XAI-Based Identification of EEG Features of Inhibition and Working Memory Activation

The most informative electroencephalographic (EEG) features related to basic Executive Functions (EFs) were identified by exploiting the SHapley Additive exPlanations (SHAP), an eXplainable Artificial Intelligence (XAI) method. In particular, Working Memory (WM) and Inhibitory Control (IC) were tested by exposing participants to N-Back and Go-NoGo tasks. Currently, eXplainable AI (XAI) methods are scarcely investigated in the framework of EEG signal processing. Specific EEG features of EFs activation are still poorly explored. EEG data from 13 healthy participants were gathered during rest condition and cognitive task executions. In particular, N-Back (activating WM) and Go-NoGo (activating IC) tasks were administered with two difficulty levels. SHAP identified Absolute Delta Power in Fp1 and Fp2 as the two most significant features for both WM and IC, when the two experiments were analyzed separately. The same two features emerged as the most significative when SHAP was applied on a classification problem involving the two tasks at the highest difficulty level. Absolute power of Delta Fp1 and Fp2 resulted statistically significantly lower in the Go-NoGo task compared to the N-Back. These results are consistent with a condition of suppression of irrelevant sensory stimuli common to the two tasks and particularly to the more challenging N-back task. The research contributes to the development of solutions for the cognitive rehabilitation of patients suffering from neurodegenerative diseases.