Calibration-free Ocular artifact reduction in EEG signals using a montage-independent deep learning model

Abstract

Ocular artifacts (OA) are the most common artifacts in electroencephalography (EEG), significantly affecting signal quality and analysis. Common approaches like indepentent component analysis (ICA) or regression-based methods address this problem but require several minutes of subject-specific EEG and electrooculography (EOG) calibration, making them impractical for real-time applications like brain–computer interfaces (BCI). In this study, we introduce EEGOAR-Net, a deep learning architecture designed to reduce OA in EEG. It address these issues while also providing flexibility across various EEG montages. Based on U-Net architecture, EEGOAR-Net was trained with contaminated EEG signals in order to reconstruct them with OA attenuated, using SGEYESUB as the reference method. In addition, a novel training methodology based on masking signals from different channels was applied to make EEGOAR-Net independent of the EEG montage used. A cross- validation analysis was conducted to assess EEGOAR-Net’s performance, demonstrating its ability to reduce EEG-EOG correlations to chance levels across most brain regions with minimal information loss. Thus, the performance of EEGOAR-Net is comparable to that of the reference method without the need for subject- specific calibration or EOG channels. Furthermore, validation on an additional dataset confirmed effective blink reduction and superior preservation of neural information compared to the state-of-the-art models: 1D- ResCNN and IC-U-Net. EEGOAR-Net’s performance across datasets and versatility across montages prove it to be a reliable and practical solution for EEG-based research and BCI applications, ensuring a notable reduction of OA on signal while maintaining the integrity of neural information.