FedCare: Privacy-Preserving Clinical Language Models
A multi-institution collaboration building federated NLP models for clinical decision support without centralizing patient data.
Overview
FedCare is a multi-year, multi-institution collaboration across four hospital systems, building clinical NLP models — for de-identification, readmission-risk prediction, and discharge summarization — that train collaboratively without any patient record ever leaving its host institution.
Approach
The project uses an attention-guided federated learning protocol (see our JMLR paper) that shares compressed attention summaries between institutions instead of full model gradients, cutting communication overhead by roughly 40% while maintaining a formal differential-privacy guarantee.
Team & collaborators
Two PhD students, one postdoctoral researcher, and clinical collaborators at four partner hospital systems. Funded by the National Institutes of Health (R01).
Status
Active. Year 3 of a 5-year award, currently validating the protocol on a fifth clinical task (adverse event detection) ahead of a planned multi-site deployment.
Back to all research
Abhiraj Bibhar
Department of Sociology and Anthropology
SRM University - Andhra Pradesh