Our federated NLP paper has been accepted to JMLR

A three-year collaboration

I’m delighted to share that our paper, “Attention-Guided Federated Learning for Privacy-Preserving Clinical NLP,” has been accepted for publication in the Journal of Machine Learning Research (JMLR). This work is the culmination of a three-year collaboration with clinical partners across four hospital systems.

What the paper shows

The paper introduces a federated training scheme that lets clinical language models learn from distributed patient records without those records ever leaving their host institution. Key contributions include:

• A lightweight attention-sharing protocol that reduces communication overhead by roughly 40% compared to standard federated averaging.
• A formal differential-privacy analysis with tight utility bounds.
• Evaluation across four real clinical NLP tasks, including de-identification and readmission-risk prediction.

Why it matters

Healthcare data is some of the most sensitive data we work with, and centralizing it for model training is often legally or ethically infeasible. This line of work aims to make privacy-preserving collaboration the default, not the exception, in clinical machine learning.

Acknowledgments

Huge thanks to my co-authors, Lin Chen and Tunde Okafor, and to our clinical collaborators for their patience through several rounds of revision. A preprint will be posted shortly; the camera-ready version and code will follow.