Multi-scale feature fusion and neural architecture search for electroencephalography-based affective-state recognition in public benchmark datasets

Introduction Electroencephalography (EEG)-based affective-state recognition may provide an objective neurophysiological complement to self-report scales and behavioral observations, but robust modeling remains difficult because EEG signals are non-stationary, temporally heterogeneous, and strongly subject-dependent. Methods This study evaluated MFANet, a multi-scale feature fusion and neural architecture search-enhanced framework for binary classification of valence, arousal, and dominance. MFANet combines parallel one-dimensional convolutional branches with different receptive fields, channel-wise attention for adaptive feature recalibration, and a neural architecture search-guided long short-term memory module for temporal modeling. The framework was evaluated on the public DEAP and DREAMER datasets using five-fold, trial-grouped subject-dependent cross-validation in which all windows from one stimulus trial remained in the same outer fold and all data-dependent operations were confined to the corresponding training data. Results In DEAP ablation experiments, the complete model achieved mean accuracies of 97.20 ± 1.63% for valence, 97.32 ± 1.27% for arousal, and 97.14 ± 1.21% for dominance. Compared with the basic convolutional-recurrent baseline, multi-scale feature fusion and the searched recurrent module improved average accuracy by 10.17 and 10.87 percentage points, respectively, corresponding to relative gains of 12.5% and 13.4%. In participant-level paired comparisons based on the archived participant-level values, MFANet exceeded the dynamic graph convolutional neural network (DGCNN) by 4.13, 3.39, and 3.71 percentage points; bootstrap 95% confidence intervals were 3.29–5.15, 2.56–4.38, and 2.74–4.85, and all Holm-adjusted p-values were <0.001. A DEAP channel-count sensitivity analysis indicated that the full 32-channel montage changed accuracy by less than 0.4 percentage points relative to the prespecified 14-channel montage. Discussion These findings suggest that joint multi-scale representation learning and adaptive recurrent architecture selection can improve benchmark EEG affective-state recognition. Because the analysis used public healthy-volunteer datasets and self-reported affective labels, the framework should be regarded as a computational method for future clinical validation rather than a diagnostic tool.