The last decade has seen widespread adoption of electronic health records (EHRs) across hospitals and clinics in the US. Physicians frequently seek answers to questions from a patient’s EHR to support clinical decision-making.
It is not too hard to imagine a future where a physician interacts with an EHR system and asks it complex questions and expects precise answers with adequate context from a patient’s record. Central to such a world is a medical question answering system that processes natural language questions asked by physicians and finds answers to the questions from all sources in a patient’s record.
I will talk about the steps we have taken towards building such a system in terms of (1) creating a large-scale dataset emrQA with over 1M questions, logical forms (that capture information/answering needs in a structured format), and answers in clinical notes, (2) building a model for semantic parsing to map questions to logical forms and (3) building an automatic medical question answering system that answers these questions.
I will also discuss the challenges and roadblocks and lay out a vision for how we can make further progress in realizing such a future.
