Last Wednesday, January 6, 2016, I had an opportunity to attend a presentation about NIH’s work with IBM’s Watson. Several groups at NIH have been exploring how Watson might benefit our operations and serve our intramural and extramural research communities. The presentation featured work that Mike Lauer, who is Deputy Director for Extramural Research, has been involved in.
The first example pertained to using Watson to do impact analyses of NIH funding. Our approach at the moment tends to be pretty crude, usually number of publications from a grant or initiative. Humans had determined that less than two-thirds of clinical trials at the National Heart Lung and Blood Institute had published results within 30 months after they were closed. They identified around five predictors of publication. See the citation here. When Watson was fed the grant applications and taught what to look for, an additional ten or so predictors were identified.
This drove home for me what Watson is good at – identifying connections. There is a limit to the human capacity to process large amounts of information looking for relationships. Not true for computers.
A related example was using the same data to look for adverse events for cardiovascular drugs. Watson identified several rare but serious adverse events that had been overlooked because they were reported in free text rather than in a table. In the data display, those dots were colored yellow and humans could click on the dot to drill down to the specific papers. While these adverse events might not rise to the threshold of FDA labelling, they might provide clues that could be used to develop pharmacogenetic studies.
Another important point was that Watson is a tool. The results have to be checked and verified by humans. However, what makes Watson distinct from other large computers is the ability to learn. If Watson is making errors, they can be corrected and the analyses rerun.
I left the presentation intrigued about Watson but without a clear understanding of how it would benefit healthcare delivery research. It seems to me that there are analyses where we would like to look at complex patterns of care using data that are not captured in standardized ways. An example might be relationships between patients, clinicians, and delivery systems.
Would love to hear ideas from others on this point.