<p>Advances in deep learning have enabled highly accurate classification and prediction across domains. However, these “black box” models offer limited insight into their decision-making processes. Existing deep models for EEG-based emotion recognition achieve high performance but lack interpretability. This hampers model understanding, trustworthiness and identification of neurophysiological patterns related to emotions. We developed an explainable Grouped Deep Echo State Network (GDESN) for classifying emotions from EEG data. The GDESN leverages the representation power of deep learning while maintaining efficiency. Local and global interpretation techniques were applied to provide model explanations. Local Interpretable Model-Agnostic Explanations (LIME) provided local explanations around predictions by learning an interpretable linear model approximating the GDESN. Eli5 generated global explanations through feature importance weights, uncovering behavior drivers. On the DEAP dataset, the GDESN achieved state-of-the-art arousal accuracy of 89.32% and valence accuracy of 91.21%. Model interpretation identified gamma and theta frequency bands, and electrodes FP1, T8 and Oz as key contributions to emotion recognition. The explainable GDESN demonstrates potential for transparent analysis of emotions from neuroimaging and provides insights into underlying neurophysiological patterns. Overall, the study addresses limitations of non-interpretable “black box” models for this task.</p>