tACSNet: A Novel Deep Convolutional Autoencoder for Transcranial Alternating Current Stimulation Artifact Removal from Concurrent EEG Signal

Objective: Transcranial Alternating Current Stimulation (tACS) introduces strong artifacts into simultaneously recorded Electroencephalography (EEG) signals, significantly impairing the ability to monitor brain activity during stimulation. This limits the advancement of closed-loop neuromodulation systems. Therefore, effective removal of tACS artifacts from EEG recordings is critically needed. Methods: We propose a deep convolutional autoencoder neural network to reconstruct clean EEG signals from tACS-contaminated EEG recordings, using a publicly available dataset containing both clean and artifact-contaminated EEG. The performance of our model was evaluated by comparing the reconstructed EEG with the ground-truth clean EEG, and benchmarked against traditional methods including template subtraction and adaptive filtering. Results: The proposed model achieved high correlation coefficients with the ground-truth signals-averaging 0.639 in the time domain and 0.880 in the frequency domain-outperforming traditional artifact removal methods. The model was particularly effective in handling high-frequency tACS artifacts (e.g., 40 Hz) compared to lower frequencies (5 Hz and 10 Hz). Conclusion: The proposed deep learning approach demonstrates robust capability in suppressing tACS-induced artifacts and accurately reconstructing clean EEG signals from singlechannel recordings. These findings validate the model's capability to recover meaningful neural activity even under strong stimulation-induced interference. Significance: This work enables monitoring of brain activity during ongoing tACS stimulation, facilitating closed-loop neuromodulation research. Additionally, we provide a preprocessed EEG+tACS dataset and establish a benchmark for future development of deep learning algorithms in this domain.