How Can We Train AI Models Together Without Sharing Patient Data

What is it about?

Strict privacy rules stop hospitals from pooling patient records in one place. Federated learning lets each hospital train locally and share only model parameters, not raw data. Most frameworks, however, support deep neural nets that are resource-hungry and hard to interpret. Our work builds the first Random-Forest-based federated toolkit in PySyft so hospitals and labs can train transparent, CPU-friendly models collaboratively while fully complying with GDPR. Tests on AIDS and diabetic-retinopathy datasets showed the federated model stays within 0.3–9 percentage points of a centralised baseline, proving that tree-based methods can compete without compromising privacy.

Featured Image

Photo by NASA on Unsplash

Photo by NASA on Unsplash

Why is it important?

- Brings interpretability to federated AI. Clinicians can inspect feature importance instead of trusting a “black box.” - Slashes hardware costs. Random Forests run on commodity CPUs—no GPUs or cloud accelerators needed—making privacy-preserving analytics affordable for smaller hospitals. - Closes a tooling gap. Until now, federated-learning libraries lacked native support for tree-based models; our open-source package (fed_rf_mk) fills that void and is already live on GitHub and PyPI. - Sets the stage for safer data sharing. By keeping data on-premise, hospitals meet GDPR and HIPAA obligations while still benefiting from multi-site evidence.

Perspectives