Machine Learning for Electronic Health Records (EHR) - Towards a Universal Patient Representation and Multi-Disease Risk Prediction
Reza Khorshidi
July 8, 2020, Wednesday, 3:00 PM - 4:00 PM EDT
The fast adoption of electronic health records (EHR) has provided the field of medicine with an unprecedented opportunity for finding patterns and insights that can influence both policy and practice of care. A patient's EHR is usually a sequence of mixed-type (i.e., numeric and nonnumeric fields) multimodal (e.g., diagnosis codes, medications, measurements, and more) data, that occur in regular intervals. While expert-defined features/markers extracted from EHR at a baseline (e.g., time of certain diagnoses, or age) have led to relatively accurate prediction of certain outcomes, the ability of such approaches in leveraging the full richness of EHR is limited. Therefore, in recent years, there has been a surge in the development of models that can automatically determine the optimal combination of EHR data for creating both general-purpose and task-specific predictors/markers. In this talk, I will introduce some of the latest developments in the field of machine learning (mostly, deep learning) and EHR. I will introduce BEHRT (Liu and Rao et al 2020, Nature Sci Rep), which is inspired by the latest developments in the field of natural language processing (NLP) and has been shown to provide superior accuracy in the prediction of disease onsets, when trained and tested on millions of patients' linked EHR data from the UK (i.e., CPRD). Furthermore, I will explore additional ideas on how such deep learning models can close the gap between predictive models and clinical practice, and help improve the personalised risk models and complex diseases patterns such as multimorbidity.
Reza is currently the chief scientist at AIG, and a principal investigator (in machine learning and medicine) at Deep Medicine program of The University of Oxford. His current research at Oxford is focused on probabilistic machine learning, and deep sequence models, for biomedical informatics, population health, and precision medicine; more specifically, he is interested in using machine learning for the development of personalised health predictions and recommendations, and an improved understanding of multimorbidity. Reza's team at AIG (i.e., Investments AI) is a group of scientists, engineers, designers and product managers/strategists, primarily focused on the development of AI-first products in the FinTech space.