Interpretable CNN for Single-Channel Artifacts Detection in Raw EEG Signals

Abstract

Electroencephalogram (EEG) signals recorded from the scalp are often affected by artifacts. Most existing artifact detection methods rely on multi-channel statistics such as inter-channel correlation. Recently, there has been a growing interest in realizing single-channel EEG systems to promote everyday use, demanding novel artifacts detection techniques. This paper presents validation results for single-channel artifacts detection in raw EEG signals using four neural architectures: a one-dimensional CNN (1D-CNN) - proposed by us, EEGNet, SincNet and EEGDenoiseNet. We used semi-synthetic EEG data corrupted with Ocular (EOG) and Myo-graphic (EMG) noise components to validate the approaches. Precisely, we contaminated the randomly chosen EEG channels with EOG and EMG artifacts in a controlled manner using a predefined Signal-to-Noise Ratio (SNR) such that the ground truth is known. We validated these models both in terms of classification performance and the interpretability of the learned features. Of the four models, 1D-CNN, EEGNet, and SincNet achieved a comparable classification accuracy (around 99%) and EEGDenoiseNet achieved as low as 64%. Analysing the learned filters for interpretability, we found both 1D-CNN and EEGNet clearly separates EOG (Delta and Theta) and EMG (Gamma) frequency bands from EEG. Instead, SincNet prioritized to learn EEG-specific features (Alpha and Beta) rather than artifact-related information still achieiving the comparable performance as the other two models. EEGDenoiseNet with kernel width of 3 was excluded from this evaluation as it is practically infeasible to perform FFT analysis. Finally, we also computed the number of training parameters for each model to evaluate which among them would be suitable for a resource-constrained wearable device and we found that 1D-CNN and SincNet are the most parameter-efficient, although not by a large margin.