Explainable EEG Microstate Classification for Cognitive Workload Estimation: A Multi-method XAI Analysis Using SHAP, LIME, DiCE, and Anchors

Cognitive workload (CWL) estimation from EEG is central to neuroergonomics and safety-critical human-machine interaction, yet high-performing models often lack interpretability. This paper applies four complementary explainable AI (XAI) methods, SHAP, LIME, DiCE, and Anchors, to a deep bidirectional LSTM classifier trained on EEG microstate features extracted from the STEW dataset, which consisted of 48 participants across two CWL conditions. Eight amplitude and temporal features were extracted per microstate across 14 EEG channels, and the classifier achieved over 84% accuracy and F1-score under both low and high CWL conditions. XAI analysis revealed consistent patterns across methods: Zero Crossing Rate, RMS, and Energy were the most informative features, with right frontal channels (FC6, F4) playing a critical role under high cognitive demand. Counterfactual analysis highlighted substantially larger feature changes required for class transitions under high CWL, while Anchor rules demonstrated near-perfect precision but limited coverage. The convergence of findings across four XAI methods strengthens confidence in the identified neurophysiological markers and advances interpretable, workload-aware AI for EEG-based cognitive monitoring.