Current UROP openings

Project title: Reinforcement learning and causal inference for patient care management
Group: Clinical Machine Learning (Professor David Sontag) in EECS
Project description:
Managing chronic diseases is important for preventing adverse outcomes and helping patients live healthy lives. Our health insurance collaborators provide care management programs to patients to improve their health outcomes and reduce cost. The goal of this project is to maximize the impact of these programs by identifying patients who would benefit from these programs for outreach. Because outreach decisions are made over time, we can model sequential decision making to learn a policy for when to reach out to patients.
The goal is to develop a reinforcement learning-based approach that incorporates causality, fairness, and interpretability. For causality, the goal is to learn a causal graph on the pathway through outreach, enrollment, healthcare utilization, outcomes, and unenrollment across time. For fairness, we would like to ensure our policy results in fair allocation of healthcare resources to minority groups. For interpretability, explaining why an action is recommended both in terms of current patient features and potential future outcomes is important for deployment.
We are seeking a UROP to drive this project through data exploration, literature review, causal framing, model development, and ideally deployment with our insurance collaborators. Because developing the correct machine learning framing for this clinical problem is essential and challenging, this project will likely take more than one semester, so longer-term commitment would be preferred. Expectations for UROP are to be motivated and passionate about the project, work on the project for at least 12 hours a week, and have prior experience or course work in ML. Email Christina Ji (cji at mit dot edu) with resume and transcript if interested.

Project title: Using machine learning and natural language processing to make clinic visit notes easy for patients to comprehend
Group: Clinical Machine Learning (Professor David Sontag) in EECS
Project description:
Ever read your clinic visit notes after a hospital procedure and wondered what in the world it meant? Your medical record and clinic notes have important information that summarize your visit, prescriptions, and next steps. However, these notes contain medical terminology which millions of patients find difficult to understand and interpret. Take the following sample note for example: “Hypocomplementemic urticarial vasculitis. Presented with hives since early 2012, a skin biopsy was consistent with vasculitis. Her laboratory reported negative serology for HBV, HVC, ANCA, dsDNA, Ro, La, Sm, RNP, normal CI inhibitor, C4 3”…….. you get the point. If you aren’t a patient with a medical background a lot of these words and acronyms might as well be gibberish. While patients could hypothetically google each word they don’t understand, it is inefficient. Our goal is to translate visit notes readily and efficiently to better understood definitions, using natural language processing methods, and deploy it for real-time, real-world use. We will be (1) identifying the appropriate NLP model (BERT, GPT-3, T5, etc), (2) Optimizing and training the model on open-source general ICU medical record data, (3) Generating additional labeled data from closed-source medical data, for retraining and testing, (4) Deploying the most appropriate model into a toolbox appropriate for use, for example, a website or chrome extension, (5) Publishing results in a journal or conference.
We are seeking 2-3 UROPs to drive this project through data extraction, model development and deployment for use. UROPs should be motivated and passionate, committing to working on the project for at least 12 hours a week, for at least 1 semester.
Experience and skills requirements:
UROP 1: prior experience or coursework in web development and comfortable with JavaScript programming, HTML and CSS. Nice to have: familiarity with Google’s quality standards, experience with Tensorflow.jS or translating python ML models to JavaScript.
UROP 2: prior experience or coursework in ML (preferably NLP-specific). Comfortable coding in Python. Nice to have: experience with Tensorflow.jS or translating Python ML models (eg PyTorch) to JavaScript.
UROP 3: A third UROP may be considered to fill any gaps.
Please email Mercy Asiedu (mnasiedu at mit dot edu) with your resume and transcript if interested.