Decoding Motor Excitability in TMS using EEG-Features:An Exploratory Machine Learning Approach

BackgroundWith the burgeoning interest in personalized treatments for brain network disorders, closed-loop transcranial magnetic stimulation (TMS) represents a promising frontier. Relying on the real-time adjustment of stimulation parameters based on brain signal decoding, the success of this approach depends on the identification of precise biomarkers for timing the stimulation optimally. ObjectiveWe aimed to develop and validate a supervised machine learning framework for the individualized prediction of motor excitability states, leveraging a broad spectrum of sensor and source space EEG features. MethodsOur approach integrates multi-scale EEG feature extraction and selection within a nested cross-validation scheme, tested on a cohort of 20 healthy participants. We assessed the frameworks performance across different classifiers, feature sets, and experimental protocols to ensure robustness and generalizability. ResultsPersonalized classifiers demonstrated a statistically significant mean predictive accuracy of 72 {+/-} 11%. Consistent performance across various testing conditions highlighted the sufficiency of sensor-derived features for accurate excitability state predictions. Subtype analysis revealed distinct clusters linked to specific brain regions and oscillatory features as well as the need for a more extensive feature set for effective biomarker identification than conventionally considered. ConclusionsOur machine learning framework effectively identifies predictive biomarkers for motor excitability, holding potential to enhance the efficacy of personalized closed-loop TMS interventions. While the clinical applicability of our findings remains to be validated, the consistent performance across diverse testing conditions and the efficacy of sensor-only features suggest promising avenues for clinical research and wider applications in brain signal classification.