King's College London, University College London

3

CCU078

Who gets ill, when, and with which diseases and outcomes, are key questions for individuals, clinicians and healthcare systems. The COVID-19 pandemic highlighted this, with a need to identify vulnerable individuals to prioritise interventions such as vaccination. Many models were developed to provide answers, but these focussed on single, narrow questions, such as the risk of death when admitted to hospital with COVID-19. This limits their ability to answer new questions, such as future pandemics or treatments, and may lead to bias when stretched too far outside their use case, for example in patients with rare diseases the model didn’t see during development. This work aims to harness recent advances in artificial intelligence (AI) to develop the world’s first national-scale generative AI model for patient medical histories - simply put, learning from past patient data, to predict the future. Sometimes called ‘Foundation models’, we will rigorously test our model’s ability to generalise to predicting a wide-range of patient-centred outcomes, such as disease/COVID-19 onset, hospitalisation and death, for all individuals, across all backgrounds and diseases. Through this ability to generalise to new questions, this model could allow us to rapidly make accurate predictions about future pandemics, as well as helping to answer ‘what if’ questions for researching the existing COVID-19 pandemic. Importantly, working with national-scale data allows us to minimise and investigate signs of bias by training the model on a diverse, representative population and investigating its performance in important groups, such as different ethnicities.

As a single model for all diseases in all individuals we anticipate this work will deliver benefits across multiple facets of healthcare, including those underserved in current research and policy. Examples are many, but could include virtual screening, identifying individuals at high risk of disease or emergency admission early, enabling preventative action to be taken, or simulating clinical trials to predict the effect of changing medications. Through securely sharing our work with approved researchers, the NHS and UK Government, we aim to scale this impact, accelerate the route to patient’s benefiting from such research and exemplify the UK Government’s AI Taskforce mission of developing world leading capability in AI and UK Sovereign foundation models. Visit the BHF Data Science Centre website for more detailed information about project outputs. https://bhfdatasciencecentre.org/projects/ccu078/

This project used the following datasets within the Trusted Research Environment for CVD-COVID-UK / COVID-IMPACT: - Sentinel Stroke National Audit Programme Clinical Dataset - MSDS (Maternity Services Data Set) - Secondary Uses Services Payment By Results - Covid-19 Second Generation Surveillance System - Hospital Episode Statistics Outpatients - Covid-19 UK Non-hospital Antibody Testing Results - Hospital Episode Statistics Accident and Emergency - Covid-19 UK Non-hospital Antigen Testing Results - COVID-19 Vaccination Status - COVID-19 Vaccination Adverse Reaction - Improving Access to Psychological Therapies Data Set - Civil Registration - Deaths - Hospital Episode Statistics Critical Care - COVID-19 SARI-Watch (formerly CHESS) - Medicines Dispensed in Primary Care (NHSBSA Data) - North East and North Cumbria - CUREd+ Linked Emergency Care Data Set (ECDS) - Emergency Department (ED) attendances - Mental Health Services Data Set (MHSDS)

03/12/2023

22/12/2023

De-Personalised

22/12/2023

TRE