Decoding Fine-Grained Emotions From EEG: A Multi-Model Framework With Fuzzy Logic-Based Adaptive Intelligence

Abstract

Emotion recognition using Electroencephalogram (EEG) signals has gained significant attention in affective computing due to its ability to capture intrinsic neural responses that are less susceptible to voluntary control and environmental disturbances. Nevertheless, EEG-based emotion recognition remains challenging because of signal noise, inter-subject variability, non-linear dynamics, and the need for interpretable decision-making. To address these challenges, this paper presents a progressive hybrid framework that integrates deep spatio-temporal learning, domain-informed feature modelling, and fuzzy adaptive reasoning for robust and fine-grained emotion recognition in the Valence–Arousal (VA) space. The proposed framework comprises three complementary models. First, a Spatio-Temporal Deep Model (STDM) based on a CNN–LSTM architecture learns spatial and temporal EEG representations in an end-to-end manner for four-class emotion recognition. Second, a Feature-Augmented Deep Model (FADM) enhances this baseline by integrating handcrafted Time and Frequency-domain EEG features with deep representations using weighted linear fusion, improving discriminative capability. Finally, a Fuzzy Logic–Based Adaptive Weighting (FLAW) model is introduced to explicitly model uncertainty and transitional affective states through Low–Moderate–High fuzzy reasoning, enabling fine-grained eight-class emotion recognition with interpretable decision logic. Extensive experiments on the DREAMER benchmark dataset, conducted under rigorous subject-independent evaluation and cross-subject validation protocols, demonstrate a clear performance progression across the models. The STDM achieves an accuracy of 86.8%, the FADM improves performance to 93.5%, and the FLAW model attains the best accuracy of 94.3%, with valence and arousal accuracies of 96.3% and 94.2%, respectively. Comparative analysis with existing methods on the DREAMER dataset demonstrates the effectiveness of the proposed framework under subject-independent evaluation. The results suggest that integrating deep spatio-temporal learning with adaptive fuzzy reasoning can improve fine-grained emotion recognition and interpretability within the experimental setting considered.