A cross-institutional diabetic complication prediction method and system based on graph-enhanced medical information architecture

Abstract

This invention relates to the field of medical information technology, specifically to a cross-institutional method and system for predicting diabetes complications based on a graph-augmented medical information architecture. The method includes constructing a graph-augmented medical information architecture for multi-center medical scenarios; standardizing and mapping heterogeneous medical data across institutions and aligning it with medical semantics to construct a heterogeneous medical knowledge graph with unified semantic representation; employing a relation-aware projection mechanism to perform graph representation learning on the knowledge graph, obtaining deep multimodal node representations; constructing a relation-aware multi-task graph attention network, inputting node representations into the network for graph attention aggregation and multi-task collaborative learning to achieve joint prediction of multiple complications of type 2 diabetes; generating complication risk evolution paths to enable interpretable analysis of the prediction results. This invention enables unified organization and collaborative utilization of cross-institutional medical data, effectively improving the accuracy, generalization ability, and interpretability of diabetes complication prediction in multi-center scenarios.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, specifically to a cross-institutional method and system for predicting diabetic complications based on a graph-enhanced medical information architecture. Background Technology

[0002] With the rapid development of medical informatics, intelligent clinical decision-making, and multi-center collaborative research, the standardized integration, knowledge sharing, and joint modeling of cross-institutional medical data have become crucial foundations for the construction of smart healthcare systems. Particularly in the scenario of predicting the risk of complications in type 2 diabetes, electronic medical records, diagnostic records, laboratory test results, and medication information contain a wealth of clinical knowledge. Achieving unified organization and collaborative analysis across institutions would help improve the accuracy of disease risk prediction and enhance decision support capabilities. However, various issues commonly exist among different medical institutions, such as inconsistent field naming, heterogeneous coding systems, differences in units of measurement, sample distribution bias, and scarcity of low-frequency complication samples. These issues make it difficult for existing methods to achieve high-quality integration, joint prediction, and interpretable inference within a unified framework, resulting in low complication prediction accuracy, insufficient model generalization ability, and limited clinical application value.

[0003] Existing medical interoperability standards (such as FHIR) primarily address the exchange of medical data at the syntactic and structural levels. However, when dealing with heterogeneous clinical concepts across multiple institutions, they still lack unified deep semantic representations, explicit entity alignment, and cross-institutional knowledge fusion mechanisms. Specifically, significant differences exist among medical institutions in data collection, encoding, and storage. Even when exchanging data following the FHIR standard, it remains difficult to achieve semantic alignment and fusion of clinical concepts across different institutions. This results in a lack of unified clinical knowledge representation across institutions, hindering the provision of high-quality graph structure inputs for deep learning models. Furthermore, traditional disease prediction methods often employ a single-task independent modeling approach, treating each complication as an independent prediction task. This fails to fully utilize the comorbidity, pathological cascades, and risk propagation paths among various diabetic complications. Consequently, the model cannot capture the intrinsic connections between complications, limiting overall prediction accuracy. Moreover, the prediction process is typically black-box inference, lacking traceable medical explanations and failing to meet the rigid interpretability requirements of clinical decision support. Furthermore, in multi-center clinical application scenarios, the differences in patient composition, disease distribution, and data collection methods among different medical institutions lead to significant data distribution bias. The scarcity of low-frequency complication samples further exacerbates class imbalance. Traditional deep learning models are prone to local overfitting in such scenarios and struggle to learn robust features that can be stably shared across institutions. Consequently, the prediction accuracy in the target institution decreases significantly, severely restricting the performance and clinical applicability of cross-institutional collaborative prediction. Summary of the Invention

[0004] The purpose of this invention is to propose a cross-institutional method and system for predicting diabetes complications based on a graph-enhanced medical information architecture. This method enables unified organization and collaborative utilization of cross-institutional medical data, improving the accuracy, generalization ability, and interpretability of diabetes complication prediction in multi-center scenarios.

[0005] According to a first aspect of the embodiments of this disclosure, a cross-institutional method for predicting diabetes complications based on a graph-enhanced medical information architecture is provided, comprising the following steps: A graph-augmented medical information architecture for multi-center medical scenarios is constructed, comprising a data access layer, a standard semantic layer, a graph construction layer, a graph representation learning layer, and a risk interpretation layer. Based on Fast Healthcare Interoperability Resources, LOINC, and pre-defined standard terminology, cross-institutional medical datasets are standardized and mapped and aligned with medical semantics. After entity-relation extraction, entity alignment, and knowledge fusion, an adjacency matrix A of cross-institutional medical entities and an initial feature vector X of patient nodes are generated. These are then fused to construct a heterogeneous medical knowledge graph G with unified semantic representation. A relation-aware projection mechanism is used to map and transform the node feature space of the heterogeneous medical knowledge graph G. Relationship association weights are applied and updated based on the attribute differences of different medical entity categories and their connection relationships to obtain a deep multimodal node representation H. A relation-aware multi-task graph attention network model M is constructed. The deep multimodal node representation H is input into the network model M for graph attention aggregation and collaborative computation. A comorbidity association constraint mechanism is introduced in the multi-task collaborative learning process of the network model M to obtain the joint predicted risk probability y of various type 2 diabetes complications. Based on the attention weight distribution α generated by the relation-aware multi-task graph attention network model M during collaborative reasoning and the joint predicted risk probability y, the graph backtracking algorithm based on attention weights is used to extract key clinical feature nodes and edges with association weights greater than a preset threshold θ from the heterogeneous medical knowledge graph G, and to generate an evolutionary risk path R from the basic case node to the early warning complication node.

[0006] In one embodiment, the standardized mapping and medical semantic alignment method is as follows: For the first Original clinical fields from medical institutions Perform standardization processing to obtain standardized fields. And establish a semantic mapping relationship between local fields and standard medical entities: in, Represents a standardized field. Represents a standard physical key. Indicates the standard resource type, The entity category is represented; based on the mapping relationship, patient entities, indicator entities, disease entities, and medication entities are extracted, and heterogeneous original fields from different medical institutions are uniformly mapped to the same standard entity space.

[0007] In one embodiment, the entity alignment method is as follows: Let the standard entity sets in the first medical institution and the second medical institution be respectively and The cross-organizational shared entity set is defined as follows: Cross-organizational entity alignment relationships are represented as follows: in, Indicates the type of medical entity; This indicates the clinical concept corresponding to the primary medical institution. Entities; This indicates the clinical concept corresponding to the second medical institution. The entities are considered equivalent entities when medical entities in different medical institutions have the same standard entity key and the same entity type. For medical entities that do not meet the entity alignment conditions, the identifier of the medical institution to which they belong is retained as the institution's private entity.

[0008] In one embodiment, the knowledge fusion method is as follows: Based on the entity alignment results, a unified knowledge representation is established for entities shared across institutions, and an institution-specific knowledge representation is established for medical entities that have not completed entity alignment. Furthermore, the structural relationships between patients and medical entities, the co-occurrence relationships between medical entities, the statistical association relationships, and the influence relationships between medical entities and complication tasks are integrated to construct a fused heterogeneous medical knowledge graph. Node set Defined as: edge set Defined as: in, These represent sets of patient nodes from different medical institutions; This represents a set of shared medical entity nodes across institutions; Each represents a set of private entity nodes for a medical institution. Represents the set of patient feature value nodes; This represents the set of task nodes related to complications.

[0009] In one embodiment, the relation-aware projection mechanism is as follows: Let the node... The initial feature vector is The node type is ,side The characteristics are First, feature recalibration is performed on the node features to obtain: in, and For learnable parameter matrix, The activation function is represented; subsequently, the node features and edge features are projected onto the unified latent space respectively: in, and These represent the node projection matrix and the edge projection matrix, respectively. This indicates node type embedding. In one embodiment, the graph attention aggregation method is as follows: Get Nodes For neighboring nodes Relationship-aware attention weight : in, Represents a node For nodes In the Relationship-aware attention score within the layer; Represents an exponential function; Represents a node The set of neighboring nodes; Represents the set of neighboring nodes Any node in; Representing neighboring nodes For nodes In the Relationship-aware attention scores within the layer; through the above normalization calculation, all pointers to nodes are calculated... The sum of the attention weights of the neighboring nodes is 1, thus reflecting the different attention of the neighboring nodes to the node. The relative contribution level is determined; and neighborhood weighted aggregation is performed based on the relationship-aware attention weights to obtain the node update representation. : in, Represents the linear rectification activation function; Indicates a node All neighboring nodes Perform summation and aggregation; Represents a node Message representation after relational type transformation; Represents a node In the The original representation in the layer is used to form residual updates together with the neighborhood aggregation results; through the above method, joint propagation of heterogeneous nodes and multi-relation edges and deep representation learning are realized to generate deep multimodal node representations. .

[0010] In one embodiment, the prediction and optimization method of the relationship-aware multi-task graph attention network model M is as follows: Task set for complications For the first The risk prediction probability for each complication task is expressed as follows: in, Indicates the patient In the mission The specific representation of the task below, Indicates task The corresponding prediction head, Let represent the Sigmoid function; the relation-aware multi-task graph attention network model is optimized using a multi-task joint loss function, the expression of which is: in, Indicates the first The loss weight of each task, Indicates the first The classification loss function corresponding to each task.

[0011] According to a second aspect of the present disclosure, a cross-institutional diabetes complication prediction system based on a graph-enhanced medical information architecture is provided, comprising: The architecture building module constructs a graph-augmented medical information architecture for multi-center medical scenarios. The graph-augmented medical information architecture includes a data access layer, a standard semantic layer, a graph construction layer, a graph representation learning layer, and a risk interpretation layer. The standardization and graph construction module, based on Fast Healthcare Interoperability Resources, LOINC and preset standard terms, standardizes and maps cross-institutional medical datasets and aligns them with medical semantics. After entity-relation extraction, entity alignment and knowledge fusion, it generates the association adjacency matrix A of medical entities across institutions and the initial feature vector X of patient nodes. The two modules are then fused to form a heterogeneous medical knowledge graph G with unified semantic representation. The relation-aware representation learning module uses a relation-aware projection mechanism to map and transform the node feature space of the heterogeneous medical knowledge graph G. It performs relation association weighting and updating based on the attribute differences of different medical entity categories and their connection relationships to obtain a deep multimodal node representation H. The multi-task joint prediction module constructs a relation-aware multi-task graph attention network model M, inputs the deep multimodal node representation H into the network model M for graph attention aggregation and collaborative computation, and introduces a comorbidity association constraint mechanism in the multi-task collaborative learning process of the network model M to obtain the joint prediction risk probability y of various type 2 diabetes complications; The interpretable path reasoning module, based on the attention weight distribution α generated by the relation-aware multi-task graph attention network model M during collaborative reasoning and the joint predicted risk probability y, uses an attention weight-based graph backtracking algorithm to extract key clinical feature nodes and edges with association weights greater than a preset threshold θ from the heterogeneous medical knowledge graph G, and generates an evolutionary risk path R from the basic case node to the early warning complication node.

[0012] According to a third aspect of the present disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and running on the memory, wherein the processor executes the program to implement the aforementioned method for predicting cross-institutional diabetes complications based on a graph-augmented medical information architecture.

[0013] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the aforementioned method for predicting cross-institutional diabetes complications based on a graph-augmented medical information architecture.

[0014] The advantages of the above technical solutions adopted in this invention compared with the prior art are as follows: 1. This invention proposes a graph-enhanced medical information architecture. This architecture achieves syntactic layer unification of heterogeneous medical data across institutions through FHIR standardized mapping. On this basis, through unified semantic representation, deterministic entity alignment, and multi-relational knowledge fusion, it maps semantically inconsistent clinical concepts in different medical institutions to the same standard entity space. While preserving the private knowledge of each institution, it realizes the unified organization and collaborative expression of shared knowledge across institutions, effectively solving the problems of semantic inconsistency and knowledge silos in multi-center medical data, and significantly improving the interoperability and sharing capabilities of cross-institutional medical data.

[0015] 2. This invention proposes a relation-aware projection mechanism. This mechanism explicitly models the semantic differences of different medical entity types and multiple relation types during graph representation learning. It maps heterogeneous node features and edge relation features to a shared latent space through a learnable projection matrix. Relationship-aware attention propagation is introduced to achieve joint updates of node information, relation information, and type information. This effectively enhances the representation ability of deep semantic relations in complex medical topologies and significantly improves the modeling effect of graph neural networks on heterogeneous medical knowledge.

[0016] 3. This invention constructs a Relationship-Aware Multi-Task Graph Attention Network (RA-MTGAT). This network introduces a comorbidity constraint mechanism during model training by sharing a graph representation foundation and a multi-task collaborative learning framework. It achieves joint risk prediction of multiple diabetic complications and can fully explore the potential comorbidity relationships and pathological transmission paths between different complications. Compared with the traditional single-task independent modeling method, it effectively improves the overall prediction accuracy of multi-tasks and the generalization ability of the model in cross-institutional scenarios.

[0017] 4. This invention constructs an interpretable path reasoning mechanism based on graph attention weights. This mechanism extracts key clinical feature nodes and propagation paths of target complications from the patient's basic information by back-analyzing the attention weight distribution learned by the model. This realizes the transformation of prediction results from "black box inference" to "white box traceability", providing intuitive and traceable risk analysis basis for clinical practice, and significantly enhancing the credibility of the model and its value in clinical decision support. Attached Figure Description

[0018] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0019] Figure 1 A flowchart of a cross-institutional method for predicting diabetes complications based on a graph-augmented healthcare information architecture; Figure 2 This diagram illustrates the overall framework of an enhanced healthcare information architecture. Figure 3 Flowchart for standardized mapping and unified semantic representation of cross-institutional medical data; Figure 4 A schematic diagram illustrating cross-institutional entity alignment and knowledge fusion; Figure 5 Structure diagram of the Relationship-Aware Multi-Task Graph Attention Network (RA-MTGAT) model; Figure 6 This is a schematic diagram of complication path reasoning based on attention weights. Detailed Implementation

[0020] The present disclosure will be further described below with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0022] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0023] It should be noted that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and systems according to various embodiments of this disclosure. It should be noted that each block in a flowchart or block diagram may represent a module, segment, or portion of code, which may include one or more executable instructions for implementing the logical functions specified in the various embodiments. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutively represented blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented using a dedicated hardware-based system that performs the specified functions or operations, or using a combination of dedicated hardware and computer instructions.

[0024] Example 1: like Figure 1 As shown, this embodiment provides a cross-institutional method for predicting diabetes complications based on a graph-augmented healthcare information architecture, including the following steps: S1. Construct a graph-augmented medical information architecture for multi-center medical scenarios, wherein the graph-augmented medical information architecture includes a data access layer, a standard semantic layer, a graph construction layer, a graph representation learning layer, and a risk interpretation layer; like Figure 2 As shown, the data access layer receives raw clinical data from different medical institutions; the standard semantic layer performs standardized mapping and medical semantic alignment; the graph construction layer performs entity extraction, relation extraction, entity alignment, and knowledge fusion; the graph representation learning layer performs relation-aware projection and graph attention propagation on the heterogeneous medical knowledge graph; and the risk interpretation layer performs critical path backtracking and interpretable analysis based on the model output. Through this layered architecture, a closed-loop processing of cross-institutional medical data is achieved, from raw heterogeneous phenotypes to a unified graph structure representation, deep feature learning, and risk interpretation output.

[0025] S2. Based on Fast Healthcare Interoperability Resources, LOINC and preset standard terminology, the cross-institutional medical dataset is standardized and mapped and aligned with medical semantics. After entity-relation extraction, entity alignment and knowledge fusion, the association adjacency matrix A of each medical entity across institutions and the initial feature vector X of the patient node are generated and fused to form a heterogeneous medical knowledge graph G with unified semantic representation. Specifically, raw clinical data is obtained from at least two medical institutions, and this raw clinical data includes at least basic patient information, diagnostic information, laboratory test information, medication information, and complication labeling information. For example... Figure 3 As shown, the original field names, field values, coding systems, and units of measurement are standardized. Let the first field be... The original fields in each medical institution are Standardized and then referred to as The semantic mapping relationship from local fields to standard medical entities is then represented as follows: in, Represents a standard physical key. Indicates the standard resource type, This indicates the entity category. Preferably, the standard terminology system includes Fast Healthcare Interoperability Resources and Logical Observation Identifiers Names and Codes. Through this method, heterogeneous original fields from different medical institutions are uniformly mapped to the same standard medical entity space.

[0026] Based on the standardized mapping results, patient entities, indicator entities, disease entities, and medication entities are extracted, and structural relationships between patients and standardized entities are constructed. For example... Figure 5 As shown, a patient feature value node is introduced between the patient node and the standard entity node to simultaneously preserve feature existence and attribute value information. The relationship is represented as follows: in, Representing patient nodes, Indicates that the patient is in a standard physical environment eigenvalue nodes on, This represents a standard entity node. Based on this, further construction is made of entity co-occurrence relationships, entity statistical correlation relationships, and entity influence relationships on complication tasks, forming a multi-relationship heterogeneous graph structure. Preferably, the node set of the heterogeneous knowledge graph is defined as: The edge set is defined as: in, Represents the set of patient nodes. Represents the set of indicator entity nodes. Represents the set of disease entity nodes. Represents the set of entity nodes related to medication use. Represents the set of patient feature value nodes. Represents the set of task nodes related to complications; Indicates structural relationships. Indicates a co-occurrence relationship. Indicates statistical correlation. Indicates the relationship between tasks. Indicates entity alignment relationship.

[0027] Preferably, deterministic entity alignment is performed on standard entities in different medical institutions based on unified standard entity keys and entity category consistency. Let the standard clinical concept sets in the first medical institution and the second medical institution be respectively... and The cross-organizational shared entity set is defined as follows: Cross-organizational entity alignment relationships are represented as follows: in, This indicates the corresponding clinical concept in the primary medical institution. The entity, This indicates the corresponding clinical concept in the second medical institution. After entity alignment is completed, a unified node merging is performed on shared entities. Entities that are not fully aligned retain their institutional source identifiers as institutional private entities. The structural relationships between patients and medical entities, co-occurrence relationships between medical entities, statistical correlation relationships, and influence relationships between medical entities and complication tasks are integrated to construct a unified semantic representation of the fused heterogeneous medical knowledge graph. in: in, and These represent sets of patient nodes in different medical institutions. Represents a set of shared entity nodes. and Each represents a set of private entity nodes of a medical institution. This represents the set of task nodes related to complications. Through the aforementioned knowledge fusion method, both cross-institutional shared knowledge and institution-specific private knowledge are preserved simultaneously within a unified semantic space, enabling the collaborative organization of multi-source heterogeneous medical knowledge.

[0028] Furthermore, an association adjacency matrix is ​​generated based on the fused heterogeneous medical knowledge graph: And the patient node initialization feature matrix: .

[0029] S3. A relation-aware projection mechanism is used to map and transform the node feature space of the heterogeneous medical knowledge graph G. Relationship association weights are weighted and updated according to the attribute differences of different medical entity categories and their connection relationships to obtain a deep multimodal node representation H. Specifically, set nodes The initial feature vector is The node type is ,side The characteristics are First, feature recalibration is performed on the node features, resulting in: in, and For learnable parameter matrix, This represents the activation function. Subsequently, the node features and edge features are projected separately to obtain the initial node representations: The relationship is then represented by edge projection: in, and These represent the node projection matrix and the edge projection matrix, respectively. This represents node type embedding. Through the above mechanism, relation-aware representations of different entity types and relation types are achieved in a unified latent space. Furthermore, during graph attention propagation, for edges... Corresponding relation type Construct relationship-aware attention weights: in, Represents a node For nodes In the Relationship-aware attention score within the layer.

[0030] Furthermore, a weighted aggregation is performed on the neighbor messages to obtain the node update representation: in, Represents a node The message representation after relation type transformation. Through the above method, joint propagation and deep representation learning of heterogeneous nodes and multiple relation edges are achieved, generating deep multimodal node representations. .

[0031] S4. Construct a relation-aware multi-task graph attention network model M, input the deep multimodal node representation H into the network model M for graph attention aggregation and collaborative computation, introduce a comorbidity association constraint mechanism in the multi-task collaborative learning process of the network model M, and obtain the joint predicted risk probability y of various type 2 diabetes complications; Specifically, deep representations of patient nodes are obtained based on attention propagation in a relation-aware graph. And construct a shared patient representation by combining the patient's original clinical characteristics: in, Indicates the patient The original clinical characteristics. Tasks for each complication. Generate task-specific representations through a task-aware mechanism. The task prediction head outputs the probability of complication risk respectively: in, Indicates task The corresponding prediction head.

[0032] In a preferred embodiment, a multi-task joint loss function is used to train the model, and its expression is: in, Indicates the number of tasks involving complications. Indicates the first The loss weight of each task, Indicates the first Each task corresponds to a classification loss function. In a preferred embodiment, the classification loss function employs Focal Loss or a weighted binary cross-entropy loss function to improve the model's prediction performance in imbalanced class scenarios.

[0033] S5. Based on the attention weight distribution α generated by the relation-aware multi-task graph attention network model M during collaborative reasoning and the joint predicted risk probability y, a graph backtracking algorithm based on attention weights is used to extract key clinical feature nodes and edges with association weights greater than a preset threshold θ from the heterogeneous medical knowledge graph G, generating an evolutionary risk path R from the basic case node to the early warning complication node, thereby achieving interpretable reasoning of the prediction results. Figure 6 As shown, the above methods enable unified semantic expression of heterogeneous medical data across institutions, construction of heterogeneous knowledge graphs, joint prediction of multiple complications of diabetes, and interpretable analysis of risk paths without changing the underlying database structure of medical institutions.

[0034] Example 2: This embodiment provides a cross-institutional diabetes complication prediction system based on a graph-augmented medical information architecture, including: The architecture building module constructs a graph-augmented medical information architecture for multi-center medical scenarios. The graph-augmented medical information architecture includes a data access layer, a standard semantic layer, a graph construction layer, a graph representation learning layer, and a risk interpretation layer. The standardization and graph construction module, based on Fast Healthcare Interoperability Resources, LOINC and preset standard terms, standardizes and maps cross-institutional medical datasets and aligns them with medical semantics. After entity-relation extraction, entity alignment and knowledge fusion, it generates the association adjacency matrix A of medical entities across institutions and the initial feature vector X of patient nodes. The two modules are then fused to form a heterogeneous medical knowledge graph G with unified semantic representation. The relation-aware representation learning module uses a relation-aware projection mechanism to map and transform the node feature space of the heterogeneous medical knowledge graph G. It performs relation association weighting and updating based on the attribute differences of different medical entity categories and their connection relationships to obtain a deep multimodal node representation H. The multi-task joint prediction module constructs a relation-aware multi-task graph attention network model M, inputs the deep multimodal node representation H into the network model M for graph attention aggregation and collaborative computation, and introduces a comorbidity association constraint mechanism in the multi-task collaborative learning process of the network model M to obtain the joint prediction risk probability y of various type 2 diabetes complications; The interpretable path reasoning module, based on the attention weight distribution α generated by the relation-aware multi-task graph attention network model M during collaborative reasoning and the joint predicted risk probability y, uses an attention weight-based graph backtracking algorithm to extract key clinical feature nodes and edges with association weights greater than a preset threshold θ from the heterogeneous medical knowledge graph G, and generates an evolutionary risk path R from the basic case node to the early warning complication node.

[0035] The above modules can be deployed on the same device or distributed devices; the division of modules is only a functional logic description and does not limit the specific physical boundaries or implementation order.

[0036] Example 3: An electronic device is provided for running the aforementioned "A cross-institutional method for predicting diabetes complications based on a graph-augmented medical information architecture". The electronic device includes: a processor, a memory, and optional communication interfaces / display devices / input devices, etc.; the memory stores a computer program that can run on the processor, and when the processor executes the program, it implements steps S1 to S5 of the method described in Embodiment 1, specifically including but not limited to: S1. Construct a graph-augmented medical information architecture for multi-center medical scenarios, wherein the graph-augmented medical information architecture includes a data access layer, a standard semantic layer, a graph construction layer, a graph representation learning layer, and a risk interpretation layer; S2. Based on Fast Healthcare Interoperability Resources, LOINC and preset standard terminology, the cross-institutional medical dataset is standardized and mapped and aligned with medical semantics. After entity-relation extraction, entity alignment and knowledge fusion, the association adjacency matrix A of each medical entity across institutions and the initial feature vector X of the patient node are generated and fused to form a heterogeneous medical knowledge graph G with unified semantic representation. S3. A relation-aware projection mechanism is used to map and transform the node feature space of the heterogeneous medical knowledge graph G. Relationship association weights are weighted and updated according to the attribute differences of different medical entity categories and their connection relationships to obtain a deep multimodal node representation H. S4. Construct a relation-aware multi-task graph attention network model M, input the deep multimodal node representation H into the network model M for graph attention aggregation and collaborative computation, introduce a comorbidity association constraint mechanism in the multi-task collaborative learning process of the network model M, and obtain the joint predicted risk probability y of various type 2 diabetes complications; S5. Based on the attention weight distribution α generated by the relation-aware multi-task graph attention network model M during collaborative reasoning and the joint predicted risk probability y, the graph backtracking algorithm based on attention weight is used to extract key clinical feature nodes and edges with association weights greater than a preset threshold θ from the heterogeneous medical knowledge graph G, and to generate an evolutionary risk path R from the basic case node to the early warning complication node.

[0037] The electronic device hardware can be one of a server, personal computer, workstation, industrial controller, edge computing device, or mobile terminal; the processor can be a general-purpose CPU, GPU, NPU, FPGA, or a combination thereof; the memory can be RAM, ROM, flash memory, or disk array. The device can interact with local / remote data storage via a communication interface. The above hardware configuration does not constitute a limitation of the present invention.

[0038] Example 4: A computer-readable storage medium storing a computer program, which, when run on a processor of an electronic device, causes the program to execute the method steps S1 to S5 described in Embodiment 1; the storage medium may be a disk, optical disk, flash memory, solid-state drive, read-only memory, random access memory, or any combination of the above media.

[0039] Those skilled in the art will understand that the modules or steps described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, which can then be stored in a storage device for execution by a computer device. Alternatively, they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. This disclosure is not limited to any particular combination of hardware and software.

[0040] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0041] While the specific embodiments of this disclosure have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of this disclosure. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of this disclosure are still within the scope of protection of this disclosure.

Claims

1. A cross-institutional method for predicting diabetes complications based on a graph-augmented medical information architecture, characterized in that, Includes the following steps: A graph-augmented medical information architecture for multi-center medical scenarios is constructed, comprising a data access layer, a standard semantic layer, a graph construction layer, a graph representation learning layer, and a risk interpretation layer. Based on Fast Healthcare Interoperability Resources, LOINC, and pre-defined standard terminology, cross-institutional medical datasets are standardized and mapped and aligned with medical semantics. After entity-relation extraction, entity alignment, and knowledge fusion, an adjacency matrix A of cross-institutional medical entities and an initial feature vector X of patient nodes are generated. These are then fused to construct a heterogeneous medical knowledge graph G with unified semantic representation. A relation-aware projection mechanism is used to map and transform the node feature space of the heterogeneous medical knowledge graph G. Relationship association weights are applied and updated based on the attribute differences of different medical entity categories and their connection relationships to obtain a deep multimodal node representation H. A relation-aware multi-task graph attention network model M is constructed. The deep multimodal node representation H is input into the network model M for graph attention aggregation and collaborative computation. A comorbidity association constraint mechanism is introduced in the multi-task collaborative learning process of the network model M to obtain the joint predicted risk probability y of various type 2 diabetes complications. Based on the attention weight distribution α generated by the relation-aware multi-task graph attention network model M during collaborative reasoning and the joint predicted risk probability y, the graph backtracking algorithm based on attention weights is used to extract key clinical feature nodes and edges with association weights greater than a preset threshold θ from the heterogeneous medical knowledge graph G, and to generate an evolutionary risk path R from the basic case node to the early warning complication node.

2. The method for predicting diabetic complications across institutions based on a graph-augmented medical information architecture according to claim 1, characterized in that, The standardized mapping and medical semantic alignment method are as follows: For the first Original clinical fields from medical institutions Perform standardization processing to obtain standardized fields. And establish a semantic mapping relationship between local fields and standard medical entities: in, Represents a standardized field. Represents a standard physical key. Indicates the standard resource type, The entity category is represented; based on the mapping relationship, patient entities, indicator entities, disease entities, and medication entities are extracted, and heterogeneous original fields from different medical institutions are uniformly mapped to the same standard entity space.

3. The method for predicting diabetic complications across institutions based on a graph-augmented medical information architecture according to claim 1, characterized in that, The entity alignment method is as follows: Let the standard entity sets in the first medical institution and the second medical institution be respectively and The cross-organizational shared entity set is defined as follows: Cross-organizational entity alignment relationships are represented as follows: in, Indicates the type of medical entity; This indicates the clinical concept corresponding to the primary medical institution. Entities; This indicates the clinical concept corresponding to the second medical institution. The entities are considered equivalent entities when medical entities in different medical institutions have the same standard entity key and the same entity type. For medical entities that do not meet the entity alignment conditions, the identifier of the medical institution to which they belong is retained as the institution's private entity.

4. The method for predicting diabetic complications across institutions based on a graph-augmented medical information architecture according to claim 1, characterized in that, The knowledge fusion method is as follows: Based on the entity alignment results, a unified knowledge representation is established for entities shared across institutions, and an institution-specific knowledge representation is established for medical entities that have not completed entity alignment. Furthermore, the structural relationships between patients and medical entities, the co-occurrence relationships between medical entities, the statistical association relationships, and the influence relationships between medical entities and complication tasks are integrated to construct a fused heterogeneous medical knowledge graph. Node set Defined as: edge set Defined as: in, These represent sets of patient nodes from different medical institutions; This represents a set of shared medical entity nodes across institutions; Each represents a set of private entity nodes for a medical institution. Represents the set of patient feature value nodes; This represents the set of task nodes related to complications.

5. The method for predicting diabetic complications across institutions based on a graph-augmented medical information architecture according to claim 1, characterized in that, The relation-aware projection mechanism is as follows: Let the node... The initial feature vector is The node type is ,side The characteristics are First, feature recalibration is performed on the node features to obtain: in, and For learnable parameter matrix, The activation function is represented; subsequently, the node features and edge features are projected onto the unified latent space respectively: in, and These represent the node projection matrix and the edge projection matrix, respectively. Indicates node type embedding.

6. The method for predicting diabetic complications across institutions based on a graph-augmented medical information architecture according to claim 1, characterized in that, The graph attention aggregation method is as follows: Get Nodes For neighboring nodes Relationship-aware attention weight : in, Represents a node For nodes In the Relationship-aware attention score within the layer; Represents an exponential function; Represents a node The set of neighboring nodes; Represents the set of neighboring nodes Any node in; Representing neighboring nodes For nodes In the Relationship-aware attention scores within the layer; through the above normalization calculation, all pointers to nodes are calculated... The sum of the attention weights of the neighboring nodes is 1, thus reflecting the different attention of the neighboring nodes to the node. The relative contribution level is determined; and neighborhood weighted aggregation is performed based on the relationship-aware attention weights to obtain the node update representation. : in, Represents the linear rectification activation function; Indicates a node All neighboring nodes Perform summation and aggregation; Represents a node Message representation after relational type transformation; Represents a node In the The original representation in the layer is used to form residual updates together with the neighborhood aggregation results; through the above method, joint propagation of heterogeneous nodes and multi-relation edges and deep representation learning are realized to generate deep multimodal node representations. .

7. The method for predicting diabetic complications across institutions based on a graph-augmented medical information architecture according to claim 1, characterized in that, The prediction and optimization method of the relation-aware multi-task graph attention network model M is as follows: Task set for complications For the first The risk prediction probability for each complication task is expressed as follows: in, Indicates the patient In the mission The specific representation of the task below, Indicates task The corresponding prediction head, Let represent the Sigmoid function; the relation-aware multi-task graph attention network model is optimized using a multi-task joint loss function, the expression of which is: in, Indicates the first The loss weight of each task, Indicates the first The classification loss function corresponding to each task.

8. A cross-institutional diabetes complication prediction system based on graph-augmented medical information architecture, characterized in that, include: The architecture building module constructs a graph-augmented medical information architecture for multi-center medical scenarios. The graph-augmented medical information architecture includes a data access layer, a standard semantic layer, a graph construction layer, a graph representation learning layer, and a risk interpretation layer. The standardization and graph construction module, based on Fast Healthcare Interoperability Resources, LOINC and preset standard terms, standardizes and maps cross-institutional medical datasets and aligns them with medical semantics. After entity-relation extraction, entity alignment and knowledge fusion, it generates the association adjacency matrix A of medical entities across institutions and the initial feature vector X of patient nodes. The two modules are then fused to form a heterogeneous medical knowledge graph G with unified semantic representation. The relation-aware representation learning module uses a relation-aware projection mechanism to map and transform the node feature space of the heterogeneous medical knowledge graph G. It performs relation association weighting and updating based on the attribute differences of different medical entity categories and their connection relationships to obtain a deep multimodal node representation H. The multi-task joint prediction module constructs a relation-aware multi-task graph attention network model M, inputs the deep multimodal node representation H into the network model M for graph attention aggregation and collaborative computation, and introduces a comorbidity association constraint mechanism in the multi-task collaborative learning process of the network model M to obtain the joint prediction risk probability y of various type 2 diabetes complications; The interpretable path reasoning module, based on the attention weight distribution α generated by the relation-aware multi-task graph attention network model M during collaborative reasoning and the joint predicted risk probability y, uses an attention weight-based graph backtracking algorithm to extract key clinical feature nodes and edges with association weights greater than a preset threshold θ from the heterogeneous medical knowledge graph G, and generates an evolutionary risk path R from the basic case node to the early warning complication node.

9. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and running thereon, characterized in that, When the processor executes the program, it implements the cross-institutional method for predicting diabetes complications based on a graph-augmented medical information architecture as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements a cross-institutional method for predicting diabetes complications based on a graph-augmented medical information architecture, as described in any one of claims 1-7.

Images