#LIGHTForum2018 Q&A: Chris Potts

Chris Potts, PhD is Professor of Linguistics at Stanford, by courtesy of Computer Science, and
Director of Stanford’s Center for the Study of Language and Information (CSLI).  At LIGHT 2018, H
ear from Chris on the plenary “From Academy to Reality — The Cutting Edge of Innovation” (May 9, 1:00 pm) 

1. Tell us about how you are applying your computer science and linguistics expertise to the healthcare field.

I’d say there is consensus in the field of natural language processing (NLP) that healthcare is one of the most important application domains for our work—maybe the single most important domain. The big conferences all regularly have numerous workshops on various aspects of NLP in healthcare and medicine, and I remember Regina Barzilay (MIT Professor in NLP and winner of a MacArthur Genius Grant) enjoining us all to focus on this area in a 2016 invited talk called “How Can NLP Help Cure Cancer?” Notice that’s a question—a call to action to a field that has arguably been distracted by less impactful problems.

The motivation is clear: some of the most valuable insights about how to improve patient outcomes are locked away in the notes doctors write about their interactions with patients, or lost in a sea of academic publications. We urgently need better automatic methods for extracting information from these texts and enabling quantitative study of what they contain.

The clinical narratives especially are really our only source for information about safety, uncertainty, and decision making. For example, one sees immediately the value in hearing about what the doctor didn’t try, or which plans were considered but had to be rejected. These negative events aren’t something you can bill for, so they don’t show up anywhere else in healthcare records!

2. So what are the near-term changes that you see these scientific and technological innovations enabling?

In the really near term, I’m hoping to see more decision making based on quantitative analysis of big data sets—text data and also more structured kinds of information. Chronic diseases seem like the place where we’ll see the biggest impact. These diseases are complex. They affect huge populations (e.g., asthma, diabetes), or they affect smaller populations very profoundly (e.g., multiple sclerosis). If we can identify patterns of care, or understand what leads people to fall our of care and put their lives at risk, we can have a massive impact on public health. These are nuanced problems where clinical insights can be usefully guided by high-quality, large-scale observational studies.

Somewhat longer term, I hope this work can help get the healthcare system out of the way of healthcare practitioners, so that they can concentrate on helping people. This could be as simple as using speech technology and NLP to alleviate the burden of entering information into digital records. The truly grand vision is that careful data curation and solid analytics will facilitate more value-based care, where we pay based on the quality of outcomes rather than on the amount of care administered. Getting this right means solving many challenging political and institutional problems, but data and analysis will be crucial.

3. How can academia and industry work together to advance healthcare innovations and improve outcomes?

I think there are lots of chances for collaboration. An example from my own experience: the team at Roam Analytics was basically able to devote an entire year to integrating heterogenous health datasets into a single, massive, densely interconnected knowledge graph. It’s very hard to sustain such projects in academia, since they require a dedicated team and a lot of resources, with relatively little in the way of immediate output like publications and findings. The result, though, is this amazing resource for understanding healthcare and medicine and building out other tools and capabilities. It’s at that point that academics should swoop in to do the kind of fine-grained scientific work that we need to realize the potential of these big data resources. Industry folks can help out at this stage too, of course, but I would hope that academic research is oriented towards the risky, cutting-edge approaches that don’t always show immediate economic rewards.



add a new comment