A deep learning approach to artifact removal in Transcranial Electrical Stimulation: From shallow methods to deep neural networks and state space models

Transcranial Electrical Stimulation (tES) is a non-invasive neuromodulation technique that generates artifacts in simultaneous EEG recordings, hindering brain activity analysis. This study analyzes Machine Learning (ML) methods for tES noise artifact removal across three stimulation types: tDCS, tACS, and tRNS. Synthetic datasets were created by combining clean EEG data with synthetic tES artifacts. Eleven artifact removal techniques were tested and evaluated using the Root Relative Mean Squared Error (RRMSE) in the temporal and spectral domains, and the Correlation Coefficient (CC). Results indicate that method performance is highly dependent on stimulation type: for tDCS, a convolutional network (Complex CNN) performed best; while a multi-modular network (M4) based on State Space Models (SSMs) yielded the best results for tACS and tRNS. This study provides guidelines for selecting efficient artifact removal methods for different tES modalities, establishing a benchmark for future research in this area and paving the way for more robust analysis of neural dynamics in advanced clinical and neuroimaging applications.