A New EEG Emotion Detection Model Based on Multichannel Binary Transition Pattern

Abstract

: Electroencephalography (EEG) emotion recognition attempts to infer affective state from brain activity. This work proposes multichannel Binary Transition Pattern (mBTP), a BP-derived descriptor that converts center-to-neighbor vector differences into ordered channel transitions and transition-table features. We also introduce a text-elicited EEG emotion dataset collected from 49 participants, with 245 records divided into 4895 15-s segments representing five emotions. The resulting explainable feature-engineering pipeline combines mBTP with cumulative weighted neighborhood component analysis (CWNCA), t-algorithm-based k-nearest neighbors (tkNN), and Directed Lobish (DLob). On the five-class task, the complete pipeline reached 88.50% accuracy under segment-level 10-fold cross-validation and 75.12% under leave-one-record-out (LORO) validation. Controlled ablations yielded 82.29% with BP, 84.11% with LBP, 86.09% after replacing CWNCA with NCA, and 86.82% after replacing tkNN with a single kNN. These results describe internal segment-and record-level validation and should not be read as subject-independent LOSO performance. Overall, mBTP provides a compact and interpretable EEG representation, while repeated-seed testing, participant-disjoint LOSO, and matched public-benchmark evaluation remain important next steps.