Research On Key Technologies Of Chronic Disease Risk Prediction Based On Complex Network And Machine Learning

With the development of medical big data and network science,patient medical information can be abstracted into network representations,such as undirected or directed comorbidity networks with diseases as nodes and statistical associations or progression relationships between diseases as edges,and patient networks with patients as nodes and associations between patients as edges,to form analyzable data.By utilizing complex network analysis techniques,comorbidity patterns of chronic diseases can be extracted from these networks to improve patient management and diagnostic decision support for medical professionals.The combination of complex network analysis technology and machine learning(ML)technology can effectively predict the future disease status of patients and provide guidance for health authorities in developing screening and prevention strategies.This study focuses on three aspects of research on chronic disease risk prediction:1.Comorbidity patterns of chronic diseases in hospitalized groups based on undirected comorbidity networks was studied.Ischemic heart disease(IHD)patients were used as the research object,and the comorbidity networks of the case and control groups were constructed using the undirected comorbidity network method.The comorbidity patterns of IHD patients were investigated by comparing the differences in the networks.Results showed that the comorbidity burden of IHD hospitalized patients increased by 50%compared to the control group.The comorbidity network of IHD patients consisted of1,941 significant associations among 71 chronic diseases.Males showed a higher comorbidity burden than females,and the network complexity increaed with age.2.A framework based on directed comorbidity network and ensemble learning technique was proposed to predict chronic disease risk for hospitalized populations.This framework used patient demographic information(sex,age)and new features extracted from the directed comorbidity network(such as node score,edge score,and rank-based score)as inputs to models predicting heart failure(HF)risk in IHD patients.A stackingbased ensemble model was constructed and compared with several ML models.Network feature comparison experiments were conducted to verify the importance of network features.Experimental results showed that the constructed model outperformed other traditional ML models,and the proposed network features exhibited better performance than other constructed features in previous study.3.The method for chronic disease risk prediction based on patient network and ML was studied.This method aimed to learn the similarity relationships among IHD patients in the network and use the extracted features from the network to predict the HF risk of IHD patients.First,a patient similarity network was constructed,and then the corresponding node features were extracted from the network.Finally,by constructing ML models and comparing the feature importance,it was shown that the proposed patient network features could improve the model performance,and the proportion of feature importance exceeded 50%.