Diabetic complication risk deduction system based on graph neural network and knowledge graph

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a diabetic complication risk deduction system based on a graph neural network and a knowledge graph, and the system comprises a physiological feature sensing module which collects physiological sensing data of a diabetic patient, carries out the time synchronization and normalization processing of the physiological sensing data, and carries out the prediction of the diabetic complication risk. Separating a stability characteristic component and a disturbance characteristic component; the pathological potential mapping module is used for mapping the stable characteristic component into a basic physiological node and mapping the disturbance characteristic component into a potential pathological activation node on the basis of a semantic structure of a knowledge graph, introducing a pathological potential function and establishing a pathological potential mapping network; the pathology activation recognition module calculates potential pathology energy of each potential pathology activation node in the pathology potential mapping network through a graph neural network, and when the potential pathology energy is greater than a preset potential pathology energy threshold, the potential pathology activation node is marked as a pathology activation source node; and early warning of diabetic complications is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a system for predicting the risk of diabetic complications based on graph neural networks and knowledge graphs. Background Technology

[0002] Existing systems for projecting the risk of diabetes complications have the following main problems: The development of diabetes and its complications involves a complex coupling of multiple pathological mechanisms, including glycemic regulation, pancreatic islet cell function, inflammatory response, and tissue metabolic function. Existing methods for assessing the risk of diabetes complications mainly rely on single-point indicators or fixed thresholds, which are insufficient to characterize the dynamic changes of pathological mechanisms over time and their causal relationships, and cannot reflect the time dependence of different physiological disturbances on the accumulation of potential pathological risks.

[0003] Traditional methods have several limitations in risk identification: they cannot identify which nodes will become the starting points of risk diffusion based on individual physiological data. Most methods only provide single-indicator risk assessments and cannot simulate the accumulation and transmission of risk among multiple factors and nodes, as well as the self-repair mechanisms of the nodes themselves. They cannot consider the efficiency of risk transmission between pathological nodes and the risk decay capacity of the nodes themselves, leading to inaccurate predictions of the evolution trajectory of diabetic complications. Existing methods are mostly based on static snapshots, ignoring the changing trends of physiological indicators in time series, and cannot effectively identify early risks, nor can they perform high-precision individualized risk extrapolation and differentiated risk path analysis for different patients.

[0004] In view of this, the present invention proposes a risk prediction system for diabetic complications based on graph neural networks and knowledge graphs to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: a risk prediction system for diabetic complications based on graph neural networks and knowledge graphs, comprising: The physiological feature perception module collects physiological perception data from diabetic patients, performs time synchronization and normalization processing on the physiological perception data, and separates the stable feature components and the perturbation feature components. The pathological potential mapping module, based on the semantic structure of the knowledge graph, maps stable feature components to basic physiological nodes and perturbation feature components to potential pathological activation nodes. It introduces a pathological potential function and establishes a pathological potential mapping network. The pathological activation identification module calculates the potential pathological energy of each potential pathological activation node in the pathological potential mapping network through a graph neural network. When the potential pathological energy is greater than the preset potential pathological energy threshold, it is marked as a pathological activation source node. The risk conservation deduction module constructs a risk flow conservation graph based on pathological activation source nodes, treats diabetes-related risks as energy flows propagating in the risk flow conservation graph, calculates the flow rate, direction and cumulative energy of risks between different pathological activation source nodes, and deduces the evolution trajectory of diabetes complications. The risk feedback evolution module uses a graph attention mechanism to weighted convergence of the evolutionary trajectories of diabetic complications, generating individualized risk trend curves for diabetic complications.

[0006] Specifically, the physiological perception data of the diabetic patients includes physiological data on glucose metabolism, cardiovascular physiological data, nervous system perception data, renal function data, and lifestyle behavior data.

[0007] Specifically, the method for time synchronization and normalization of physiological perception data includes: A sliding time window is used to synchronize the collected physiological perception data to form a time-synchronized physiological signal matrix; the physiological signal matrix is ​​then normalized to obtain a set of physiological perception feature vectors with a unified structure.

[0008] Specifically, the method for separating the stability characteristic component and the disturbance characteristic component includes: The physiological perception feature vector set is subjected to feature decomposition to separate stable feature components and perturbation feature components. Specifically, a combination of weighted moving average and principal component analysis is used to extract the low-frequency variation part of the physiological perception feature vector as the stable feature component, while a combination of high-pass filtering and residual analysis is used to extract the high-frequency variation part of the physiological perception feature vector as the perturbation feature component.

[0009] Specifically, the method for constructing the pathological potential mapping network includes: The pathological mechanism elements of diabetes are extracted from a pre-set database and used as objects to construct knowledge graph nodes and edges. The extracted pathological mechanism elements are abstracted into nodes and causal relationships between nodes in the knowledge graph, forming a semantic hierarchical structure. The nodes of the knowledge graph include basic physiological nodes and potential pathological activation nodes. Stable feature components are mapped to basic physiological nodes of the knowledge graph, and perturbation feature components are mapped to potential pathological activation nodes. A pathological potential function is introduced to perform integral modeling on the dynamic change curve of each potential pathological activation node within a preset time interval, and to calculate the risk potential energy accumulated by each potential pathological activation node in the time dimension. When the risk potential energy of a potential pathological activation node is greater than zero, it indicates that the potential pathological activation node is in an accumulation state; when the risk potential energy is equal to zero, it indicates that the potential pathological activation node is in a balanced state; when the risk potential energy is less than zero, it indicates that the potential pathological activation node is in a self-repair state. The risk potential energy accumulated over time by potential pathological activation nodes is assigned to the corresponding nodes; the nodes, edges, and risk potential energy of the knowledge graph are integrated to form a complete pathological potential mapping network.

[0010] Specifically, the method for obtaining the potential pathological energy of each potential pathological activation node includes: The risk potential energy accumulated over time for each potential pathological activation node is used as a node feature. The node features also include node type and historical physiological indicators. A graph neural network (GNN) is used to iteratively message pass and weight the features of each node with the features of its neighboring nodes. The node features are updated through a nonlinear mapping function to generate the state representation of the node in the current iteration. Repeatedly perform message passing and node feature update iterations until the preset iteration round is reached. The node features after the final iteration are converted into scalar values ​​through the output mapping function to obtain the potential pathological energy of each potential pathological activation node.

[0011] Specifically, the method for marking the pathological activation source nodes includes: A preset potential pathological energy threshold is set. The potential pathological energy of each potential pathological activation node is compared with the preset potential pathological energy threshold. When the potential pathological energy is greater than the preset potential pathological energy threshold, the potential pathological activation node is marked as a pathological activation source node.

[0012] Specifically, the method for constructing the risk flow conservation graph includes: The marked pathological activation source nodes are used as the pathological node set of the risk flow conservation graph. The causal relationship and semantic hierarchy between the pathological mechanism elements extracted in the knowledge graph are used to determine the direction of the directed edges between pathological nodes in the risk flow conservation graph. The risk flow conservation graph is formed based on the pathological node set and the directed edges.

[0013] Specifically, the method for extrapolating the evolutionary trajectory of diabetic complications includes: The risk potential energy corresponding to the pathological node at the initial time point is taken as the instantaneous risk intensity of the pathological node. A risk flow conservation equation is constructed, the rate of change of the instantaneous risk intensity of any pathological node in the risk flow conservation diagram is calculated, the risk flow conservation equation is solved iteratively, and the risk flow rate and direction between pathological nodes are updated in real time to obtain the complete risk flow path. The instantaneous risk intensity of pathological nodes is integrated over a preset time interval to obtain the cumulative energy of risk between pathological nodes; based on the direction of risk between pathological nodes and the cumulative energy, risk flow paths with cumulative energy greater than a preset cumulative energy threshold are identified as high-risk paths; the evolutionary trajectory of diabetes into different diabetic complications is predicted based on the high-risk paths.

[0014] Specifically, the method for generating individualized risk trend curves for diabetes complications includes: The evolutionary trajectory of diabetic complications is extracted into the pathological nodes corresponding to the risk flow conservation graph. Based on the flow rate, direction and cumulative energy between pathological nodes, a node feature vector set is constructed and linearly transformed to obtain risk features. The similarity of the node feature vectors of adjacent pathological nodes is calculated by cosine similarity. The calculated similarity values ​​are normalized to obtain the attention weight coefficient between adjacent pathological nodes. The risk features of adjacent pathological nodes are weighted and aggregated using normalized attention weight coefficients to form weighted aggregated risk features. A risk feedback matrix is ​​constructed based on the weighted aggregated risk features, and the risk feedback matrix is ​​subjected to time-series fitting and smoothing to generate individualized risk trend curves for diabetic complications.

[0015] Compared with the prior art, the present invention has the following beneficial effects: This invention introduces a pathological potential function to integrate the dynamic change curves of potential pathological activation nodes over time, calculating risk potential energy and achieving cumulative risk modeling. It can distinguish whether a node is in an accumulation state, a equilibrium state, or a self-repair state, providing more refined temporal information for risk management. By comparing the risk potential energy with a preset threshold, potential pathological activation nodes are marked as pathological activation source nodes, clarifying the starting point of risk diffusion and providing a basis for individualized intervention. Networked modeling can depict the complex coupling relationships between pathological mechanism elements, improving the accuracy of risk projection. The system integrates the dynamic changes of physiological indicators, the causal relationships between nodes, and the cumulative risk potential energy, enabling more accurate identification of risk nodes and avoiding misjudgments caused by single indicators or static thresholds.

[0016] By integrating the instantaneous risk intensity of a node over a time interval, the cumulative risk of that node can be obtained, enabling modeling of the cumulative energy of risk in the network. Based on the direction of risk flow and cumulative energy, risk flow paths with cumulative energy exceeding a threshold can be identified, clarifying potentially high-risk pathological pathways in diabetes and providing a basis for early intervention. Using high-risk pathways, the impact of pathological nodes on downstream nodes can be deduced, thereby predicting the possible evolution of diabetes into different types of complications and achieving personalized risk trajectory prediction. By establishing a causal relationship structure between nodes of pathological mechanism elements through a knowledge graph and using graph neural networks for propagation analysis, comprehensively considering the coupling strength and self-repair capabilities between nodes, the dynamic evolution process of diabetic complications can be simulated more accurately. Dynamic modeling and cumulative risk analysis can effectively distinguish between short-term fluctuations and long-term risk accumulation, avoiding misjudgments caused by traditional single-indicator or static threshold methods and improving the reliability of clinical decision-making. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of the diabetes complication risk prediction system based on graph neural network and knowledge graph of the present invention; Figure 2 This is a schematic diagram of the process for extrapolating the risk of diabetic complications based on graph neural networks and knowledge graphs according to the present invention. Detailed Implementation

[0018] 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.

[0019] Example 1

[0020] Please see Figure 1 As shown, this embodiment provides a system for extrapolating the risk of diabetes complications based on graph neural networks and knowledge graphs, specifically including the following steps: The physiological feature perception module collects physiological perception data from diabetic patients, performs time synchronization and normalization processing on the physiological perception data, and separates the stable feature components and the perturbation feature components. The pathological potential mapping module, based on the semantic structure of the knowledge graph, maps stable feature components to basic physiological nodes and perturbation feature components to potential pathological activation nodes. It introduces a pathological potential function and establishes a pathological potential mapping network. The pathological activation identification module calculates the potential pathological energy of each potential pathological activation node in the pathological potential mapping network through a graph neural network. When the potential pathological energy is greater than the preset potential pathological energy threshold, it is marked as a pathological activation source node. The risk conservation deduction module constructs a risk flow conservation graph based on pathological activation source nodes, treats diabetes-related risks as energy flows propagating in the risk flow conservation graph, calculates the flow rate, direction and cumulative energy of risks between different pathological activation source nodes, and deduces the evolution trajectory of diabetes complications. The risk feedback evolution module uses a graph attention mechanism to weighted convergence of the evolutionary trajectories of diabetic complications, generating individualized risk trend curves for diabetic complications.

[0021] Physiological data for diabetic patients includes glucose metabolism data, cardiovascular data, nervous system data, renal function data, and lifestyle data.

[0022] Physiological perception data of diabetic patients is the basic input data for risk extrapolation. Its scope should cover physiological signals reflecting multiple dimensions such as the body's metabolic state, cardiovascular state, neurological state and kidney function, and have the characteristics of multimodality, temporality and physiological correlation. Among them, the physiological data of glucose metabolism include blood glucose levels (fasting blood glucose, two-hour postprandial blood glucose), glycated hemoglobin (HbA1c), insulin secretion, blood lipid indicators (total cholesterol, triglycerides, high-density lipoprotein HDL and low-density lipoprotein LDL), as well as urinary glucose and urinary protein concentrations, which are used to characterize the glucose and lipid metabolism balance and insulin regulation ability of diabetic patients. Cardiovascular physiological data, including blood pressure (systolic, diastolic, and mean arterial pressure), heart rate and heart rate variability (HRV), pulse wave transit time (PWV), arteriosclerosis index, and blood oxygen saturation, are used to describe the hemodynamics and circulatory health status of diabetic patients and identify potential cardiovascular risk patterns. Neurosensory data, including electrical skin response (EDA), peripheral nerve conduction velocity (NCV), plantar temperature distribution (acquired via an infrared thermal imaging sensor), and tactile sensitivity, are used to identify the risk of neurological decline and peripheral nerve damage in diabetic patients. Renal function data include urinary microalbumin, serum creatinine and estimated glomerular filtration rate, blood electrolyte concentration (sodium, potassium, calcium, phosphorus, etc.), blood viscosity and red blood cell deformability, which are used to reveal the changing trends of renal function compensation capacity and blood circulation efficiency. Lifestyle data, including dietary records and carbohydrate intake, sleep quality and duration, physical activity levels (steps, exercise intensity, sedentary time), stress levels, and mood fluctuation index, are used to construct behavioral-physiological interaction characteristics of diabetic patients.

[0023] Methods for time synchronization and normalization of physiological sensory data include: A sliding time window was used to synchronize the collected physiological sensory data, forming a time-synchronized physiological signal matrix. The physiological signal matrix was then normalized to obtain a set of physiological sensory feature vectors with a uniform structure. Continuous physiological signals were standardized using Z-score, interval-type physiological signals were normalized using Min-Max, and nonlinear fluctuating physiological signals were normalized using a smoothing normalization algorithm to achieve uniformity in the feature scale of the physiological signals.

[0024] Methods for separating stability characteristic components from disturbance characteristic components include: The physiological perception feature vector set is subjected to feature decomposition to separate stable and perturbation feature components. Specifically, a combination of weighted moving average and principal component analysis is used to extract the low-frequency variation portion of the physiological perception feature vector as the stable feature component, used to characterize the long-term physiological equilibrium state of diabetic patients. A combination of high-pass filtering and residual analysis is used to extract the high-frequency variation portion of the physiological perception feature vector as the perturbation feature component, used to characterize abnormal physiological fluctuations caused by short-term stress factors such as dietary intake, exercise, emotional fluctuations, or drug intervention.

[0025] It should be noted that the low-frequency variation component refers to the slowly varying trend component in the physiological signal power spectrum, whose energy is mainly distributed between 0-0.3Hz, and is used to characterize the individual's stability state; the high-frequency variation component refers to the rapidly fluctuating component in the physiological signal power spectrum, whose energy is mainly distributed between 0.3-2Hz, and is used to characterize the instantaneous disturbance response. The physiological signal power spectrum is obtained by performing a fast Fourier transform on the time-synchronized physiological signal matrix and then performing frequency domain analysis.

[0026] Methods for constructing pathological potential mapping networks include: The pathological mechanism elements of diabetes are extracted from a pre-defined database (such as clinical literature and physiological indicator databases). These pathological mechanism elements include glycemic regulation and pancreatic islet function. Cellular function, inflammatory response, and tissue metabolic function; serving as objects for constructing knowledge graph nodes and edges; abstracting extracted pathological mechanism elements into nodes and causal relationships between nodes in the knowledge graph, forming a semantic hierarchical structure; The nodes of the knowledge graph include basic physiological nodes and potential pathological activation nodes. Basic physiological nodes are located at the bottom layer and are used to characterize the normal physiological function indicators of the body. Potential pathological activation nodes are located at the upper or middle layer and are used to characterize the physiological perturbation indicators that may trigger pathological states. Stability feature components are mapped to the basic physiological nodes of the knowledge graph, and perturbation feature components are mapped to potential pathological activation nodes. A pathological potential function is introduced to perform integral modeling on the dynamic change curve of each potential pathological activation node within a preset time interval, and to calculate the risk potential energy accumulated by each potential pathological activation node in the time dimension. The pathological potential function is: ;in, Indicates potential pathological activation nodes In time interval The accumulated risk potential energy; Represents the first in a knowledge graph One potential pathological activation node; The time point that represents the upper limit of integration; Let be the integral variable, representing an instantaneous point in time within the time interval; Indicates potential pathological activation nodes At the point of time The dynamic change curve, that is, the physiological indicators corresponding to potential pathological activation nodes (such as blood glucose regulation indicators, pancreatic islet... Actual observed values ​​of cell function indicators, inflammation-related indicators, and tissue metabolic function indicators over time; Indicates potential pathological activation nodes The deviation between the corresponding physiological indicators and the preset health reference range; It represents the pathological acceleration coefficient, reflecting the strength of the contribution of the physiological indicator to the development of the pathology, such as a rapid rise in blood sugar or an increase in inflammatory markers; It represents the pathological dissipation coefficient, reflecting the body's self-repair ability; Indicates potential pathological activation nodes At the point of time The instantaneous rate of change; When the risk potential energy of a potential pathological activation node is greater than zero, it indicates that the node is in an accumulation state, driving risk propagation in the pathological potential mapping network and influencing other nodes. When the risk potential energy is zero, the node is in a balanced state, neither absorbing nor outputting risk, essentially acting as a stable node. When the risk potential energy is less than zero, the node is in a self-repairing state, absorbing risk and mitigating its spread within the network.

[0027] The risk potential energy accumulated over time by potential pathological activation nodes is assigned to the corresponding nodes; the nodes, edges, and risk potential energy of the knowledge graph are integrated to form a complete pathological potential mapping network.

[0028] This approach addresses the following technical problems of existing technologies: Most existing methods rely on single-point indicators or fixed thresholds for judgment, neglecting the dynamic causal relationships between pathological mechanisms such as blood glucose regulation, pancreatic islet cell function, inflammatory responses, and tissue metabolic functions. They cannot characterize the time dependence of different physiological disturbances on the accumulation of potential pathological risks. Traditional methods cannot distinguish between stable and disturbed states based on individual physiological data, nor can they determine which nodes will become the starting points for risk diffusion. Existing technologies typically only provide single-indicator risk assessments, lacking modeling of the accumulation, transmission, and self-repair mechanisms of risk among multiple factors and nodes. They cannot simulate the evolutionary trajectory of diabetic complications under the coupling of pathological mechanisms, nor can they provide individualized risk predictions; traditional assessments are mostly based on static snapshots, ignoring the changing trends of physiological indicators in time series, leading to the ineffective identification of early risks.

[0029] Compared to existing technologies, the advantages are as follows: By introducing a pathological potential function, the dynamic change curve of potential pathological activation nodes over time is integrated to calculate risk potential energy, achieving cumulative risk modeling. It can distinguish whether a node is in an accumulation state, a equilibrium state, or a self-repair state, providing more refined temporal information for risk management. By comparing the risk potential energy with a preset threshold, potential pathological activation nodes are marked as pathological activation source nodes, clarifying the starting point of risk diffusion and providing a basis for individualized intervention. Networked modeling can depict the complex coupling relationships between pathological mechanism elements, improving the accuracy of risk projection. The system integrates the dynamic changes of physiological indicators, the causal relationships between nodes, and the cumulative risk potential energy, enabling more accurate identification of risk nodes and avoiding misjudgments caused by single indicators or static thresholds.

[0030] Methods for obtaining the potential pathological energy of each potential pathological activation node include: The risk potential energy accumulated over time for each potential pathological activation node is used as a node feature. The node features also include node type and historical physiological indicators. A graph neural network (GNN) is used to iteratively message-pass and weightedly fuse each node feature with the features of its neighboring nodes. The node features are updated using a nonlinear mapping function (such as the ReLU activation function) to generate the embedding representation of the node in the current iteration. Repeatedly perform message passing and node feature update iterations until the preset iteration round is reached. The node features after the final iteration are converted into scalar values ​​through the output mapping function (such as linear transformation or weighted summation) to obtain the potential pathological energy of each potential pathological activation node.

[0031] Methods for labeling pathological activation source nodes include: A preset potential pathological energy threshold is set. The potential pathological energy of each potential pathological activation node is compared with the preset potential pathological energy threshold. When the potential pathological energy is greater than the preset potential pathological energy threshold, the potential pathological activation node is marked as a pathological activation source node.

[0032] Methods for constructing risk flow conservation graphs include: The marked pathological activation source nodes are used as the pathological node set of the risk flow conservation graph. The causal relationship and semantic hierarchy (such as the hierarchy of the bottom basic physiological nodes, the middle or upper potential pathological activation nodes) between the extracted pathological mechanism elements in the knowledge graph are used to determine the direction of the directed edges between pathological nodes in the risk flow conservation graph. The risk flow conservation graph is formed based on the pathological node set and the directed edges.

[0033] Methods for extrapolating the evolutionary trajectory of diabetic complications include: The risk potential energy corresponding to the pathological node at the initial time point is taken as the instantaneous risk intensity of the pathological node. A risk flow conservation equation is constructed, the rate of change of the instantaneous risk intensity of any pathological node in the risk flow conservation diagram is calculated, the risk flow conservation equation is solved iteratively, and the risk flow rate and direction between pathological nodes are updated in real time to obtain the complete risk flow path. The risk flow conservation equation is: ;in, Representing pathological nodes in the risk flow conservation graph At the point of time The instantaneous rate of change of risk intensity; Indicates pathological nodes The instantaneous risk intensity, for pathological nodes At the initial time point The corresponding risk potential energy; Indicates pathological nodes A set of neighboring pathological nodes with risk transmission connections; Indicates pathological nodes To pathological nodes The risk transmission coefficient describes the efficiency of risk transmission between pathological nodes. Representing pathological nodes in the risk flow conservation graph At the point of time The corresponding risk potential energy; Representing pathological nodes in the risk flow conservation graph At the point of time The corresponding risk potential energy; Indicates pathological nodes The self-inhibition coefficient describes the risk attenuation ability of a pathological node itself. Indicates neighboring pathological nodes With pathological nodes At the point of time The risk potential energy difference determines the direction of risk flow between pathological nodes. Risk originates from pathological nodes Flow to pathological nodes ; The instantaneous risk intensity of pathological nodes is integrated over a preset time interval to obtain the cumulative energy of risk between pathological nodes; based on the direction of risk between pathological nodes and the cumulative energy, risk flow paths with cumulative energy greater than a preset cumulative energy threshold are identified as high-risk paths; the evolutionary trajectory of diabetes into different diabetic complications is predicted based on the high-risk paths.

[0034] This approach addresses the following technical problems of existing technologies: Traditional methods often rely on single-point indicators or fixed thresholds to assess risk, failing to characterize the dynamic changes of diabetic pathological mechanisms over time. They cannot calculate the transmission rate, direction, and cumulative effect of risk between different pathological nodes, leading to difficulties in early risk identification. Traditional methods do not consider the efficiency of risk transmission between pathological nodes or the risk attenuation capacity of the nodes themselves, resulting in inaccurate predictions of the evolution trajectory of complications under the coupling of pathological mechanisms. Existing methods cannot perform high-precision individualized complication risk extrapolation based on individual physiological data and lack the ability to analyze differentiated risk pathways among different patients.

[0035] Compared to existing technologies, the advantages are as follows: By integrating the instantaneous risk intensity of a node over a time interval, the cumulative risk of the node can be obtained, enabling modeling of the cumulative energy of risk in the network. Based on the risk flow direction and cumulative energy, risk flow paths with cumulative energy exceeding a threshold can be identified, clarifying potential high-risk pathological pathways in diabetes and providing a basis for early intervention. Using high-risk pathways, the impact of pathological nodes on downstream nodes can be deduced, thereby predicting the possible evolution of diabetes into different types of complications and achieving personalized risk trajectory prediction. By establishing a causal relationship structure between pathological mechanism element nodes through a knowledge graph and using graph neural networks for propagation analysis, comprehensively considering the coupling strength and self-repair capabilities between nodes, the dynamic evolution process of diabetic complications can be simulated more accurately. Dynamic modeling and cumulative risk analysis can effectively distinguish between short-term fluctuations and long-term risk accumulation, avoiding misjudgments caused by traditional single-indicator or static threshold methods and improving the reliability of clinical decision-making.

[0036] Methods for generating personalized risk trend curves for diabetes complications include: The evolutionary trajectory of diabetic complications is extracted into the pathological nodes corresponding to the risk flow conservation graph. Based on the flow rate, direction and cumulative energy between pathological nodes, a set of node feature vectors is constructed and linearly transformed to obtain risk features. The similarity of the node feature vectors of adjacent pathological nodes is calculated by cosine similarity. The calculated similarity values ​​are normalized (e.g., by the Softmax function) to obtain the attention weight coefficient between adjacent pathological nodes. The risk features of adjacent pathological nodes are weighted and aggregated using normalized attention weight coefficients to form weighted aggregated risk features. A risk feedback matrix is ​​constructed based on the weighted aggregated risk features. The risk feedback matrix is ​​then subjected to time-series fitting and smoothing to generate individualized risk trend curves for diabetic complications, reflecting the potential complication development trends of diabetic patients at different time stages.

[0037] The preset potential pathological energy threshold is set by staff. By collecting the potential pathological energy of different potential pathological activation nodes, the average value of the potential pathological energy of multiple potential pathological activation nodes is taken as the preset potential pathological energy threshold.

[0038] This embodiment introduces a pathological potential function to integrate the dynamic change curves of potential pathological activation nodes over time, calculating risk potential energy and achieving cumulative risk modeling. It can distinguish whether a node is in an accumulation state, a equilibrium state, or a self-repair state, providing more refined temporal information for risk management. By comparing the risk potential energy with a preset threshold, potential pathological activation nodes are marked as pathological activation source nodes, clarifying the starting point of risk diffusion and providing a basis for individualized intervention. Networked modeling can depict the complex coupling relationships between pathological mechanism elements, improving the accuracy of risk projection. The system integrates the dynamic changes of physiological indicators, the causal relationships between nodes, and the cumulative risk potential energy, enabling more accurate identification of risk nodes and avoiding misjudgments caused by single indicators or static thresholds.

[0039] By integrating the instantaneous risk intensity of a node over a time interval, the cumulative risk of that node can be obtained, enabling modeling of the cumulative energy of risk in the network. Based on the direction of risk flow and cumulative energy, risk flow paths with cumulative energy exceeding a threshold can be identified, clarifying potentially high-risk pathological pathways in diabetes and providing a basis for early intervention. Using high-risk pathways, the impact of pathological nodes on downstream nodes can be deduced, thereby predicting the possible evolution of diabetes into different types of complications and achieving personalized risk trajectory prediction. By establishing a causal relationship structure between nodes of pathological mechanism elements through a knowledge graph and using graph neural networks for propagation analysis, comprehensively considering the coupling strength and self-repair capabilities between nodes, the dynamic evolution process of diabetic complications can be simulated more accurately. Dynamic modeling and cumulative risk analysis can effectively distinguish between short-term fluctuations and long-term risk accumulation, avoiding misjudgments caused by traditional single-indicator or static threshold methods and improving the reliability of clinical decision-making.

[0040] Example 2

[0041] Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. A method for extrapolating the risk of diabetic complications based on graph neural networks and knowledge graphs is provided, including: S1. Collect physiological perception data of diabetic patients, perform time synchronization and normalization processing on the physiological perception data, and separate the stable feature components and the perturbation feature components. S2. Based on the semantic structure of knowledge graph, stable feature components are mapped to basic physiological nodes, and perturbation feature components are mapped to potential pathological activation nodes. A pathological potential function is introduced to establish a pathological potential mapping network. S3. Calculate the potential pathological energy of each potential pathological activation node in the pathological potential mapping network through a graph neural network. When the potential pathological energy is greater than the preset potential pathological energy threshold, it is marked as a pathological activation source node. S4. Construct a risk flow conservation graph based on pathological activation source nodes, regard diabetes-related risks as energy flows propagating in the risk flow conservation graph, calculate the flow rate, direction and cumulative energy of risks between different pathological activation source nodes, and deduce the evolution trajectory of diabetes complications. S5. Using graph attention mechanism, the evolution trajectory of diabetic complications is weighted and converged to generate an individualized risk trend curve for diabetic complications.

[0042] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0043] The above description is merely a preferred embodiment of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for users of ordinary technical skills, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A diabetes complication risk deduction system based on a graph neural network and a knowledge graph, characterized in that, The method comprises the following steps: a physiological characteristic sensing module collects physiological sensing data of a diabetic patient, performs time synchronization and normalization processing on the physiological sensing data, and separates stable characteristic components and disturbance characteristic components; a pathological potential mapping module maps the stable characteristic components to basic physiological nodes and the disturbance characteristic components to potential pathological activation nodes based on the semantic structure of a knowledge graph, introduces a pathological potential function, and establishes a pathological potential mapping network; a pathological activation recognition module calculates the potential pathological energy of each potential pathological activation node in the pathological potential mapping network through a graph neural network, and marks a potential pathological activation node as a pathological activation source node when the potential pathological energy is greater than a preset potential pathological energy threshold; a risk conservation deduction module constructs a risk flow conservation graph according to the pathological activation source node, regards the risks related to diabetes as energy flow propagating in the risk flow conservation graph, calculates the flow rate, direction and cumulative energy of the risks between different pathological activation source nodes, and deduces the evolution trajectory of the diabetic complications; a risk feedback evolution module generates an individualized diabetic complication risk trend curve by using a graph attention mechanism to weight and converge the evolution trajectory of the diabetic complications.

2. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 1, characterized in that, The physiological sensing data of the diabetic patient includes glucose metabolism physiological data, cardiovascular physiological data, nervous system sensing data, kidney function data and life behavior data.

3. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 2, characterized in that, The method for performing time synchronization and normalization processing on the physiological sensing data comprises the following steps: time synchronization is performed on the collected physiological sensing data by using a sliding time window to form a time-synchronized physiological signal matrix; and normalization processing is performed on the physiological signal matrix to obtain a set of physiological sensing feature vectors with uniform structure.

4. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 3, characterized in that, The method for separating the stable characteristic components and the disturbance characteristic components comprises the following steps: feature decomposition is performed on the set of physiological sensing feature vectors to separate the stable characteristic components and the disturbance characteristic components; a weighted moving average and principal component analysis are combined to extract the low-frequency change part of the physiological sensing feature vectors as the stable characteristic components; and a high-pass filter and residual analysis are combined to extract the high-frequency change part of the physiological sensing feature vectors as the disturbance characteristic components.

5. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 4, characterized in that, The method for constructing the pathological potential mapping network comprises the following steps: pathological mechanism elements of diabetes are extracted from a preset database as objects for constructing nodes and edges of a knowledge graph; the extracted pathological mechanism elements are abstracted as nodes and causal relationships between nodes in the knowledge graph to form a semantic hierarchical structure; the nodes of the knowledge graph include basic physiological nodes and potential pathological activation nodes, the stable characteristic components are mapped to the basic physiological nodes of the knowledge graph, and the disturbance characteristic components are mapped to the potential pathological activation nodes; a pathological potential function is introduced to perform integral modeling on the dynamic change curve of each potential pathological activation node within a preset time interval, and calculate the risk potential energy accumulated by each potential pathological activation node in the time dimension; when the risk potential energy of a potential pathological activation node is greater than zero, the potential pathological activation node is in an accumulation state; when the risk potential energy is equal to zero, the potential pathological activation node is in a balance state; and when the risk potential energy is less than zero, the potential pathological activation node is in a self-repairing state. The risk potential accumulated by each potential pathological activation node in the time dimension is assigned to the corresponding node; the nodes, edges and risk potential of the knowledge graph are integrated to form a complete pathological potential mapping network.

6. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 5, characterized in that, The method for obtaining the potential pathological energy of each potential pathological activation node comprises: The risk potential accumulated by each potential pathological activation node in the time dimension is taken as a node feature, and the node feature further comprises a node type and a historical physiological indicator; each node feature is weighted and fused with neighbor node features of the node by using a GNN graph neural network through iterative message passing; the node feature is updated through a nonlinear mapping function to generate a state representation of the node at a current iteration round; The message passing and node feature updating iteration are repeatedly performed until a preset iteration round is reached, and the node feature after the final iteration is converted into a scalar value through an output mapping function to obtain the potential pathological energy of each potential pathological activation node.

7. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 6, characterized in that, The method for marking the pathological activation source node comprises: A preset potential pathological energy threshold is set, and the potential pathological energy of each potential pathological activation node is compared with the preset potential pathological energy threshold; when the potential pathological energy is greater than the preset potential pathological energy threshold, the potential pathological activation node is marked as a pathological activation source node.

8. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 7, characterized in that, The method for constructing the risk flow conservation graph comprises: The marked pathological activation source node is taken as a set of pathological nodes of the risk flow conservation graph, the direction of a directed edge between the pathological nodes in the risk flow conservation graph is determined by using the causal relationship and the semantic hierarchical structure between the extracted pathological mechanism elements in the knowledge graph, and the risk flow conservation graph is formed according to the set of pathological nodes and the directed edge.

9. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 8, characterized in that, The method for deducing the evolution trajectory of the diabetic complication comprises: The risk potential corresponding to the pathological node at an initial time point is taken as an instantaneous risk intensity of the pathological node, a risk flow conservation equation is constructed, the change rate of the instantaneous risk intensity of any pathological node in the risk flow conservation graph is calculated, the risk flow conservation equation is iteratively solved, the risk flow rate and direction between the pathological nodes are updated in real time, and a complete risk flow path is obtained; The instantaneous risk intensity of the pathological node is integrated in a preset time interval to obtain the accumulated energy of the risk between the pathological nodes; according to the direction and the accumulated energy of the risk between the pathological nodes, a risk flow path with accumulated energy greater than a preset accumulated energy threshold is identified as a high-risk path; and the evolution trajectory of the diabetic complication evolving into different diabetic complications is predicted according to the high-risk path.

10. The graph neural network and knowledge graph based risk inference system of diabetic complications according to claim 9, characterized in that, The method for generating the individualized diabetic complication risk trend curve comprises: The pathological nodes corresponding to the evolution trajectory of the diabetic complication in the risk flow conservation graph are extracted, a node feature vector set is constructed based on the flow rate, direction and accumulated energy between the pathological nodes, and linear transformation is performed to obtain a risk feature; the node feature vectors of adjacent pathological nodes are calculated for similarity by using cosine similarity, and the calculated similarity values are normalized to obtain attention weight coefficients between the adjacent pathological nodes. The normalized attention weight coefficient is used to weight and aggregate the risk features of adjacent pathological nodes to form weighted and aggregated risk features; and a risk feedback matrix is constructed based on the weighted and aggregated risk features, and the risk feedback matrix is subjected to time sequence fitting and smoothing processing to generate an individualized diabetes complication risk trend curve.

Images