Research On Disease Prediction Model Based On Hypergraph Neural Network

Accurate and interpretable prediction of health events plays an important role in the care planning process for patients by healthcare providers.Electronic Health Record(EHR)contains historical health information about patients and are often used in disease prediction studies.In recent years,enhancements in the availability of EHR datasets have led to a number of advances in deep learning for disease prediction.However,the existing deep learning-based approaches still have challenges in the following aspects: 1)How to effectively utilize a priori expert knowledge from external disease domains;2)How to collaboratively learn patient features and disease features,especially higher-order features;3)Lack of utilization of unstructured textual data;and 4)The identification of common disease concomitant patterns in the disease prediction process needs to be further enhanced.To address the above issues,we propose a collaborative hypergraph learning model to explore the patient-disease interactions by combining relevant medical domain knowledge to be able to capture the higher-order structural features of patients and diseases.The experimental results show that the proposed approach achieves better prediction results compared to the currently available disease prediction models.The main work of this thesis is reflected in the following three aspects:1.Based on the MIMIC-Ⅲ medical dataset,we conducted data screening and cleaning for the data formats required for the model,and constructed patient-disease hypergraphs,diseasepatient hypergraphs,and three dynamic sub-hypergraphs through data processing techniques.This thesis also models the unstructured text data and patients’ symptom information contained in MIMIC-Ⅲ and designs modules to embed them into the model proposed in this thesis to achieve improved model prediction performance.2.A deep learning model for disease prediction based on hypergraph(Electric Health Record to Hyper Graph,EHR2HG)is proposed.First,we optimize the initialization conditions of the model using the existing disease classification information.Then,by constructing hypergraph relationships between diseases and patients,we consider all patients and all diseases together rather than as separate individuals,allowing the model to mine and model hidden higher-order entity relationships,such as complications or groups of patient categories.Finally,patient medical record text data contain a large amount of unstructured,lengthy information that is rarely used in currently available models,and our proposed EHR2 HG model models this type of information and utilizes the information in it for the output of disease prediction results.We evaluated the proposed model on a real EHR dataset and the results showed that EHR2 HG can achieve better prediction results compared to several common disease prediction algorithms currently available.3.A Dynamic sub-Hyper Graph-based deep learning model for Disease Prediction(DHG4DP)is proposed.Firstly,we extracted the symptom information of patients in EHR,and the extracted symptom information was used for the creation of patient embedding.Then,by dividing the large hypergraph into three sub-hypergraphs,thus DHG4 DP can both globally mine the relationship between patients and diseases and discover specific disease concomitant patterns.In addition,the model correlates the sub-hypergraphs with patient time-series information to construct dynamic sub-hypergraphs,and organically fuses the disease relationships learned from the three dynamic sub-hypergraphs with the segmented disease types to finally achieve improved disease prediction accuracy.The experimental results show that our proposed DHG4 DP model is more advantageous in extracting higher-order relationships between diseases and temporal relationships between diseases,and the disease prediction accuracy is significantly improved compared with existing models.