This demo collects patient information in an open-ended and conversational format and then writes a preliminary medical note based on patient responses.

Many GPT-3 applications focus on GPT answering user queries; here, we flip it around, with the system asking the user instead.

I think open-ended conversational question-asking (that maintains long-term coherence, for which we use SNOMED-CT ontology and some rules) is especially important in medicine. Most existing systems rely on multiple-choice questions with a heavy amount of medical reasoning, but that’s very hard to get right. You can’t enumerate all the possible ways a patient can be sick, which is why many doctors begin the patient interview by letting the patient tell their medical story.

This is also not supposed to replace doctors. Many emergencies and other complicated medical cases require strong clinical reasoning to collect medical history, which a system like this lacks. Here, we just try to collect some information before the encounter to help contextualize the physician about the patient from the get-go and to empower doctors and nurses to do triage.

A couple of notes on the technical side: (1) This is openai dependent (each turn is a variable, but often 3-5 openai calls, if openai is over loaded, this means that the responses can time out). (2) The chatbot session url is a permalink. It can be shared/refreshed/etc. If it times out, you can try to continue the discussion at a later time (just go/refresh the permalink).