Laine Thomas, PhD, associate director of DCRI Biostatistics, sat down with David Page, PhD, chair of Duke Biostatistics & Bioinformatics, who is new to his role as of this year, for a discussion on the future of biomedical data science, as well as how machine learning techniques will be incorporated in this work.
LAINE: David, I’d like to talk about your vision for Duke Biostatistics & Bioinformatics in terms of where we’re going in some exciting new areas, like artificial intelligence and machine learning. Let’s start with a brief explanation of
artificial intelligence and machine learning.
DAVID: Artificial intelligence is automating human thought and actions. Machine learning is the adaptation piece of that. I view machine learning as the creation of algorithms that will analyze data and give back insights. Those insights could be predictors of the future, better explanations of past events, or indicators of cause-and-effect relationships. Although I’m trained as a computer scientist, I see machine learning and data science emerging just as much from statistics as from computer science and engineering.
LAINE: Can you talk us through your vision for collaboration among different groups on campus doing work in biomedical data science?
DAVID: A tremendous opportunity is to use cross-campus partnerships to recruit new faculty and staff. I think the DCRI and Duke Biostatistics & Bioinformatics can be critical players here because we offer outstanding data access, as well as translational opportunities for people who want to make a real difference in society.
LAINE: Speaking of intersections on campus, I want to segue a bit into other intersections—those between machine learning and causal inference, just because that’s the area my work is in. Have you worked on any problems that sit in this intersection?
DAVID: For the last 10 years, I’ve been working on using machine learning for discovery of adverse drug events. That project is primarily focused on using observational data such as EHR data or claims data. We’ve tried to build on previous statistical and artificial intelligence methods, and our best results have come from algorithms that combine insights from both traditions.
LAINE: I’m working on a project where we are developing new methods to analyze the comparison of hysterectomy and myomectomy in the COMPARE-UF registry of women with uterine fibroids. This DCRI registry emphasizes personalized medicine, where we’re trying to estimate individual treatment effects while accounting for differences between groups. Our approach also combines both traditions. We have a machine learning element, but it’s currently a tool that’s separate from the causal inference. We’re fitting models using machine learning methods, then the causal inference phase occurs. I can imagine some ways to make them more integrated. In my mind, the next step is to make them actually interact with each other instead of treating them as pre- and post-processing.
DAVID: That’s really exciting. I’d love to hear from you about what other areas you see for future intersection between the two fields as we continue to seek collaboration opportunities.
LAINE: In the DCRI Biostatistics group, I see a number of potential areas where machine learning could improve how we approach problems in causal inference. One area is precision medicine. As our purpose in causal inference changes—instead of trying to estimate averages, we’re now trying to estimate individual treatment effects—we need to figure out how to do that better, and machine learning could help. Another opportunity is hospital profiling. Because our work in this space is in demand from the patient perspective, I expect it will continue to expand and could integrate machine learning to improve our rating models.
DAVID: Thanks for those insights, Laine. Both of those sound like exciting opportunities to find the right hybrid approach that could improve how we do things across the board, and I look forward to working with you in these areas.
This article originally appeared in the DCRI’s 2018-2019 Annual Report. View more articles from this publication.