On July 1, 2019, the Department of Epidemiology welcomed Andrew Beam, PhD as assistant professor, with secondary appointments in the Department of Biomedical Informatics at Harvard Medical School and the Department of Newborn Medicine at Brigham and Women’s Hospital. Learn more about Dr. Beam and his research plans in the following Q&A.
Can you tell us a little about your research background? How did you get started in your current field?
First, it’s probably worth defining what I consider to be my primary field, given the somewhat circuitous path my academic journey has taken. I am principally concerned with improving, streamlining, and automating decision-making in healthcare through the use of quantitative, data-driven methods. The name applied to this activity will vary greatly based on with whom you are speaking. Computer scientists will call this artificial intelligence (AI) and/or machine learning, statisticians will call it regression analysis, and epidemiologists frame this problem in terms of causal inference. These communities come at this core problem of decision-making and reasoning using data with different tools and different assumptions, but it is my belief that fundamentally we are all working on very similar things. However, despite these shared interests, these communities remain largely isolated, and either relearn or are unaware of advances going on in their sister-disciplines. I am of the belief there is an extremely rich intersection between the various tribes of “data science” and this intersection is currently under-explored. I am excited to see what new ideas are created as these communities begin to cross-pollinate with one another, and I consider working at this intersection to be central to my new role here in the department of epidemiology.
Given that context, my interest in these questions began when I was a computer science and engineering undergraduate and I took an introductory AI course at N.C. State University. The idea that you could instantiate principles of intelligence in software was something that I found immediately exciting, as were the ethical and philosophical issues inherent to AI. In a piece of advice that I’m still grateful for today, one of my computer science professors advised me to strongly consider getting some significant exposure to statistics as part of my graduate training. This was around 2007-2008, and I think he could see the data-tsunami just beginning to crest. His advice was prescient, and I ended up getting a masters degree in statistics while doing research at the US Environmental Protection Agency in Durham, North Carolina. I finished up graduate school in bioinformatics and computer science, where I wrote my dissertation on machine learning methods for genome-wide association studies, so I felt like I had come full circle.
My interests in healthcare and medical decision-making also developed while in graduate school. My wife was a medical student at the time, and watching her complete her training was eye-opening for me. I had taken for granted that most medical decisions were evidence-based and made by physicians who had access to state-of-the-art technical tools. This fiction quickly dissolved once I had an “insider’s” view of the healthcare system and, as it turned out, none of my preconceptions about medicine were true! Many decisions in the hospital lack a solid evidence base and technology is often an impediment for physicians, not an enabler. This was an extreme moment of clarity for me. I decided that I could explore the ideas in AI, machine learning, and statistics that I found so energizing while applying these methods to healthcare to address some very important societal problems. What could be more exciting than that? At this point, the last piece of this puzzle I was missing was some actual expertise in healthcare and medicine. I was fortunate to have the opportunity to do a post-doctoral fellowship with Zak Kohane at Harvard Medical School where I gained knowledge about the particular challenges posed by healthcare data, along with a healthy spirit of insurgency to improve the status quo that anyone who knows Zak will recognize.
What questions/problems are you working on that you are most excited to explore?
We are in the midst of the so-called “deep learning revolution,” which has changed fields like computer vision and natural language processing, and is now beginning its assault on healthcare. Even though these models still leave something to be desired (more on that below), I think that there are many problems in healthcare that these algorithms can help us with today. One area that I’m particularly excited about is the neonatal intensive care unit (NICU). The NICU cares for preterm infants (babies born before 37 weeks of gestation) and, in my opinion, is a wonder of modern medicine. What’s more, NICU infants are data generating machines! All of the data modalities on which deep learning has been so successful, e.g. imaging, text, time series data, etc. NICU infants generate in spades. Using some large sources of NICU data and with the help of clinical collaborators, I’m working to develop predictive models of NICU patients.
Moving forward, I am really excited to widen my methodological aperture and think how we might improve machine learning algorithms to allow them to be both more capable and more robust. Even though they have performed some impressive feats, current deep learning approaches actually have a pretty impoverished model of the world. As my friends in statistics and epidemiology are quick to remind me, all deep learning is doing is extremely fancy curve fitting and input-output mapping. That is of course true, but the surprise for many of us has been just how many “hard” problems (like reading an x-ray) can be solved by highly complex but still quite dumb pattern matchers. However, there are classes of problems that are not solvable by current deep learning approaches, even in principle. Moreover, at least in its current framing, nearly all of modern machine learning techniques are designed to give predictions, but what doctors often want are decisions.
Many problems of this type are ones that come up in medicine routinely. Bringing it back to the NICU, a typical question a clinician might face is “If I extubate this baby, what will happen?” This type of causal question is not something that even the best deep learning model can answer. Obviously, the causal inference group here in the department is world-class and have spent decades thinking about this very sort of question. Working with and learning from these folks to understand how we might enable machine learning models of the future to answer this type of question was one of the key reasons I was so excited to join the department.
Seeing how you came from such a diverse academic background (with degrees in computer science, engineering, statistics & informatics), what draws you now to public health and to your new role in the Department of Epidemiology? What linkages do you see between AI, machine learning, and big data with medicine and public health?
I partially answered this in the previous question, but I strongly believe that these kinds of approaches (both machine learning and traditional statistical methods), can fundamentally change healthcare and public health for the better. So, at least for me, the link with public health and medicine is crystal clear. For the vast majority of people, healthcare is inaccessible (not enough time with their doctor) and unaffordable. The current healthcare status quo is unacceptable and unsustainable, so something must change. It may sound quixotic, but I sincerely believe that the work we are doing is part of the solution.
What plans do you have for the first few years of your new role here at the Harvard Chan School?
The first few years will be probably occupied by the typical activities of starting a new lab: writing as many grants and papers as possible! Beyond that, I really do hope to facilitate news kinds of collaborations between the different communities of healthcare data scientists. I am excited to learn from my new colleagues in the epidemiology department and at the school of public health more broadly.
Can you tell us one thing that colleagues may not know already about you?
Like the other newest member of the Epidemiology Faculty, Michael Mina, I also live in Jamaica Plain near the pond! My wife and I just welcomed our first daughter into the world at the end of July, so it’s been a pretty busy couple of months. Outside of raising our daughter, I am very fortunate to collaborate with my wife on many of my NICU projects, as she is the in last year of her neonatology fellowship at Boston Children’s. She has been amazing at helping me understand the challenges faced by physicians and keeps the machine learning projects we work on firmly grounded in clinical reality.