I volunteered at the Machine Learning for Health Care conference that was hosted this year at the University of Michigan on August 8th-10th. I had such an incredible time connecting with so many different people from a super diverse range of career stages and fields of work. Day one was a community data challenge. We were introduced to two different datasets that contained electronic medical records, and encouraged to form teams that included clinical and data people. My team ended up being mostly people from UMich. We were an equal balance of clinical and data people. Our team pitched the "Minimize Opioids" idea. Here is a picture of Michael Burns delivering our ideas to the group and judges.
We decided to approach this using a reinforcement learning framework by training an agent to minimize patient pain rating scale, while simultaneously minimizing the amount of pain medication administered. We will focus on surgical patients during their hospital stay. The idea of this challenge was not to code and solve the problem in a single day, but rather formulate the problem and make plans to collaborate together over the next few months. I researched the idea of Reinforcement Learning models to manage medication schedules and found several inspiring papers. At MLHC 2018, Gregory Yauney et. al from MIT had a paper about chemotherapy and clinical trial dosing (link). In a 2018 paper, Lu Wang et al. combined reinforcement learning and supervised learning for optimizing effective prescription and low mortality (link). In 2016 Shamim Nemanti et al. had a paper using reinforcement learning to help with misdosing medications with sensitive therapeutic windows, such as heparin (link). We were given $1,000 worth of Google Cloud Platform credits to continue our work with. The team seems very motivated to continue working together, so I am very excited to see what we come up with. An email from Jenna Wiens this morning told us that we can receive free registration to MLHC 2020 if we have a research track paper accepted that uses these data. I think that is a wonderful incentive and a realistic timeline of accomplishing this project.
The next two days consisted of several keynotes. There was a gender balance, four women and four male keynote speakers, bravo to the organizers for making this happen! My favorite part of any conference is the networking opportunity that happens during coffee breaks, poster sessions, etc. I met many people from the University of Michigan who do machine learning with clinical data, PhD students at the University of Washington, University of Toronto, MIT, and CMU. I was able to have conversations with Mert Sabuncu, Roy Perlis, and Marzyeh Ghassehmi about applying to PhD programs and the topics I plan on working on in grad school (prediction in psychiatry). Every one I spoke with was really encouraging about my ideas, and gave me a lot of confidence about my upcoming applications. Despite being an "outsider" in this crowd because I am not an MD/clinical person, and my education is not in computer science, I felt more than just fitting in with the crowd but that I belong in this community. While I planned to just be a listening sponge during the conference, soaking up as much information about machine learning in health care that my brain could handle, I can say that I contributed a lot of ideas and inspiration to the community too. Everything about MLHC 2019 was a success and I am so grateful for the path I have chosen to continue working in this direction. I wholeheartedly believe that this interdisciplinary collaboration between machine learning and health care will make the world a better place.