Machine learning validation of EEG+tACS artefact removal

Kohli, Siddharth; Casson, Alexander J · 2020 · Crossref

DOI: 10.1088/1741-2552/ab58a3

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of validating electroencephalography (EEG) data recorded during simultaneous transcranial alternating current stimulation (tACS). While tACS allows for closed-loop brain stimulation, the resulting high-amplitude stimulation artifacts obscure the low-amplitude EEG signals. Although various artifact removal algorithms exist, there is significant debate regarding their efficacy and the presence of residual artifacts, largely because on-person testing lacks a ground-truth baseline for comparison. The authors propose a novel validation method using machine learning to determine if cleaned EEG data retains genuine neural information rather than being dominated by residual stimulation artifacts. The underlying hypothesis is that if residual artifacts dominated the signal, they would be independent of the cognitive task performed, making it impossible to distinguish between different experimental conditions using machine learning. To test this, ten participants performed two working memory tasks—a visual n-back task and a backward digit recall task—during simultaneous EEG recording and tACS application. The experimental design included baseline periods, task periods with sham stimulation, and task periods with active tACS (5 Hz, 1 mA). An 8-channel EEG system recorded data, which was then processed using a Superposition of Moving Averages (SMA) algorithm to remove tACS artifacts. The cleaned EEG data was segmented into 5-second windows and analyzed using Discrete Wavelet Transform to extract power and Shannon entropy features across theta, alpha, beta, and gamma frequency bands. These features served as inputs for Linear Discriminant Analysis (LDA) classifiers. The study employed four classification models: three subject-dependent models (separating baseline, baseline+stimulation, and specific tasks with/without stimulation) and one subject-independent model trained on pooled data from all participants. The results demonstrated that the LDA classifiers could significantly differentiate between the various experimental conditions. For the four-class models, classification accuracies ranged from 65% to 94%, far exceeding chance levels. Specifically, the classifier successfully distinguished between the n-back and backward digit recall tasks during stimulation, achieving accuracies exceeding 72%. The subject-independent classifier also performed well, with a mean accuracy of 67.7%, which was more than three times the chance level of 18.2%. These findings indicate that the cleaned EEG signals contained distinct neural signatures associated with different cognitive states and stimulation conditions. The significance of this work lies in providing a new, robust method for validating tACS artifact removal algorithms. The ability to classify different cognitive tasks and stimulation states suggests that residual artifacts, if present, do not dominate the cleaned EEG signal. This adds confidence that true neural information is preserved after artifact removal, supporting the feasibility of using simultaneous EEG+tACS for closed-loop applications. The authors conclude that this machine learning-based validation approach should be used alongside existing methods, such as phantom testing and visual inspection, to build greater assurance in the utility of artifact removal techniques.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• drowsiness detection algorithms