Multi-disease risk prediction models for population-scale personalised screening

An increased emphasis on disease prevention is essential to improve population health and reduce healthcare resource needs. Key requirements for effective population scale prevention programmes are an optimal balance between simplicity of screening and accuracy of risk prediction at individual level. We developed models of varying complexity to predict the incident risk of 24 diseases in the UK Biobank population, testing combinations of modalities including established clinical variable-based risk scoring, 1H-NMR metabolomics, polygenic risk scores (PRS), and self-answered questionnaires with individual electronic health record (EHR)-based past medical history (PMH). Our results show that prediction models that utilise just questionnaire and PMH data ("No Needle" model) or with added metabolomics and PRS ("Single Blood Draw" model) exhibit robust discriminative performance, at least as good or better than a comprehensive clinical variable model or established cardiovascular disease (CVD) risk scores. The "No Needle" model was also validated in an independent prospective cohort, the Estonian Biobank. We also tested scenarios for deployment of these models to improve the effectiveness of the NHS Health Check, a population-scale UK screening programme. Our results suggest high potential for less resource-intensive approaches than current clinically-based paradigms for prediction of incident disease.