MIT Artificial Intelligence and Wellness Seed Grants
Seed grants in AI and Wellness in the Schwarzman College of Computing support early-stage, interdisciplinary research that explores how artificial intelligence can shape the future of wellness. Funded by Panasonic Well, the program encourages collaboration between computing faculty and scholars from other fields to develop innovative, AI-driven approaches to advancing human well-being.
2025 Projects
- “CardioCopilot: A Language, Assistant, and Benchmark for Personalized Cardiovascular Wellness
Wojciech Matusik, Mechanical Engineering, CSAIL
Peter Szolovits, EECS, IMES
CardioCopilot is a symbolic-AI platform for personalized cardiovascular wellness that reimagines how preventive-care protocols are authored, personalized, and evaluated. There are three integrated components: CardioDSL, a Python-embedded domain-specific language for encoding wellness logic with temporal rules and patient macros; CardioAssist, a language model–based assistant that translates plain English into executable CardioDSL while simulating personalized trajectories; and CardioBench, a benchmark for evaluating protocol reconstruction, adaptation, and predictive performance. Together, they target a fast-moving, low-regulation zone of care—lifestyle and biometrics coaching—making this stack ideal for early deployment. CardioCopilot is powered by hybrid corpora: structured prompting over WHO and ACC/AHA guidelines, and generative synthesis via top-tier LLMs. The platform fine-tunes an open-weights model on ~3,000 DSL–English pairs and evaluates on diverse comorbidity profiles to assess editing time, protocol concordance, and personalization accuracy. The resulting tools support scalable, explainable wellness planning—bridging static clinical literature with dynamic, real-world adherence. - “DREAMS: Dynamic Real-time Edge-AI for Autonomous Modulation of Sleep with Wearable Neurotechnology
Laura Lewis, IMES, RLE, EECS
Suraj Cheema, RLE, EECS, DMSE
With the rapid advancement of AI, there is a barrage of models for making decisions with partial and noisy information. Understanding how context-sensitive AI can be built while considering diverse users is becoming very important, especially in applications where personalized health interventions need to be allocated to a diverse user population. In mental health platforms, AI can provide many benefits of scalability and impact, but given the diverse mental health needs of users, there are serious concerns about the current models creating disparate impact. The goal is to develop new AI models and algorithms for trajectory data on user health and user engagement, that can tackle concerns due to (i) missing data, (ii) unequal length of user interactions, (iii) active learning techniques with costly sensitive information, and (iv) allocating cautious interventions for the diverse set of users to maintain the population needs. - AI-Driven Analysis of Sleep Architecture with Application to Mental Health and Antidepressant Use
Dina Katabi, EECS
Depression remains one of the leading causes of disability worldwide, yet clinicians still lack objective tools to monitor the effectiveness of antidepressant treatment. Current standard care relies heavily on patient questionnaires and clinical interviews, which are subjective and offer limited insight into whether medications are taken as prescribed or how they affect the patient’s physiology. Self-reporting can also be unreliable or inaccessible, especially for individuals experiencing withdrawal or older adults with cognitive impairments. Thus, this project propose a practical, science-based solution: an AI model that analyzes sleep data as a “digital biomarker” for antidepressant use and mental health. The long-term vision is to transform mental health care into a proactive and precision-guided system, enabling individuals to receive tailored, real-time support in their daily lives—ushering in a new standard of home-based, data-informed care. - AI (LOOPS) Linking Organisms, Outcomes, Pollution, and Systems
Sara Beery, EECS, CSAIL
Marzyeh Ghassemi, EECS, LIDS, IMES
Social determinants of health (SDOH) – the conditions under which a person lives, works, and ages – are intricately linked with physical health. For instance, people in areas with high noise pollution, traffic congestion, or pollutants are at an increased risk for cardiovascular disease, diabetes, and dementia. While this link between the environment and human health is known, the interplay between the two is difficult to tease apart, and the environmental factors that cause increased rates on noncommunicable diseases, such as cancer clusters, are hard to identify. Characterizing the relationship between different axes of the environment and human health can aid in the development of public policy to improve public health outcomes. This study aims to build a diverse dataset at a granular, census-tract level to understand how different environmental exposures affect health outcomes in neighborhoods across the United States. Researchers will compile this dataset across government, academic, and non-profit sources to cover environmental metrics, social determinants of health, and health outcomes. This multidisciplinary and intersectional approach will allow for identification of inequities at the geographic, environmental, and health level. Further, researchers will test the ML models developed across different time and geographic domains to understand how environmental factors can identify and forecast health events. - Modeling and Improving Individual Wellness with Multimodal AI
Pattie Maes, MIT Media Lab
Paul Liang, MIT Media Lab and EECS
Current user interfaces and systems, including conversational AI, fundamentally lack an understanding of human wellness in context—how wellness goals guide actions, how routines sustain or erode wellness, and how environments shape decisions. This gap limits their ability to respond meaningfully to the dynamics of daily life, particularly in real-time, mobile settings where the system should respond with relevant information and services without requiring elaborate, explicit commands. To address this, this project proposes that AI-based user interfaces move past surface-level inputs and begin modeling human wellness. One way to model human wellness is through capturing holistic experiences from the user’s perspective, i.e., egocentric capturing, where behavioral signals are extracted through multimodal information via wearable cameras, physiological sensors, and voice I/O devices. This project proposes a new framework where multimodal AI systems adaptively sense individuals’ wellness states over time, enabling proactive, context-aware, and history-informed interactions while remaining efficient for real-time on-edge deployment. These wellness signals into structured representations that capture lived episodes, habitual routines, and procedural action flows. Then, using context in real-time, the systems can detect deviations and deliver proactive information and interventions aligned with the user’s wellness goals. The resulting context-aware, conversational, and personalized AI system can enable several novel applications, such as proactive habit support and cognitive support for older adults.