stratification and early warning system for diabetic nephropathy according to claim 1, characterized in that, The molecular omics data shall at least include transcriptome sequencing data and epigenome methylation sequencing data; The longitudinal clinical time-series data include at least the urine albumin-to-creatinine ratio, estimated glomerular filtration rate, and glycated hemoglobin collected over time. The static baseline data includes at least the patient's age, duration of diabetes, and history of hypertension. The data fusion preprocessing module further performs data alignment and standardization operations, specifically by resampling and aligning the longitudinal clinical time-series data based on a unified time axis, and standardizing the molecular omics data and clinical time-series data to eliminate dimensional differences, thereby generating a time-aligned and scale-uniform multimodal fusion dataset.

3. The early risk stratification and early warning system for diabetic nephropathy according to claim 2, characterized in that, The multi-scale causal network construction module first extracts biological aging-related features from the molecular omics data, specifically by calculating the expression aging index of a specific gene set and the acceleration value of the genome-wide methylation aging clock; The multi-scale causal network construction module further employs a constraint-based causal discovery algorithm, using the biological aging-related characteristics as candidate dependent variables, the longitudinal clinical time-series data as candidate outcome variables, and introducing the static baseline data as conditional variables, to learn the causal structure under time-series constraints. During the causal structure learning process, a directional causal graph model is output. Nodes represent feature variables at different scales, and directed edges represent potential causal relationships between variables with temporal order. The weights of the edges are calculated in subsequent causal effect quantification steps.

4. The early risk stratification and early warning system for diabetic nephropathy according to claim 3, characterized in that, The dynamic risk trajectory calculation module includes a recurrent neural network prediction model with embedded causal priors. The input to the recurrent neural network prediction model is a multimodal fusion data slice arranged in chronological order, which has been processed by the data fusion preprocessing module. The network structure of the recurrent neural network prediction model is constrained by the causal graph model output by the multi-scale causal network construction module. Specifically, trainable parameterized connections are established only between neural network units corresponding to nodes with directed edge connections in the causal graph model. For connections that do not exist in the causal graph model, the corresponding neural network connection weights are fixed to 0. The recurrent neural network prediction model learns the specific parameterization and strength of the relationships represented by each directed edge in the causal graph model through training; During the inference phase, the dynamic risk trajectory calculation module takes the observation data of the preceding time points as input, performs multi-step forward extrapolation through the recurrent neural network prediction model, predicts the values ​​of key clinical indicators such as the urinary albumin-to-creatinine ratio and the estimated glomerular filtration rate at multiple future time points, and simultaneously calculates the comprehensive risk index. The comprehensive risk index is a weighted function of the deviation of the future predicted urinary albumin-to-creatinine ratio from the estimated glomerular filtration rate relative to its baseline value, and the uncertainty of the predicted trajectory.

5. The early risk stratification and early warning system for diabetic nephropathy according to claim 4, characterized in that, The hierarchical early warning decision module receives the predicted trajectory of the future comprehensive risk index and its uncertainty range output by the dynamic risk trajectory calculation module. The hierarchical early warning decision module has three preset risk thresholds, which correspond to low risk, medium risk and high risk states respectively. The hierarchical early warning decision module compares the predicted comprehensive risk index trajectory with these thresholds in real time. When the predicted trajectory enters the medium-risk threshold range at any future observation time, the hierarchical early warning decision module generates a first-level early warning signal and outputs an analysis report prompting attention to changes in molecular aging characteristics. When the predicted trajectory remains in the medium-risk range for more than two preset observation periods, or directly enters the high-risk threshold range, the hierarchical early warning decision module generates a secondary early warning signal and outputs a list of key driving factors derived from causal network backtracking analysis, as well as targeted intervention simulation suggestions.

6. The early risk stratification and early warning system for diabetic nephropathy according to claim 5, characterized in that, The constraint-based causal discovery algorithm employs a time-extended PC algorithm combined with a linear non-Gaussian acyclic model; the execution process of this causal discovery algorithm is as follows: Based on the time order constraint, variables at later time points are prohibited from being the cause of variables at earlier time points; By using the conditional independence test, given a partial set of conditional variables, we can gradually eliminate unconditional edges between variables. The retained edges are inferred for directionality, and the causal effect strength of the edges is quantitatively estimated using a linear non-Gaussian acyclic model, thus obtaining a weighted causal graph.

7. The early risk stratification and early warning system for diabetic nephropathy according to claim 6, characterized in that, The recurrent neural network prediction model with embedded causal priors adopts a gated recurrent unit network structure; In the gated recurrent unit network structure, each feature variable in the causal graph model corresponds to a specific hidden state unit; The connection matrix between network layers is forced to be a sparse trainable matrix consistent with the adjacency matrix structure of the causal graph model. In the causal graph model, the matrix elements corresponding to directed edge connections are trainable parameters, otherwise they are constant zero values. The loss function of the recurrent neural network prediction model consists of two parts: the first part is the mean square error between the predicted clinical indicator value and the true value, and the second part is the uncertainty estimate represented by the variance of the predicted trajectory.

8. The early risk stratification and early warning system for diabetic nephropathy according to claim 7, characterized in that, The process for generating intervention measure simulation suggestions in the hierarchical early warning decision-making module is as follows: When the Level II warning signal is triggered, the system identifies the key causal path pointing to the high-risk state based on the causal graph model obtained by the multi-scale causal network construction module. The system performs counterfactual reasoning on the causal graph model to simulate the impact on the predicted trajectory of downstream clinical indicators if the upstream molecular aging characteristic variables in the list of key driving factors undergo positive changes. The simulation results quantitatively predict the possible magnitude and time of decline in the risk index under different intervention intensities, thereby forming simulation recommendations for intervention measures with priority ranking.

9. The early risk stratification and early warning system for diabetic nephropathy according to claim 8, characterized in that, The system also includes a module for continuous model learning and updating; The model continuous learning and updating module is connected to the data fusion preprocessing module and the multi-scale causal network construction module, and is used to continuously integrate newly generated patient data after system deployment; The model continuous learning and updating module is started periodically to incrementally learn the causal discovery algorithm using new data in order to update the causal graph structure. The newly added data is then used to incrementally train the prediction model in the dynamic risk trajectory calculation module to optimize the model parameters.

10. A method for early risk stratification and early warning of diabetic nephropathy, characterized in that, Performed by the system according to any one of claims 1 to 9, the method includes the following steps: Data fusion preprocessing steps: Collect molecular omics data, longitudinal clinical time-series data, and static baseline data of the target individuals; align the longitudinal clinical time-series data based on a unified time axis, and standardize all data to generate a multimodal fusion dataset; The steps of the multi-scale causal network construction module are as follows: calculate biological aging characteristics from the molecular omics data; use the aging characteristics, clinical time-series indicators and baseline data as variables, apply a time-constrained causal discovery algorithm to learn and construct a weighted directed causal graph model from the molecular scale to the organ function scale; Dynamic risk trajectory calculation steps: The individual's historical multimodal fusion data is input into a causal prior recurrent neural network prediction model according to the time series; the network structure of the recurrent neural network prediction model is constrained by the causal graph model; through model deduction, the key clinical indicator values ​​and comprehensive risk index trajectory of the individual at multiple future time points are predicted; The tiered early warning decision-making process involves comparing the predicted trajectory of the future comprehensive risk index with preset low, medium, and high risk thresholds; generating risk warning signals of different levels based on the threshold range and duration reached by the predicted trajectory, and outputting corresponding causal analysis reports and intervention simulation suggestions.

Images