This repository contains the source code and documentation for our recent study on the channel adaptation methodology in electroencephalography (EEG) signals using few-shot learning in wearable biomedical systems.
Paper Link: Read the Manuscript
Lab Website: Embedded Systems Laboratory (ESL) at EPFL
In our study, we tackle the challenge of efficiently selecting the optimal set of electrodes for EEG-based seizure detection. Traditional methods require extensive computational resources, exploring limited electrode combinations. We propose a novel approach that dramatically speeds up this process by using a method inspired by few-shot learning, allowing us to evaluate all possible combinations without the need to retrain the network extensively. Our technique reduces what would traditionally take months into just a few hours on a single GPU, achieving high accuracy with significantly less effort.
The research utilizes the TUH EEG Seizure Corpus (TUSZ), an open-source dataset for the seizure detection task.
- Efficient Exploration of Electrode Configurations: By leveraging few-shot learning, our method can explore all potential electrode combinations quickly and effectively.
- High Performance with Reduced Training Time: Achieves performance matching with previous research results but in much less time.
- Practical Application in Epileptic Seizure Detection: Demonstrates the feasibility of our approach in real-world settings, potentially transforming the development and usability of wearable EEG systems.
For more information or to discuss this research, feel free to contact Alireza Amirshahi at Alireza's EPFL profile or Professor David Atienza at Prof. David Atienza's EPFL profile.