Generative Pre-Training on MEDical event streams in Intensive Care

Background: Artificial intelligence (AI) holds great promise for improving patient care, but challenges related to data irregularity and complexity have hindered its translation into clinical practice. Modelling rich longitudinal electronic health records (EHRs) such as those found in intensive care units (ICUs) remains especially difficult, as they represent a complex interplay between the patient’s health and clinical decisions made in response.Objectives: We aim to develop a robust AI framework for flexible prediction of any outcome in the ICU and beyond. We will pioneer a class of generative pre-trained models optimised for complex EHR data (Objective 1). Our approach will be rigorously benchmarked across outcomes and hospitals (Objective 2) on a secure, federated infrastructure that ensures data privacy (Objective 3). Methods: Our approach treats EHR data as a stream of clinical events in continuous time and uses generative pre-training to learn both the likely time and the content of the next event. From generated sequences of future events, we will derive explainable predictions for any clinical outcome, including those not encountered before (zero-shot learning). Our approach will be powered by the largest set of harmonised multicentre ICU data to date, covering up to 1 million patients with 33 billion clinical events. We will refine existing methods in federated learning to allow for secure decentralised training of our models at scale.Innovation & Impact: Our work will introduce a novel paradigm for AI-based clinical risk prediction, setting the stage for a new era of flexible, general-purpose AI in medicine. Our unprecedented multicentre benchmarks will provide an urgently needed baseline for meaningful innovations in the field, while our federated approach will facilitate secure model building across institutions and borders.