A Machine Learning Pipeline for Evaluating Pre-Stimulus EEG Features and Their Impact on Post-Stimulus TMS-EEG Responses

Transcranial magnetic stimulation (TMS) combined with electroencephalography (EEG) offers a direct window into cortical excitability and connectivity. Among several TMS-evoked potential (TEP) metrics, the absolute area under the curve (AUC) reflects the total energy of the post-stimulus response. However, the extent to which pre-stimulus neural features influence these responses remains unclear. This study presents a transparent machine learning pipeline to systematically evaluate how pre-stimulus spectral power and complexity measures predict post-stimulus TEPs. Methodological choices-including scaling techniques and feature selection methods (mutual information, LASSO, f-regression)-were rigorously compared and validated. The results demonstrate that gamma band power and Higuchi Fractal Dimension (HFD) consistently yield the highest predictive relevance. The best performance was achieved using the full feature set, suggesting a complementary role of spectral and nonlinear features. These findings emphasize the importance of methodological transparency and robust validation in TMS-EEG analysis. Future work may benefit from incorporating subject-specific modeling and connectivity-based features to enhance both prediction and neuroscientific insight.