Data without Pizza #3 - AI and Health talk by Peter Szolovits (MIT)

It's been a while, but now it's time for the third online AMDS meeting! Still without pizza, but with an attractive program.

This edition features:

** AI in Health - Peter Szolovits (MIT)

Peter Szolovits is a full professor at MIT (Boston) and heads the clinical decision making group at the Computer Science and AI Lab (CSAIL). He has been one of the early pioneers in the area of AI in medicine and has won numerous awards for his contributions. In this talk, he will provide a historical view of the development of AI applications in medicine, ending with a presentation of some of the projects he and his group is currently working on. A more detailed description of the talk can be found below.

*****

The idea that computers could be helpful in medicine dated back at least to Turing. My interest in clinical decision support was kindled by an insightful though over-optimistic projection by Schwartz in 1970 about how computers would change medical practice and practitioners “by augmenting … the intellectual functions of the physician.” Because machine learning (ML) methods were then primitive and computerized data to learn from were unavailable, we turned to “expert systems” methods that based decision models on knowledge learned from human experts. Although some early programs could demonstrate impressive examples, they tended to be very fragile when cases they were asked to analyze fell outside the “sweet spot” of the expertise they embedded. However, starting in the 1990s and especially by 2010, electronic medical record systems started to be more widely deployed, so clinical data became plentiful, though difficult to share among institutions. At the same time, the growth of computing power allowed neural network-based “deep models” to be trained, and these, if exposed to a broad range of actual clinical cases, proved more robust than the earlier expert systems. Nevertheless, actual use of ML-based decision support tools is still rare, though some image interpretation models are now commercially available and approved for use.

I will review a few of our many research efforts now under way to address my perceived needs in this field. One area is creation of multi-modal models that bring together imaging and text data which, together, can create better image interpretation models. A second is exploration of new potential imaging biomarkers that may correlate with difficult-to-assess clinical conditions. A third is exploiting huge recent advances in natural language processing to support question answering about records of clinical observations. A fourth introduces ways to incorporate a priori medical knowledge into ML models that better reflect human understanding. My list is illustrative, not comprehensive.

Looking further ahead, I believe the major drivers of advances in health care are likely to come from (1) better minimally- or non-invasive measurement technologies that allow us to take a closer look at details of how the body works in health and illness, (2) greater insights into the genetic bases of disease, and (3) improved use of routinely-collected (non-trial) data to learn the best methods of intervention. Each of these goals faces challenging technical and conceptual problems, but each also offers great opportunities for AI and computing.