Medical professionals struggle with the burden of processing more and more data, mostly in the form of free text - good thing they have LLMs, which did so nicely on medical licensing examinations, to help them, right?
Wrong.
In order to create systems that integrate into real-life medical professional workflows, LLMs need to have much more than medical knowledge. We need to be able to evaluate their performance, add guardrails, explainability, and human-in-the-loop feedback to the system. All these things require not only a deep understanding of medicine, but also a deep understanding of doctors.

In this lecture, I will discuss these challenges based on my experience as a data scientist in Microsoft Healthcare group, and suggest promising ways to solve them. I will show how to use medical ontologies to achieve grounding; how to use semantic structuring for smart precision and recall medical free text evaluation; and how to use various augmentation techniques in the medical domain.

This talk is intended for data scientists interested in the medical data science domain, as well as for data scientists coping with similar problems in other domains.