Predicting Dementia Risk Using Longitudinal Electronic Health Records Data

EHR data are collected as part of routine clinical practice and can be used to train predictive models that estimate and stratify disease risk at the population level. This is particularly valuable for conditions like dementia, where advances in disease-modifying treatments and early interventions have the potential to substantially reduce disease burden. Longitudinal analysis of EHR data can provide valuable insights into dementia risk. In this work, we investigate the utility of statistical learning, graph neural networks, and large language modelling approaches to predict dementia five years before diagnosis using time-course medical history data. We evaluate the performance and utility of five different modelling approaches using data from the UK Biobank (n=9,537) and present a risk stratification model.