An Explainable CNN–Transformer Framework for Subject-Independent EEG-Based Emotion Recognition

Abstract

Recognizing emotional states from electroencephalogram (EEG) signals is an active area of research in affective computing and brain–computer interaction. EEG is useful for this task because it records brain activity with high temporal resolution, but the signals are complex and can vary considerably from one person to another. These variations make it difficult to build a model that performs consistently on participants who were not seen during training. In addition, deep learning models can provide accurate predictions without clearly showing which parts of an EEG recording influenced the decision. This paper presents an explainable CNN–Transformer framework for subject-independent EEG emotion recognition. The CNN component learns local temporal and spatial patterns, while the Transformer encoder uses self-attention to model relationships across the learned representation. The two representations are fused before classification, and an attention stage is used to emphasize informative features. The DEAP dataset is considered as the primary benchmark, with EEG-only binary valence classification used as the initial task. A Leave-One-Subject-Out (LOSO) strategy is adopted so that the participant used for testing is excluded from model training. Performance is assessed using accuracy, precision, recall, F1-score, and ROC-AUC. SHAP and Integrated Gradients are considered for examining the contribution of EEG channels and temporal regions. Baseline and ablation experiments are also included to study the value of the individual model components. The numerical values currently included in the manuscript are provisional development-stage values and must be replaced by measurements from the completed DEAP experiment before publication.