A Privacy-Preserving and Explainable Approach for Possible Epilepsy Seizure Detection

Epileptic seizure detection using electroencephalogram data plays a crucial role in the timely diagnosis and treatment of epilepsy. Existing deep learning models often suffer from two critical limitations that hinder their clinical applicability: lack of explainability and violation of patient data privacy. In this paper, we propose a novel privacy-preserving and explainable technique for possible seizure detection that addresses both challenges simultaneously. The proposed method transforms EEG time-series data into image representations, enabling the use of Vision Transformers for classification purposes. To ensure patient data confidentiality, we adopt a Federated Learning paradigm that allows multiple clients (e.g., different hospitals) to collaboratively train a global model without sharing raw data. Furthermore, we integrate an explainability module based on Attention Rollout to visualize the decision-making process of the model and highlight the regions of the electroencephalographic image most influential in the prediction. Experimental results on a publicly available dataset demonstrate that the proposed approach achieves interesting classification performance while preserving privacy and providing prediction explainability, with the aim to boost the application in real-world clinical environments of deep learning.