EEG-Based Classification of Affective Dimensions: Neural Dissociability of Valence, Arousal, and Dominance in Oscillatory EEG Features

How many neural dimensions does emotional experience require? The circumplex model (Russell, 1980) reduced emotion to two axes, valence and arousal, collapsing the PAD model’s third axis, dominance, on psychometric grounds. This paper tests that reduction computationally using the DEAP dataset (N = 32 subjects, 1280 trials, 32-channel EEG). Two findings carry formal or exploratory evidential weight. Beta-only and gamma-only single-band models formally exceed matched null distributions for cross-subject valence classification (both p = 0.010), with no equivalent spectral concentration at any band for arousal or dominance, establishing high-frequency oscillations as the only cross-subject stable spectral signature for valence. Valence-dominance feature importance correlation remains near zero across four independent classifier conditions spanning two architectures and two evaluation regimes (−0.040 to +0.040), with no condition producing the positive overlap expected if the two dimensions shared their neural feature basis, constituting model-independent exploratory evidence for neural dissociability of dominance from valence. Three secondary analyses correctly scope the dataset’s limits: frontal alpha asymmetry is non-determinable on DEAP due to its compressed valence distribution; the cross-subject generalization gap is directionally coherent but not statistically separable at N = 32; and EEG-peripheral configuration differences are directional only. Together, these results constitute exploratory computational evidence that affective space contains more neural structure than a two-dimensional circumplex captures.