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

Brain state-dependent transcranial magnetic stimulation (TMS) holds promise for enhancing neuromodulatory effects by synchronizing stimulation with specific features of cortical oscillations derived from real-time electroencephalography (EEG). However, conventional approaches rely on open-loop systems with static stimulation parameters, assuming that pre-determined EEG features universally indicate high or low excitability states. This one-size-fits-all approach overlooks individual neurophysiological differences and the dynamic nature of brain states, potentially compromising therapeutic efficacy. We present a supervised machine learning framework that predicts individual motor excitability states from pre-stimulus EEG features. Our approach combines established biomarkers with a comprehensive set of spectral and connectivity measures, implementing multi-scale feature selection within a nested cross-validation scheme. Validation across multiple classifiers, feature sets, and experimental protocols in 50 healthy participants demonstrated a mean prediction accuracy of $71 \; \pm \; 7$ %. Hierarchical clustering of top predictive EEG features revealed two distinct participant subgroups. The first subgroup, comprising approximately 50% of participants, showed predictive features predominantly in alpha and low-beta bands in sensorimotor regions of the stimulated hemisphere, aligning with traditional associations of motor excitability and the sensorimotor $\mu $ -rhythm. The second subgroup exhibited predictive features primarily in low and high gamma bands in parietal regions, suggesting that motor excitability is influenced by broader neural dynamics for these individuals. Our data-driven framework effectively identifies personalized motor excitability biomarkers, holding promise to optimize TMS interventions in clinical and research settings. Additionally, our approach provides a versatile platform for biomarker discovery and validation across diverse neuromodulation paradigms and brain signal classification tasks.