While artificial intelligence offers immense potential for personalized healthcare, its application is hindered by strict data privacy regulations, high cloud-computing latency, and the extreme hardware heterogeneity of edge devices. Although recent breakthroughs in resource-constrained Federated Learning (FL) address some of these limitations, they currently rely on unimodal data. This project extends these techniques to handle the multimodal complexities of healthcare data, essentially focusing on the rich, unutilized physiological signals from smartphones and wearables. By performing efficient on-device training, raw data never leaves the user’s device, ensuring strict privacy while utilizing local updates to collaboratively improve a global model. Ultimately, this project aims to develop an efficient, privacy-preserving FL framework that delivers proactive, equitable, and personalized care to a diverse patient demographic.

I am participating in SuperUROP because I want to build practical, distributed AI systems that put privacy first. After taking multiple ML courses and learning about just how much data is required for training AI models, I see the importance of building frameworks that protect users in sensitive fields like healthcare. Through this project, I also look forward to learning how to optimize these complex models for the severe hardware constraints of edge devices.