Learning Signal Representations for EEG Cross-Subject Channel Selection and Trial Classification

Abstract

EEG is a non-invasive powerful system that finds applications in several domains and research areas. Most EEG systems are multi-channel in nature, but multiple channels might include noisy and redundant information and increase computational times of automated EEG decoding algorithms. To reduce the signal-to-noise ratio, improve accuracy and reduce computational time, one may combine channel selection with feature extraction and dimensionality reduction. However, as EEG signals present high inter-subject variability, we introduce a novel algorithm for subject-independent channel selection through representation learning of EEG recordings. The algorithm exploits channel-specific 1D-CNNs as supervised feature extractors to maximize class separability and reduces a high dimensional multi-channel signal into a unique 1-Dimensional representation from which it selects the most relevant channels for classification. The algorithm can be transferred to new signals from new subjects and obtain novel highly informative trial vectors of controlled dimensionality to be fed to any kind of classifier.