Current epileptic detection methods primarily utilize machine learning, wavelet transforms, and multi-feature fusion, yet often inadequately distinguish between interictal and ictal status. To address the limitation, this study proposes a novel enhanced EEG decoding framework that systematically integrates time-domain, frequency-domain, and nonlinear characteristics of epileptic signals. Our method incorporates three key innovations: (1) comprehensive multi-domain feature fusion architecture, (2) multi-head attention mechanism enabling adaptive feature weighting, and (3) multi-task learning architecture for classification of healthy, interictal, and ictal status. Experiments on EEG data from five healthy subjects and five epileptic patients demonstrated robust performance with 94.00% classification accuracy (sensitivity=94.00%, specificity=93.75%). The results exhibited the framework’s capability in discriminative feature extraction and effective identification of clinically significant temporal variations across signal types. The proposed methodology advances EEG decoding precision through its synergistic integration of complementary feature domains and attention-based dynamic learning, establishing new benchmarks for automated epileptic state differentiation.