Identification of EEG Features of Transcranial Electrical Stimulation (tES) Based on eXplainable Artificial Intelligence (XAI)

Electroencephalographic (EEG) features of transcranial Electrical Stimulation (tES) effects in Multiple Sclerosis (MS) patients were identified. Machine Learning and eXplainable Artificial Intelligence (XAI) algorithms were the used methods for the EEG feature selection. Current tES-based treatments lack of adaptivity to their effects on individuals. Real-time modification of electrical stimulation parameters based on the trend of specific EEG features may represent a new perspective in terms of personalized medicine for MS patients. This preliminary study aimed to identify EEG features reflecting the effect of tES treatment. Ongoing analyses are exploring the correlation between the identified EEG features and the expected clinical outcomes. Five MS patients underwent non-pharmacological treatment combining Theory of Mind (ToM) training with tES or sham treatment. Pre-and post-treatment variation in EEG features were assessed both in Eyes Opened (EO) and Eyes Closed (EC) conditions. In particular, absolute and relative power across six frequency bands, and Posterior Dominant Rhythm (PDR) amplitude and frequency were explored. The Sequential Feature Selection (SFS) algorithm in combination with Support Vector Machine (SVM) classifier identified (i) difference of absolute powers in high beta band in T3 channel, (ii) difference of absolute powers in gamma band in T3 channel, and (iii) difference of absolute powers in gamma band in O2 channel in EO condition, and (i) difference of absolute powers in gamma band in T3 channel, (ii) difference of absolute powers in gamma band in C4 channel, and (iii) difference of PDR amplitude in O2 channel in EC condition as the most discriminating tES from sham treatment (67.5 % accuracy in EO and 83.33 % in EC conditions, respectively). The SHapley Additive exPlanations (SHAP) algorithm highlighted PDR amplitudes in O2 and O1 as most informative features.