FT-DropBlock: A Novel Approach for Spatiotemporal Regularization in EEG-based Convolutional Neural Networks

Abstract

The use of Deep Learning (DL) and digital signal processing in Brain-Computer Interfaces (BCIs) to improve the analysis of Electroencephalogram (EEG) data has shown promise, but overfltting of DL models remains a challenge, particularly in Convolutional Neural Networks (CNNs) designed for spatiotemporal data. This paper introduces Features-Time DropBlock (FT-DropBlock), a novel adaptation of DropBlock regularization tailored for EEG processing, targeting the spatiotemporal dimensions represented by features and time points to achieve structured regularization. Contiguous blocks of features and their associated time points are strategically dropped, promoting robust learning by encouraging spatiotemporal coherence and reducing overfitting. To our knowledge, at the time of writing, no previous use of DropBlock in EEG-based CNNs has been reported in the literature. Our approach is the first to use DropBlock to enforce structured regularization across features and time dimensions to preserve spatial consistency within feature maps and ensure temporal regularization. We tested our approach by replacing conventional Dropout with FTDropBlock at selected regularization stages in three EEG-based CNNs. Experimental results conducted on the publicly available BCI Competition IV 2a Dataset show that our approach demonstrates significant improvements over traditional Dropout regularization in classification accuracy and robustness against overfitting, highlighting the effectiveness of this targeted FT-DropBlock strategy for EEG-based CNNs.