Chronic Disease Prediction Using Graph Machine Learning

Abstract

Chronic diseases are a growing concern worldwide, significantly impacting public health and healthcare systems. This research employs Australian administrative claims data to predict chronic diseases using a novel graph-based approach that considers the complex relationships between patients and diseases. The thesis begins with a comprehensive review of existing graph machine learning in disease prediction, aiming to provide a solid foundation of knowledge. It then proposes a “Patient Network” and applies a variety of machine learning models for prediction tasks. Among these, the Random Forest model stands out for its superior performance. Next, to effectively manage intricacies in patient relationships, the study proposes a chronic disease prediction framework using Graph Neural Networks (GNNs). It develops a “Weighted Patient Network” that uncovers latent relationships among patients, enhancing predictive accuracy. Moreover, this research addresses the challenge of link prediction in the context of comorbidities by comparing various graph machine learning models with traditional similarity-based methods. The findings reveal that combining hand-crafted feature techniques with eXtreme Gradient Boosting is most effective. Furthermore, the research extends to chronic diseases and their comorbidities using graph theory combined with a content-based recommender system. This approach demonstrates that models incorporating network features yield lower prediction errors. Notably, the graph convolution matrix completion model is highlighted for its minimal error rates. Overall, the thesis underscores the effectiveness of graph machine learning models in enhancing chronic disease prediction, showing potential benefits for healthcare stakeholders by identifying patient cohorts at high risk of developing chronic diseases. This enables early intervention strategies to prevent these conditions, ultimately leading to reduced healthcare expenses for both providers and patients.