Measuring the accuracy of ICA-based artifact removal from TMS-evoked potentials

Background: The analysis and interpretation of transcranial magnetic stimulation (TMS)-evoked potentials (TEPs) relies on successful cleaning of the artifacts, which typically mask the early (0-30 ms) TEPs. Independent component analysis (ICA) is possibly the single most utilized methodology to clean these signals. Objective: ICA-based cleaning is reliable provided that the input data are composed of independent components. Differently, in case the underlying components are to some extent dependent, ICA algorithms may yield erroneous estimates of the components, resulting in incorrectly cleaned data. We aim to ascertain whether TEP signals are suited for ICA. Methods: We present a systematic analysis of how the properties of simulated artifacts imposed on measured artifact-free TEPs affect the ICA results. The variability of the artifact waveform over the recorded trials is varied from deterministic to stochastic. We measure the accuracy of ICA-based cleaning for each level of variability. Results: Our findings indicate that, when the trial-to-trial variability of an artifact component is small, which can result in dependencies between underlying components, ICA-based cleaning biases towards eliminating also non-artifactual TEP data. We also show that the variability can be measured using the ICA-derived components, which in turn allows us to estimate the cleaning accuracy. Conclusion: As TEP artifacts tend to have small trial-to-trial variability, one should be aware of the possibility of eliminating brain-derived EEG when applying ICA-based cleaning strategies. In practice, after ICA, the artifact component variability can be measured, and it predicts to some extent the cleaning reliability, even when not knowing the clean ground-truth data.