Explainable AI for EEG-based classification of inner, pronounced, and visualized tasks

Electroencephalography (EEG) is a widely used non-invasive technique for monitoring neural activity, yet its practical impact in clinical and assistive contexts is constrained by the requirement for specialized expert interpretation. This dependence on manual analysis generates significant delays in neurological evaluation, limits scalability in public healthcare systems, and restricts timely access to neurocognitive assessment. Although AI-based approaches can automate aspects of EEG interpretation, most existing solutions lack explainability and show limited cross-subject generalization—two constraints that undermine trust, clinical adoption, and equitable large-scale deployment. Explainable AI models, supported by appropriate preprocessing, windowing, and feature-selection strategies, can assist preliminary EEG interpretation while preserving neurophysiological interpretability and improving performance across subjects. Using a high-density EEG dataset from 10 participants, we evaluate five classical Machine-Learning classifiers—Random Forest, Naive Bayes, Gradient Boosting, Multi-Layer Perceptron, and Support Vector Machine—to discriminate between inner speech, pronounced speech, and visualized cognitive imagery. We systematically analyze how windowing strategies, normalization procedures, and feature-selection methods affect classification performance. Model explainability is examined with SHAP and LIME to identify task-relevant features and cortical patterns. Normalization significantly improves cross-subject performance, while within-subject models achieve higher accuracy, indicating that inter-subject variability remains a central challenge for scalable EEG-based systems. SHAP and LIME analyses reveal feature distributions consistent with theoretical expectations for each task, confirming that the models capture neurophysiologically meaningful distinctions. This study offers a proof of concept for interpretable AI-driven EEG systems capable of supporting citizen-centered applications such as neurorehabilitation, assistive communication, and early triage in public healthcare. By enhancing transparency and generalization, the findings advance the development of scalable neuro-assistive technologies designed to reduce diagnostic bottlenecks, expand accessibility, and strengthen citizen well-being.