View in EDS

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

Saved in:
Bibliographic Details
Title: A Privacy-Preserving and Explainable Approach for Possible Epilepsy Seizure Detection
Authors: Tavolato, Paul, Schölnast, Hubert, Eigner, Oliver, Santone, Antonella, Cesarelli, Mario, Martinelli, Fabio, Agnello, Patrizia, Petyx, Marta, Mercaldo, Francesco
Publisher Information: Springer, 2025.
Publication Year: 2025
Subject Terms: 102016 IT-Sicherheit, 102020 Medical informatics, 102020 Medizinische Informatik, 102016 IT security
Description: Epileptic seizure detection using electroencephalogramdata plays a crucial role in the timely diagnosis and treatmentof epilepsy. Existing deep learning models often suffer fromtwo critical limitations that hinder their clinical applicability:lack of explainability and violation of patient data privacy. In thispaper, we propose a novel privacy-preserving and explainable tecniquefor possible seizure detection that addresses both challengessimultaneously. The proposed method transforms EEG timeseriesdata into image representations, enabling the use of VisionTransformers for classification purposes. To ensure patient dataconfidentiality, we adopt a Federated Learning paradigm thatallows multiple clients (e.g., different hospitals) to collaborativelytrain a global model without sharing raw data. Furthermore, weintegrate an explainability module based on Attention Rollout tovisualize the decision-making process of the model and highlightthe regions of the EEG image most influential in the prediction.Experimental results on a publicly available dataset demonstratethat the proposed approach achieves interesting classificationperformance while preserving privacy and providing predictionexplainability, with the aim to boost the application in real-worldclinical environments of deep learning.
Document Type: Conference object
Language: English
Access URL: https://ucrisportal.univie.ac.at/de/publications/74ded268-ee3d-47d3-ad4b-b84ffb0a3668
Accession Number: edsair.dris...00911..cf20e51f0cdf671ed7712bb33ed96279
Database: OpenAIRE
Description
Abstract:Epileptic seizure detection using electroencephalogramdata plays a crucial role in the timely diagnosis and treatmentof epilepsy. Existing deep learning models often suffer fromtwo critical limitations that hinder their clinical applicability:lack of explainability and violation of patient data privacy. In thispaper, we propose a novel privacy-preserving and explainable tecniquefor possible seizure detection that addresses both challengessimultaneously. The proposed method transforms EEG timeseriesdata into image representations, enabling the use of VisionTransformers for classification purposes. To ensure patient dataconfidentiality, we adopt a Federated Learning paradigm thatallows multiple clients (e.g., different hospitals) to collaborativelytrain a global model without sharing raw data. Furthermore, weintegrate an explainability module based on Attention Rollout tovisualize the decision-making process of the model and highlightthe regions of the EEG image most influential in the prediction.Experimental results on a publicly available dataset demonstratethat the proposed approach achieves interesting classificationperformance while preserving privacy and providing predictionexplainability, with the aim to boost the application in real-worldclinical environments of deep learning.