Interpretable and probabilistic deep learning models for heart failure risk prediction using temporal and multi-modal UK electronic health records (Mr Yikuan Li)

Recent advances in artificial intelligence (especially deep learning) and the growing access to large-scale datasets such as electronic health records (EHR) from millions of individuals have provided an unprecedented opportunity for medical research.

However, despite striking progress in the earlier works, several limitations remain that have been hindering the usefulness of models for discovery and their wider application in clinical practice.

In the context of EHR, existing approaches have largely relied on a fraction of information available in the datasets (typically disease and medications from hospital records) and have ignored the inherent temporal characteristics of medical records.

Additionally, deep learning predictions remain largely deterministic, thus ignoring the uncertainty of estimates, which is important for clinical decision making.

Finally, deep learning models are perceived as ‘black-box’ models with little opportunity for explaining healthcare phenomena which is often crucial for discovering disease determinants.

This proposal aims to (1)develop a framework for heart failure risk prediction, as an example for other complex diseases, using temporal and multi-modal UK electronic health records; (2)develop methodologies for probabilistic modelling to quantify the reliability of predictions, with provision of measures of uncertainty; and (3) develop methodologies to make deep learning models more interpretable towards risk factor analysis.