Background

Accurate identification of individuals at highest risk of disease is key to determining population groups who should be prioritised for vaccination. Routinely collected electronic health record (EHR) data are a critical resource for identifying population groups at high risk of disease, as exemplified by the COVID-19 pandemic. Current studies tend to use curated ‘codelists’ to define population subgroups with clinical conditions. While this approach facilitates statistical analysis and clinical interpretation, it suffers from key drawbacks including: (i) the grouping of conditions that may vary by quality or severity; and (ii) the potential omission of informative diagnostic codes that are absent from curated codelists. A data-driven machine learning approach that treats each diagnostic or medication code as an individual piece of information may offer a powerful complement to existing methods.

Project objectives

1. Compare risk prediction based on high-dimensional machine learning versus regression-based approaches in a case study
2. Evaluation the performance of machine learning for risk prediction across different diseases and population subgroups

Skills/methods

1. Machine learning using routine electronic health record data
2. Regression-based risk prediction based on clinical code lists