Federated Learning Approach for Explainable Multilayer Graph Neural Networks

Stroke outcomes vary widely, and predicting severity early is critical for guiding treatment. EEG offers a fast, bedside window into brain function, but turning it into reliable, interpretable predictions is a challenge, especially when patient data cannot be freely shared across hospitals.

Our team built a multilayer Graph Neural Network (GNN) that learns from EEG connectivity patterns to predict stroke severity (NIHSS). The model not only reaches high accuracy but also explains why it makes each prediction, highlighting the brain connections most responsible.

To respect privacy, we scaled this system across hospitals using our federated learning platform CAFEIN. This lets centers collaborate without exchanging raw EEG: each site trains locally, while global knowledge is shared securely. The result? Centralized-level performance, but with data staying at its source.

Our role in this project

Our role in this project is to accurately predict stroke severity from resting-state EEG, while providing clear, connection-level explanations that clinicians can easily audit. At the same time, we ensure that the training process remains privacy-preserving across multiple hospitals, achieving performance comparable to pooled data.

Collaborators