Watson better learn to read crabby handwriting and minds. The fundamental issue in applying AI to medical data is the low information dataset. We are working on a couple of smaller AI projects in my hospital and finding that even when electronic, most entries in the record are narrative not discrete.
Applying ranges and logic to data pulled out of narrative records is tricky and leads to unusual responses. Even when discrete, the data set may be difficult to use. My favorite example seen nationally is blood transfusion data. It can be in units of 250ML, 500ML or just ML. Users don't look at the unit label for a field. If we ask for the most precise measurement in ML, we get amounts of 1 or 2. We know nobody got a thimble of blood.
AI interventions have to be quick. It's no good telling the provider later he made a mistake. The goal is to steer them in the right direction before they act without presenting them with so much info they throw in the towel. Not easy and in infancy.