Chronic disease co-disease new incidence risk assessment method, system and equipment based on link prediction, and medium

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a chronic disease co-disease new risk assessment method, system, device and medium, and the method comprises the steps: employing a time sequence ternary group to model a disease new path, introducing a multi-disease co-existence coefficient to quantify the new risk, and carrying out the grading, learning disease network representation by adopting a relational graph convolutional network (R-GCN), and predicting a new risk; secondly, designing a cross-time window joint training strategy, performing hot start training on the model by using embedded representation of historical data, expanding representation information of an encoder for incremental data, fusing the representation information into a disease network, and integrating data to iteratively optimize the model; according to the method, through systematic modeling of the patient medical history data and the disease association network, the potential law of the new occurrence of the common disease is disclosed, and an analysis tool and direction guidance are provided for subsequently exploring the common disease mechanism of a specific disease and identifying a high-risk evolution mode.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and in particular relates to a method, system, device and medium for assessing the risk of new comorbidities of chronic diseases based on link prediction. Background Technology

[0002] Chronic non-communicable diseases (hereinafter referred to as "chronic diseases") do not refer to a specific disease, but rather to a general term for a class of non-communicable diseases characterized by insidious onset, long course, and complex etiology. Comorbidity, as an important developmental feature of chronic diseases, involves complex pathological interactions across multiple factors, systems, and levels, posing a severe challenge to the traditional single-disease-oriented treatment model. Against this backdrop, promoting a shift in the medical model from a disease-centered approach to a health-centered approach, and implementing a comprehensive chronic disease prevention and control strategy that prioritizes prevention and combines prevention with treatment, has become a key direction for the development of the health and medical care sector.

[0003] In this transformation process, how to accurately and personally assess the new risk of chronic disease comorbidities and achieve prospective intervention has become an urgent problem to be solved. Research shows that many chronic diseases share common pathological mechanisms and biological processes at the molecular and physiological levels, constituting the intrinsic basis for the occurrence and development of comorbidities. More importantly, in long-term disease management, the comorbid state of patients is not static; the dynamic changes in their disease control status continuously affect their overall health trajectory, triggering new comorbidity risks. Previous studies have conducted preliminary analyses of new comorbidity phenomena in common chronic diseases such as cardiovascular and cerebrovascular diseases, diabetes, and hypertension, further highlighting the universality and importance of this issue. To address this challenge, machine learning and network analysis techniques provide powerful tools for uncovering complex relationships between diseases by revealing hidden attributes in disease networks. Link prediction methods, especially those combined with graph neural network models, can predict the probability of a link forming between two points in a network based on a known network structure, demonstrating their potential in the field of disease prediction. However, existing research still has limitations in addressing the dynamic evolution of comorbidities and the continuous growth of data in real-world scenarios: First, the problem of new comorbidities is becoming increasingly prominent, but traditional link prediction methods are limited by static analysis of network topology, making it difficult to capture the temporal patterns of dynamic new occurrences and evolution of diseases as they progress; Second, medical data is growing explosively, but models usually rely on fixed historical data for training and lack the ability to integrate and update incremental medical data, resulting in insufficient generalization and timeliness in long-term health management scenarios; Third, existing risk studies are mostly focused on endpoint events such as death or serious complications, and lack risk assessment of "what new chronic diseases will an individual develop in the future under existing comorbidities."

[0004] To address the aforementioned issues, this invention proposes a method for assessing the risk of new comorbidities in chronic diseases based on link prediction. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device, and medium for assessing the risk of new comorbidities of chronic diseases based on link prediction, so as to solve the problems existing in the prior art.

[0006] To achieve the above objectives, this invention provides a method for assessing the risk of new comorbidities in chronic diseases based on link prediction, comprising:

[0007] S1: Obtain the CHARLS dataset; the CHARLS dataset includes the types, times, and descriptions of chronic diseases; perform multi-stage probability sampling on the CHARLS dataset based on different time scales to obtain the source dataset and the new dataset;

[0008] S2: The source dataset is split based on the order of onset of disease. Existing disease changes are encoded as head entities, and newly added diseases are encoded as tail entities. Relationships and relative temporal sequences are established between the head and tail entities to complete the construction of the temporal triplet structure. The relationship is the risk level of new onset.

[0009] S3: Based on different new risk levels, perform stratified sampling of the temporal triplet structure in the source dataset to obtain the training set;

[0010] S4: Construct a first prediction model and a second prediction model based on a relational graph neural network, input the training set into the first prediction model to predict the risk of new occurrences, and train it based on the target loss function;

[0011] S5: Based on the new dataset, the encoding representation system corresponding to the source dataset is expanded to obtain a joint training set;

[0012] S6: Use the entity and relation embedding vectors corresponding to the trained first prediction model as the initialization parameters of the second prediction model; train the second prediction model based on the joint training set to obtain the trained second prediction model;

[0013] S7: Perform a new risk assessment task for chronic disease comorbidities based on the trained second prediction model.

[0014] Optionally, the process of classifying the risk level of new cases in step S2 is as follows:

[0015] A multi-disease coexistence coefficient is introduced to quantify the risk of new cases and classify them into levels.

[0016] Optionally, the training process of the first prediction model in S4 specifically includes:

[0017] For different new risk levels, relational subgraphs are constructed and trained to achieve deep learning of each risk feature;

[0018] By aggregating the results of different relational subgraphs and combining them with the objective loss function, the disease modeling under different risks is iterated repeatedly to complete the training of the model.

[0019] Optionally, step S5 specifically includes:

[0020] Each entity in the new dataset is examined to distinguish between known entities and newly added entities. The original encoding of known entities is retained, and new encoding is appended to newly added entities. The encoded entities are then added to the encoding representation system to obtain a joint training set.

[0021] Optionally, step S7 specifically includes:

[0022] Obtain the head entity set representing known disease changes in the data to be predicted. Combine the head entity set with the data corresponding to each new risk level by performing a Cartesian product to form all possible ordered pairs. Input these pairs into the trained second prediction model for prediction and output the top three possible tail entities.

[0023] Calculate the link prediction score, normalize all scores, and then filter out time-series triple structures with scores greater than 0.5 to form candidate triples;

[0024] Input unknown entities that are not yet connected in the data to be predicted into the model, and output a unique risk with reference value to check the uniqueness of risk for candidate triples;

[0025] By removing contradictory data that has already appeared in the head entity in the prediction results, logical consistency screening is achieved, and the final prediction results are obtained.

[0026] On the other hand, to achieve the above objectives, the present invention provides a chronic disease comorbidity risk assessment system based on link prediction, comprising:

[0027] The data acquisition module is used to acquire the CHARLS dataset, which includes the types, times, and descriptions of chronic diseases. Multi-stage probability sampling is performed on the CHARLS dataset based on different time scales to obtain the source dataset and the new dataset.

[0028] The triplet construction module is used to split the source dataset based on the order of disease onset, encode existing disease changes as head entities, encode new diseases as tail entities, establish relationships and relative temporal order between head and tail entities, and complete the construction of the temporal triplet structure; the relationship is the new risk level;

[0029] The basic model training module is used to perform stratified sampling of the time-series triplet structure in the source dataset according to different new risk levels to obtain a training set; a first prediction model and a second prediction model are constructed based on a relational graph neural network, the training set is input into the first prediction model to predict the new risk, and training is performed based on the target loss function;

[0030] The joint training module is used to expand the encoding representation system corresponding to the source dataset according to the new dataset to obtain a joint training set; use the entity and relation embedding vectors corresponding to the trained first prediction model as the initialization parameters of the second prediction model; and train the second prediction model based on the joint training set to obtain the trained second prediction model.

[0031] The risk prediction module is used to perform a new risk assessment task for chronic disease comorbidities based on the trained second prediction model.

[0032] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the aforementioned link prediction-based method for assessing the risk of new comorbidities of chronic diseases.

[0033] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for assessing the risk of new comorbidities of chronic diseases based on link prediction.

[0034] The technical effects of this invention are as follows:

[0035] This invention constructs a framework for assessing the risk of new onset of chronic disease comorbidities based on link prediction. This framework integrates temporal features to construct triples and trains a graph neural network model to predict the risk of new onset of diseases under existing comorbidities. To integrate incremental medical data and achieve continuous model optimization, this invention further employs a joint training mechanism of encoder iterative optimization and model parameter representation embedding, enabling the model to adapt to new data, iterate parameters, and update accordingly. Furthermore, case studies preliminarily validate the feasibility and application potential of link prediction technology in chronic disease comorbidity management. Attached Figure Description

[0036] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0037] Figure 1 This is a research framework for assessing the risk of new comorbidities of chronic diseases based on link prediction, as described in this invention.

[0038] Figure 2The number and proportion of various chronic disease samples in the embodiments of the present invention;

[0039] Figure 3 This is an example of disease time-series feature data based on triplet representation in an embodiment of the present invention;

[0040] Figure 4 This is a relational graph convolutional network model for risk level differentiation in this embodiment of the invention;

[0041] Figure 5 This is a schematic diagram of joint training of the model based on cross-time window data in an embodiment of the present invention;

[0042] Figure 6 These are the evaluation results of different models and different dimensions in the CHARLS dataset in the embodiments of the present invention;

[0043] Figure 7 This describes the prediction effect of high-risk links in the embodiments of the present invention;

[0044] Figure 8 This refers to the prediction effect of medium-risk links in the embodiments of the present invention;

[0045] Figure 9 This is the prediction effect of low-risk links in the embodiments of the present invention. Detailed Implementation

[0046] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as a limitation of the present invention, but rather as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0047] like Figure 1 - Figure 9 As shown, this embodiment provides a method for assessing the risk of new comorbidities of chronic diseases based on link prediction, including:

[0048] S1: Obtain the CHARLS dataset; the CHARLS dataset includes the types, times, and descriptions of chronic diseases; perform multi-stage probability sampling on the CHARLS dataset based on different time scales to obtain the source dataset and the new dataset;

[0049] S2: The source dataset is split based on the order of onset of disease. Existing disease changes are encoded as head entities, and newly added diseases are encoded as tail entities. Relationships and relative temporal sequences are established between the head and tail entities to complete the construction of the temporal triplet structure. The relationship is the risk level of new onset.

[0050] S3: Based on different new risk levels, perform stratified sampling of the temporal triplet structure in the source dataset to obtain the training set;

[0051] S4: Construct a first prediction model and a second prediction model based on a relational graph neural network, input the training set into the first prediction model to predict the risk of new occurrences, and train it based on the target loss function;

[0052] S5: Based on the new dataset, the encoding representation system corresponding to the source dataset is expanded to obtain a joint training set;

[0053] S6: Use the entity and relation embedding vectors corresponding to the trained first prediction model as the initialization parameters of the second prediction model; train the second prediction model based on the joint training set to obtain the trained second prediction model;

[0054] S7: Perform a new risk assessment task for chronic disease comorbidities based on the trained second prediction model.

[0055] To capture complex disease progression relationships and integrate incremental medical data, this embodiment proposes a link prediction-based framework for assessing the risk of new comorbidities in chronic diseases. First, a temporal triplet model is used to model new disease pathways. A multi-disease coexistence coefficient is introduced to quantify and classify the risk of new occurrences. A relational graph convolutional network (R-GCN) is used to learn the disease network representation to predict the risk of new occurrences. Second, a joint training strategy across time windows is designed. The model is warm-started using the embedded representations of historical data. The encoder's representation information is expanded for incremental data and integrated into the disease network. The model is iteratively optimized by integrating the data. The R-GCN-based method achieves a maximum accuracy of 83.53% on the Hits@3 metric (the proportion of correct answers in the first three predictions), outperforming representation learning and traditional neural networks. The joint training model RGCN-Jo++, which uses representation embedding, performs better in learning incremental data features. The data scope and quantity are relatively limited, and rare diseases are not included. As the amount of new data increases, joint training will consume significant training resources; therefore, incremental learning and other methods can be combined in the future. The link prediction method based on graph neural networks provides strong support for predicting the risk of new comorbidities in chronic diseases.

[0056] Traditional link prediction methods are mainly based on algorithms such as node similarity, path similarity, and random walk similarity. These methods are intuitive and easy to implement, but have poor scalability. In recent years, methods based on representation learning (RL) and neural networks have been widely applied to link prediction problems. The former transforms entities and relations from symbolic representations to numerical representations, capturing the latent characteristics between entities and relations by learning low-dimensional embeddings. It can be divided into translational distance models represented by TransE and tensor decomposition models represented by DistMult and ComplEx. The latter transforms the link prediction task into a deep learning task, converting the link prediction problem into a binary classification task. Compared with traditional neural networks, graph neural networks (GNNs) further improve the predictive power and generalization performance of the model. Its main idea is information propagation, that is, for each node in the graph data, it aggregates the features of its neighbors to better explore the deep relationships between entity nodes. Common models include graph convolutional networks (GCNs), graph attention networks (GATs), and graph autoencoders (GAEs).

[0057] Based on the above research, traditional methods typically assume that the network is static and struggle to adapt to dynamic changes in data. Therefore, this embodiment employs a link prediction method based on graph neural networks to deeply explore the complex relationships among chronic disease comorbidities in graph networks.

[0058] This embodiment aims to break through the traditional static analysis framework and predict the probability of a patient developing a new specific chronic disease on the basis of existing comorbidities from the perspective of longitudinal observation and relative comparison of individual health status. That is, the risk of new comorbidity, and constructs a chronic disease comorbidity risk assessment framework based on link prediction. It is mainly reflected in the following three aspects: (1) Deep integration of temporal features. By explicitly modeling the order of onset, a triple containing temporal information is constructed, and the temporal nodes are isolated to train the model, so as to more accurately capture the path from the current comorbidity state to the future new disease, in order to respond to the prediction needs of "what new diseases will occur in the future". (2) Joint training of historical and incremental data. A method for cross-time window data integration and model iterative update is designed so that the model can integrate incremental medical data and dynamically adjust parameters, in order to solve the challenge of "how to effectively utilize incremental medical data". (3) Classification of comorbidity risk levels. A multi-disease coexistence coefficient is introduced to measure the risk of new comorbidity and classify levels, providing a reference for the management of chronic disease comorbidity risk and providing a new assessment dimension for the existing risk assessment system.

[0059] like Figure 1As shown, this embodiment presents a research framework for assessing the risk of new comorbidities of chronic diseases. It is divided into three stages: basic model training, joint training, and risk assessment. The joint training provides a solution that can dynamically adapt to new data based on the basic model training.

[0060] 1. Training of the basic model that integrates temporal features:

[0061] (1) Construction of triples by splitting the order of disease onset:

[0062] This embodiment uses the China Health and Retirement Longitudinal Study (CHARLS) as the research data, which is available through open-source access at https: / / charls.pku.edu.cn / . This dataset targets Chinese community residents aged 45 and above, and samples were selected using multi-stage probability sampling. Since the pilot survey began in 2008, follow-up has been conducted approximately every two years using Computer-Assisted Personal Interviews (CAPI). The survey covers most common chronic diseases; Table 1 shows 14 of these chronic diseases and their corresponding numbers in this embodiment. The assessment of residents' health status in the data includes changes in pre-existing diseases over a preceding period (including improvement, maintenance, and deterioration), as well as new disease occurrences at the current time point. The sample size and proportion of each type of chronic disease (taking 2018 as an example) are shown below. Figure 2 As shown.

[0063] When the number is 14, it means that no new one of the above diseases has been contracted.

[0064] Based on the preceding changes and current occurrences of the disease, this embodiment constructs a temporal triplet structure, utilizing the relationship between head and tail entities to characterize temporal features. Specifically, as... Figure 3 As shown, the original data containing multiple new diseases is first split into multiple independent data records. On this basis, the changes in the original diseases are encoded as head entities, the new diseases are encoded as tail entities, and a clear correspondence and relative time sequence are established between the two.

[0065] (2) Relationship graph convolutional network model for risk level differentiation: Existing research has divided risk stratification health management into three risk levels: high, medium, and low. This embodiment calculates the risk value based on the multimorbidity coefficient (MC), with the specific formula as follows:

[0066]

[0067] The values ​​of each parameter are shown in Table 2.

[0068] MC measures the relationship between multiple diseases based on disease co-occurrence. If a disease occurs... The probability is At the same time, diseases occur and The probability is ,but This indicates that the probability of two diseases occurring simultaneously is higher than the probability of them occurring independently. Conversely, there may be no significant correlation. Based on this, this embodiment sets the boundary between low risk and medium-high risk to 1. For further subdivision of medium-high risk, since there is no consensus in academia, this embodiment considers interpretability and sample distribution characteristics, setting the threshold for dividing medium-risk and high-risk to 2, meaning the probability of two diseases occurring simultaneously is twice the probability of them occurring independently. After this division, the proportion of high-risk samples in the data over many years has consistently been between 20% and 25%, consistent with empirical observation. In summary, the risk level classification criteria in this embodiment are as follows: risk values ​​between 0 and 1 are low risk, between 1 and 2 are medium risk, and above 2 are high risk. The risk classification results are used for relation representation in triples. When partitioning the dataset, stratified sampling is performed according to risk level, provided that each head entity appears at least once in the training set, to ensure the stability of the category distribution.

[0069] The applicability of Relational Graph Convolutional Network (R-GCN) models for risk level differentiation is mainly reflected in the following two aspects: First, the core of the R-GCN model lies in the discriminative representation and learning of different relationships, thus enabling the decomposition of disease evolution relationships under different risk levels for the same disease. From the model architecture perspective, the R-GCN model explicitly distinguishes various relationships and constructs separate relationship subgraphs for training. Existing research has utilized the discriminative learning characteristics of the R-GCN model to explore heterogeneous network associations between microorganisms, drugs, and diseases. Second, the R-GCN model can aggregate the results of different relationship subgraphs, allowing for iterative disease modeling under different risks. From the model's forward propagation perspective, the final representation of each node and relationship is not isolated but originates from the aggregation results of these different relationship subgraphs, enabling in-depth mining of nodes and relationships. For layers... Nodes in If the relationship type of its association is And its neighbors are The specific formula for the forward propagation of the model is:

[0070]

[0071] in, It is a node In the layer The hidden layer state, It is a weight matrix. It is a normalized parameter. It is an activation function.

[0072] Therefore, this embodiment utilizes the R-GCN mechanism for distinguishing different relationship types, representing different risk levels of multiple chronic diseases as different relationships in the graph, and constructs risk partitioning subgraphs for high-risk, medium-risk, and low-risk conditions (e.g., Figure 4 (As shown). The model learns information in the risk partitioning subgraphs to achieve deep learning of each risk feature, and then aggregates them to generate the final representation of each node and relationship. In addition, in order to maintain the temporal characteristics of the disease triples, this embodiment introduces an isolation mechanism for the head entity (i.e., the preceding node) and the tail entity (i.e., the following node), and strictly restricts the connection of nodes in the same temporal space to ensure that the relative order of nodes is maintained.

[0073] To evaluate the model's performance, this embodiment uses the differences in modeling approaches as the dividing criterion, selecting classic models from the perspectives of translation distance, tensor decomposition, and traditional neural network models for comparison (see Table 3) to verify the superiority of the proposed method compared to various modeling paradigms. TransE, being simple, efficient, and highly scalable, is a pioneering work in the field of knowledge graph completion and has been widely validated in numerous studies, providing a reference standard for the basic structure of this embodiment's experiments. DistMult and ConvE excel in symmetric semantic associations (e.g., the correlation between diseases) and nonlinear complex interaction patterns, respectively, and can help verify the modeling capabilities of the proposed method in different aspects.

[0074] 2. Joint training based on cross-time window data:

[0075] In real-world healthcare scenarios, patient health data accumulates continuously over time, with new comorbidities and disease associations constantly emerging. To address this challenge and enable the model to adapt to the dynamically changing healthcare environment and fully utilize historical knowledge, this embodiment designs a joint training strategy based on cross-time window data. The core objective of this strategy is to solve the problem of how to effectively update and integrate knowledge when faced with incremental data containing new entities and relationships.

[0076] Incremental data typically includes two types: one is triples where all entities have appeared in historical data and only new relationships have been established; the other includes new disease sets of entities not recorded in historical data. To integrate both old and new data into a unified modeling framework, the encoder first needs to be iteratively updated to represent the incremental information. The encoder examines all entities in the new data, distinguishing between known entities and newly added entities. Known entities retain their unique encodings obtained during the basic training phase to ensure the continuity of knowledge representation; newly added entities have new encodings appended to their existing encoding sequences, seamlessly integrating them into the existing representation system. This process enables the encoder to be compatible with dynamically growing entity sets.

[0077] After iteratively expanding the encoder, further representation embedding techniques are used to achieve knowledge transfer and rapid model adaptation. For example... Figure 5 As shown, joint training uses the entity and relation embedding vectors obtained after the basic model training converges as the initialization parameters of the new model, including entities and relations known from historical data in the incremental data; for newly added entities, parameters are randomly initialized. The model is fine-tuned based on the joint training set containing new data in a "warm-start" state that incorporates historical knowledge. This approach not only significantly accelerates the training convergence process but also enables the model to robustly maintain disease association patterns learned from historical data while absorbing features from new data, thus achieving robust learning from incremental data.

[0078] In summary, through a joint training mechanism combining encoder iteration and representation embedding, the model constructed in this embodiment possesses the ability to continuously learn and dynamically evolve. To verify the effectiveness of this strategy, this embodiment concatenates the training and validation sets to form a joint training set and a joint validation set, respectively. For the test set, the independence of the source test set and the new test set is maintained for subsequent separate testing. The following section will compare and analyze the joint training model without representation embedding with the model using representation embedding, and compare the results with the experimental model in Table 3, thereby comprehensively verifying the joint training effect and the model's ability to learn from new data.

[0079] 3. Risk assessment of new comorbidities in chronic diseases based on link prediction

[0080] This embodiment assesses the risk of link prediction for newly emerging unknown relationships in chronic disease comorbidities, aiming to uncover latent relationships and improve existing knowledge. First, it obtains the head entity set representing known disease changes in the dataset and performs a Cartesian product with each risk set to form all possible ordered pairs. These pairs are input into a trained model for prediction, outputting the top three possible tail entities and calculating link prediction scores. After normalizing all scores, triples with scores greater than 0.5 are selected as candidate triples. Further screening of candidate triples involves two aspects: First, due to the Cartesian product result, the same disease change may output the same result with different risks. To prevent contradictions between different risks in the output results, this embodiment inputs unconnected unknown entity pairs from the dataset into the model, outputting unique risks with reference value to check the uniqueness of risk for candidate triples. Second, since "existing diseases" cannot be "added" again, contradictory data already appearing in the head entities in the prediction results need to be removed to achieve logical consistency screening, thus obtaining the final prediction result. Combining these steps, the model can provide personalized disease risk prediction results for specific disease changes. Based on this, the risk is divided into high, medium and low risk, and the number of chronic diseases comorbidities in the prediction results is statistically analyzed and visualized, which can realize a refined risk assessment of chronic disease comorbidities.

[0081] 4. Experimental Results and Analysis:

[0082] (1) Model training results: The operating system used in the experimental environment was Windows 11, the server CPU was Intel Core i9-13900K, and the GPUs were NVIDIA GeForce RTX 4090 and NVIDIA GeForce RTX 3090. The development language was Python (V3.10.9), and the underlying architecture was PyTorch framework (V2.1.0+cu118) and DGL (V2.2.1+cu118).

[0083] Experimental Data: This embodiment involves three types of data: historical data (or source data), which refers to the data used in the early stages of model training; incremental data (or new data), which is data newly collected after the baseline time point; and cross-time window data (or joint data), which is a collection of historical data and incremental data, containing complete datasets collected under different observation time windows. Each data collection time point of CHARLS is recorded as one period. This embodiment selects data from the two most recent periods (i.e., 2018 and 2020) as research cases, and records them as source data and new data according to the order of data collection. Through disease decomposition, a total of 25,260 source data entries and 23,599 new data entries were formed. Information on chronic diseases covered in the source data is shown in Table 1. The new data adds "Parkinson's disease" to the source data, but since it accounts for only 0.54% and is relatively small, this embodiment decides to adopt an exclusion strategy for this disease, that is, temporarily ignoring the newly emerging disease and focusing on new changes in existing diseases. During the research, both the source data and the new data were independently split into their respective training, validation, and test sets in an 8:1:1 ratio. The base model was trained and validated using the training, validation, and test sets corresponding to the source data, while the jointly trained model was trained using the fused joint training and validation sets, and validated using the relatively independent source and new test sets.

[0084] Evaluation Metrics: Commonly used evaluation metrics for link prediction include Mean Rank (MR), Mean Reciprocal Rank (MRR), Hits@k, and MAP. MR calculates the average rank of positive examples among all possible entities, MRR calculates the average of the reciprocal ranks of positive examples among all possible entities, and Hits@k calculates the proportion of positive examples ranked within the top k. In link prediction tasks, higher MRR, lower MR, and higher Hits@k indicate better model performance. This embodiment uses four metrics—MRR, MR, Hits@1, and Hits@3—for model evaluation, with the specific formula as follows:

[0085]

[0086]

[0087]

[0088] in, This is an indicator function; it is 1 if the condition is true, and 0 otherwise.

[0089] In this embodiment, the batch size is set to 128 samples, the epoch value is set to 300 rounds, the learning rate is 0.001, and the embedding size of R-GCN and the baseline model ranges from 50 dimensions to 500 dimensions, with training performed once every 50 dimensions.

[0090] Figure 6 The evaluation results of different models across different dimensions of the dataset show that the embedding dimension has a certain impact on model performance. The DistMult and TransE models show similar trends, with the DistMult model slightly outperforming the TransE model overall. The ConvE model performs relatively stably in low dimensions but fluctuates in high dimensions, with a slight decrease in its evaluation metrics. Low-embedded-dimensional R-GCN models struggle to capture sufficient features, resulting in relatively poor performance. However, as the dimension increases from 150 to 200, the R-GCN model's evaluation metrics gradually surpass all other models. High-embedded-dimensional R-GCN models fully leverage the advantages of graph neural networks. Specifically, the 300-dimensional RGCN model achieves a Hits@1 accuracy of 71.62%, and the 450-dimensional RGCN model achieves a Hits@3 accuracy of 83.53%, performing best among all models. Overall, high-dimensional R-GCN models demonstrate good fit in link prediction tasks involving chronic disease comorbidities, achieving a higher level compared to other models.

[0091] Based on the evaluation results of the basic model, this embodiment conducts joint training experiments on the 450-dimensional graph neural network R-GCN model. The model without representation embedding is labeled RGCN-Jo, and the model with representation embedding is labeled RGCN-Jo++. The results are shown in Table 4. Overall, the RGCN-Jo++ model performs better than the RGCN-Jo model and converges faster during training. The models show different performance on the source dataset and the new dataset. RGCN-Jo performs better on the source test set, while RGCN-Jo++, having acquired sufficient prior information, is more inclined to capture new data features. Furthermore, compared with the average results of the basic training model, the joint training, although less effective than the optimal results (see Table 4), still shows better performance. Figure 6 The indicator has declined slightly, but it is still better than the average level of most models, and its advantage lies in its ability to adapt to more complex situations in new data.

[0092]

[0093] (2) Case study of risk assessment for new comorbidities of chronic diseases: This section applies the link prediction method to the field of chronic disease comorbidity for application analysis. It not only demonstrates the application potential of link prediction technology in risk assessment for new comorbidities of chronic diseases, but also provides strong support for the formulation of differentiated medical plans and the optimization of prevention strategies in the future.

[0094] Personalized disease prediction: Using previously trained models, personalized prediction results of potential new risks are output for specific disease changes. The link score represents the relative probability of forming the link. The predicted links with higher scores among various risks are shown in Table 5.

[0095] Note: See Table 1 for disease codes. This indicates that the situation remains unchanged. This indicates that the situation is improving. This indicates that the situation has worsened.

[0096] [Example of two co-occurring diseases]: If you already have hypertension and arthritis and your condition has improved, the risk level of developing new liver disease is medium risk, with a link score of approximately 0.8041.

[0097] [Example of three co-occurring diseases]: If you already have hypertension, liver disease and kidney disease, and the liver disease condition is stable while the others are improving, you are at high risk of developing emotional and mental health problems. Your link score is approximately 0.8565.

[0098] To illustrate the results of this embodiment within a more comprehensive assessment framework, this embodiment compares the prediction results of new comorbidity risk in chronic diseases with the classic Charlson Comorbidity Index (CCI) score. The CCI, as a mature tool for assessing the overall risk of comorbidity to severe clinical outcomes, is primarily designed for severe physical illnesses with a clear impact on short-term prognosis. Many common chronic diseases and mental health problems, while not directly jeopardizing short-term survival, significantly affect quality of life and long-term health trajectory and are not included in the main assessment scope of the CCI. The comparison reveals a significant difference between the high-risk new comorbidity combinations identified in this embodiment and the high-scoring CCI group, demonstrating the complementarity of the two in assessment dimensions. For example, some comorbidity combinations with lower CCI scores (i.e., lower traditional mortality risk) show a higher risk of new emotional or mental health problems under the prediction framework of this embodiment. Therefore, this embodiment, from the perspective of new comorbidity risk, captures potential health evolution risks related to long-term quality of life that are not fully covered by traditional prognostic tools, providing a new assessment dimension for existing risk assessment systems.

[0099] Analysis of comorbidity incidence prediction results: The prediction results are visualized by distinguishing between high, medium, and low risks. Circles and squares represent existing and newly added diseases, respectively. Solid and dashed lines represent known and predicted relationships of comorbidity incidence risks, respectively. The sum of these two relationships is represented by the node size of newly added diseases, while the color intensity of nodes representing existing diseases indicates the number of comorbidities. The Fruchterman-Reingold algorithm is used to optimize the node layout, ensuring that connected nodes attract each other in space.

[0100] ① High-risk. A high-risk relationship refers to a condition with a higher risk of developing a new disease within an existing comorbid state. This type of link can be considered a dangerous disease evolution relationship, requiring high vigilance and timely intervention from patients with such chronic comorbidities. The prediction results for high-risk links are as follows: Figure 7 As shown, the distribution is relatively moderate, with a large proportion of known relationships. Common chronic diseases such as "dyslipidemia," "diabetes," and "gastric diseases" have a large base (i.e., the nodes in the figure are relatively large). This embodiment mainly focuses on the connection density of the orange dashed lines in the figure (i.e., the proportion of the number of node dashed line links to the total number of dashed lines) to analyze the correlation degree of different disease evolution and predict relationships. It can be seen that the predicted relationships are mainly concentrated on "emotional and mental problems," accounting for 30.69%, followed by diseases such as "cancer and other malignant tumors" (18.81%) and "liver diseases" (9.90%).

[0101] As shown in Table 5, the overall prediction scores for high-risk links were relatively high, reaching a maximum of 0.8565. In terms of the number of pre-existing diseases, these links generally had three or more, while malignant tumors such as cancer (numbered 3) corresponded to four or more pre-existing diseases. Regarding the relationships between diseases, emotional and mental health issues were closely linked to the "three highs" (hypertension, hyperlipidemia, and hyperglycemia) (numbered 0, 1, and 2), and multiple "three highs" diseases often co-occurred with these issues.

[0102] Therefore, for the health service provider, it is recommended to increase psychological supervision and emotional counseling when the number of comorbid chronic diseases exceeds three, pay more attention to the patient's emotional and mental problems, and cooperate with regular cancer screening when the number exceeds four to prevent the rapid deterioration of physical condition and the occurrence of fatal risks; for the academic community, research on "brain insulin disease" can be increased.

[0103] ②Medium risk. The prediction results for medium-risk links are as follows: Figure 8 As shown, the predictive relationships are relatively dense. Among them, "dyslipidemia," "liver disease," and "kidney disease" account for a relatively high proportion of the prediction results, at 15.70%, 15.62%, and 13.43%, respectively.

[0104] Overall, the link prediction scores for medium-risk individuals were lower than those for high-risk individuals, but still remained at a high level. Among the top eight results in link prediction scores, liver disease (number 5) appeared most frequently, and patients with liver disease were at risk of developing additional memory-related diseases (number 11) and malignant tumors such as cancer (number 3). Previous studies have focused on the relationship between liver disease and cognitive impairment, but the underlying mechanisms require further investigation.

[0105] Therefore, for the supply side of health services, it is recommended to increase the proportion of liver disease screening when treating multiple diseases, and to increase the depth of popular science education on relevant prevention knowledge for at-risk groups; for the academic community, it is possible to explore the possible disease evolution mediation mechanisms of liver diseases and the potential linkage mechanisms between liver diseases and other diseases.

[0106] ③ Low risk. For example... Figure 8 As shown, the prediction results for low-risk links are relatively sparse. Among the predicted relationships (i.e., the dashed lines), "arthritis or rheumatism" accounts for a relatively high proportion, reaching 83.33%, followed by "liver disease" at 13.33%.

[0107] Because the low-risk data itself consists of various less common disease evolution scenarios, the number of prediction results is relatively small, the link robustness level is low, and the overall link prediction score is at a low level. Among the top 8 results in link prediction score, arthritis or rheumatism (numbered 12) appears frequently, and multiple "three highs" diseases (numbered 0, 1, and 2) may appear simultaneously or partially. Therefore, for the health service supply side, it is recommended to enhance the breadth of popular science knowledge on the prevention of arthritis or rheumatism, reaching more people, and providing more lifestyle guidance for relevant patients.

[0108] In summary, this embodiment focuses on the practical problem of assessing the risk of new comorbidities in chronic diseases. Based on the CHARLS cross-time window dataset, it constructs temporal triples by splitting the order of disease onset and introduces a multi-disease co-occurrence coefficient to classify risk levels. A relational graph convolutional network model fusing temporal features was trained, achieving a peak accuracy of 83.53% on the Hits@3 metric. Based on this foundational model, RGCN-Jo and RGCN-Jo++ models fusing joint data were constructed, with the latter using representation embeddings to better learn incremental data features. This embodiment verifies the effectiveness of the link prediction-based research framework for assessing the risk of new comorbidities in chronic diseases, providing new methodological support and knowledge contributions for understanding the emergence of new comorbidities in chronic diseases.

[0109] In terms of application scenarios, the results of this embodiment primarily serve two major directions: academic medical research and intelligent health management. For medical researchers, this method, through systematic modeling of patient medical history data and disease association networks, reveals the potential patterns of new comorbidities, providing analytical tools and directional guidance for subsequent research into the comorbidity mechanisms of specific diseases and the identification of high-risk evolution patterns. For health service providers (such as primary healthcare workers, medical institutions, and health management platforms), clinically validated knowledge of new comorbidity risks can be integrated into clinical decision support systems (CDSS) or health monitoring platforms. Based on historical data, high-risk groups for new comorbidities can be identified and defined, assisting in pinpointing their health needs and subsequently pushing customized health management recommendations, personalized physical examination plans, and targeted prevention strategies, achieving early warning of diseases and medical decision support. This application helps improve the accuracy of chronic disease prevention and control and optimize the efficiency of health resource allocation.

[0110] At the methodological level, current work primarily focuses on disease association networks to assess the risk of new comorbidities in chronic diseases. Future work could further integrate individual patient characteristics to construct more refined, personalized predictive models. Simultaneously, addressing the incremental learning challenge for rare chronic diseases, future research should explore more efficient encoder optimization and few-shot learning mechanisms. At the data and computational level, this embodiment validates the framework's feasibility based on the CHARLS dataset, but the data scope and quantity are relatively limited. With the continuous accumulation of multi-source, large-scale health data, the resource consumption and efficiency of model training will become challenges. Future work could introduce incremental learning, federated learning, and other methods to support sustainable and efficient model iteration and application deployment in ultra-large-scale data environments.

[0111] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for assessing the risk of new comorbidities in chronic diseases based on link prediction, characterized in that, include: S1: Obtain the CHARLS dataset; the CHARLS dataset includes the type, time, and description of chronic diseases; Multi-stage probability sampling of the CHARLS dataset was performed based on different time scales to obtain the source dataset and the new dataset. S2: The source dataset is split based on the order of onset of disease. Existing disease changes are encoded as head entities, and newly added diseases are encoded as tail entities. Relationships and relative temporal sequences are established between the head and tail entities to complete the construction of the temporal triplet structure. The relationship is the risk level of new onset. S3: Based on different new risk levels, perform stratified sampling of the temporal triplet structure in the source dataset to obtain the training set; S4: Construct a first prediction model and a second prediction model based on a relational graph neural network, input the training set into the first prediction model to predict the risk of new occurrences, and train it based on the target loss function; S5: Based on the new dataset, the encoding representation system corresponding to the source dataset is expanded to obtain a joint training set; S6: Use the entity and relation embedding vectors corresponding to the trained first prediction model as the initialization parameters of the second prediction model; train the second prediction model based on the joint training set to obtain the trained second prediction model; S7: Perform a new risk assessment task for chronic disease comorbidities based on the trained second prediction model.

2. The method for assessing the risk of new comorbidities of chronic diseases based on link prediction according to claim 1, characterized in that, The process of classifying the risk level of new cases in step S2 is as follows: A multi-disease coexistence coefficient is introduced to quantify the risk of new cases and classify them into levels.

3. The method for assessing the risk of new comorbidities of chronic diseases based on link prediction according to claim 1, characterized in that, The training process of the first prediction model in S4 specifically includes: For different new risk levels, relational subgraphs are constructed and trained to achieve deep learning of each risk feature; By aggregating the results of different relational subgraphs and combining them with the objective loss function, the disease modeling under different risks is iterated repeatedly to complete the training of the model.

4. The method for assessing the risk of new comorbidities of chronic diseases based on link prediction according to claim 1, characterized in that, Step S5 specifically includes: Each entity in the new dataset is examined to distinguish between known entities and newly added entities. The original encoding of known entities is retained, and new encoding is appended to newly added entities. The encoded entities are then added to the encoding representation system to obtain a joint training set.

5. The method for assessing the risk of new comorbidities of chronic diseases based on link prediction according to claim 1, characterized in that, Step S7 specifically includes: Obtain the set of head entities representing known disease changes in the data to be predicted. Combine the head entity set with the data corresponding to each new risk level by performing a Cartesian product to form all possible ordered pairs. Input these pairs into the trained second prediction model for prediction and output the top three possible tail entities. Calculate the link prediction score, normalize all scores, and then filter out time-series triple structures with scores greater than 0.5 to form candidate triples; Input unknown entities that are not yet connected in the data to be predicted into the model, and output a unique risk with reference value to check the uniqueness of risk for candidate triples; By removing contradictory data that has already appeared in the head entity in the prediction results, logical consistency screening is achieved, and the final prediction results are obtained.

6. A system for assessing the risk of new comorbidities in chronic diseases based on link prediction, characterized in that, include: The data acquisition module is used to acquire the CHARLS dataset; the CHARLS dataset includes the type, time, and description of chronic diseases. Multi-stage probability sampling of the CHARLS dataset was performed based on different time scales to obtain the source dataset and the new dataset. The triplet construction module is used to split the source dataset based on the order of disease onset, encode existing disease changes as head entities, encode new diseases as tail entities, establish relationships and relative temporal order between head and tail entities, and complete the construction of the temporal triplet structure; the relationship is the new risk level; The basic model training module is used to perform stratified sampling of the time-series triplet structure in the source dataset according to different new risk levels to obtain a training set; a first prediction model and a second prediction model are constructed based on a relational graph neural network, the training set is input into the first prediction model to predict the new risk, and training is performed based on the target loss function; The joint training module is used to expand the encoding representation system corresponding to the source dataset according to the new dataset to obtain a joint training set; use the entity and relation embedding vectors corresponding to the trained first prediction model as the initialization parameters of the second prediction model; and train the second prediction model based on the joint training set to obtain the trained second prediction model. The risk prediction module is used to perform a new risk assessment task for chronic disease comorbidities based on the trained second prediction model.

7. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to enable the electronic device to perform a method for assessing the risk of new comorbidities of chronic diseases based on link prediction, according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements a method for assessing the risk of new comorbidities of chronic diseases based on link prediction as described in any one of claims 1-5.

Images