Applications of AI methods to the "real world" problems of medical decision making formed major intellectual roots of the "expert systems" developments of the 1980's, contributed strongly to the development of probabilistic inference methods, and helped to focus AI applications research away from old "toy" problems to ones with rich structure and demanding requirements. Despite these intellectual successes and many valuable technical results, we have not had much impact on the way medicine is actually practiced. Indeed, the most "dangerous" aspect of medicine is that so much of practice continues to be idiosyncratic, has few checks and balances, and is uninformed by scientific evidence. Even record-keeping, seemingly a prerequisite for systematic quality improvement, is still remarkably poor except in a few areas such as lab data.
In this informal talk, I want to outline what research I think is most critically needed today to overcome the lack of impact. These appear to be less about deep reasoning than about making routine things really simple and thus fitting in with, rather than obstructing the clinical workflow. Some key concepts that will play an important role:
Once we put all these in place (!), then we can afford to return to deep reasoning about difficult medical cases, which I still think is where a lot of the fun is. Discussion and arguments will be encouraged.