Getting the most from electronic health records

Peter Szolovits (MIT), Preethi Raghavan (IBM Research), and Jennifer Liang (IBM Research)


Electronic health records (EHRs) are invaluable to physicians in facilitating clinical decision-making, but maintaining and extracting needed information from them is a big challenge. In fact, primary care physicians may spend more time working with EHRs than face-to-face with patients in clinic visits. AI has the power to help physicians discover insights from EHRs much faster but is impeded by the lack of shared, large-scale, annotated datasets to build and train deep learning models. MIT and IBM scientists are working together to remove this barrier. They have already created a new method for generating datasets for question-answering from EHRs. Going forward, they will expand on this method; use it to build, test, and release new models; and encourage the community to join the effort to advance medical AI technology. The team will also develop a shared resource that could be used to construct a timeline of disease-related events for an individual patient from structured and unstructured EHR data, a task for which deep learning models do not currently exist. The model they are developing will extract a partial timeline for an EHR, then automatically complete it using disease progressions learned from large-scale structured databases like MIMIC.