Explainable self-supervised learning unveils neurophysiological signatures of proficiency in imagined speech

This study explores the neural correlates of individuals based on their proficiency in Imagined Speech (IS), a cognitive process with profound implications for brain–computer interfaces and assistive technologies. Our approach combines advanced signal processing techniques, machine learning, Explainable Artificial Intelligence (XAI), and topographical analysis, in order to unveil specific EEG rhythms and cortical regions crucial for successful IS production. Our findings demonstrate that these groups can be differentiated with high accuracy (0.884 on our held-out test set) using a combination of complex measures, fractal behaviors, and spectral characteristics derived from EEG signals. Furthermore, the application of XAI, combined with topographic analysis, allowed us to differentiate EEG epochs between two groups with high and low proficiency level respectively highlighting the neurophysiological differences between the two groups.