Temporal graph-based convolutional neural networks for electronic health records

Graph theory offers a powerful framework for using the relational dependencies in Electronic Health Records (EHRs) to enhance machine learning (ML) predictions of health outcomes and diagnoses. This thesis explores and advances graph-based ML approaches, with applications for the prediction of future hip and knee replacement risk.

A systematic literature review identified 832 studies, with 18 using patient-level graph representations of EHRs for predicting health outcomes. This review showed that current graph-based EHR models have limited clinical applicability due to high risk of bias.

A novel Temporal Graph-Based Convolutional Neural Network (TG-CNN) model was developed. Initially applied to student dropout prediction in online courses, this approach demonstrated state-of-the-art performance. Extending this method to medical data, TG-CNNs were applied to predict hip and knee replacement risks, one and five years in advance. Temporal graphs, constructed from primary care event codes from EHRs, captured temporal relationships between symptoms, diagnoses, and prescriptions. Models achieved AUROC values up to 0.967 for hip replacement and 0.955 for knee replacement.

To improve model interpretability, four explainable methods were explored, including gradient based and feature-mapping approaches. These methods provided visual insights into TG-CNN predictions, highlighting the influence of key EHR features such as prescriptions. While clinicians found these visualisations informative, further simplification is needed to support real-world clinical decision-making.

This thesis demonstrates that graph-based representations improve the predictive performance and interpretability of ML models in healthcare. The TG-CNN model offers the potential to enhance patient care and management through earlier and more accurate predictions. Future work should focus on improving model explainability and translation into clinical practice.