EEG-based emotion recognition using phase-space reconstruction with Poincaré sections: a study on the AMIGOS dataset.

This study developed a novel, fully reproducible EEG-based framework for binary emotion recognition that combines phase-space reconstruction with Poincaré sections to capture the nonlinear dynamics of brain activity during prototypical emotional states. The method was applied to the publicly available AMIGOS dataset. EEG recordings from 33 participants were downsampled to 128 Hz, bandpass-filtered (4-45 Hz), cleaned of ocular and muscular artifacts using independent component analysis (ICA), and segmented into 1-s non-overlapping windows. Strict labeling thresholds (valence ≥ 6 and arousal ≥ 6 for Happy; valence ≤ 4 and arousal ≤ 4 for Sad) were enforced to isolate extreme high-valence/high-arousal (HVHA) versus low-valence/low-arousal (LVLA) states. A hybrid feature set integrating Poincaré-derived geometric measures with classical spectral power and frontal asymmetry indices underwent rigorous two-stage selection. The final support vector machine with radial basis function kernel (SVM-RBF) achieved 98.21 ± 0.54% accuracy, 96.42 ± 1.06% sensitivity, and 100% specificity in strict subject-independent 7-fold cross-validation. Symmetric selection of 14 channels significantly enhanced feature separability (paired Wilcoxon signed-rank test, Bonferroni-corrected p = 7.4 × 10-8). Independent validation on the DEAP dataset using the identical pipeline yielded 97.68% accuracy, confirming generalizability. The near-perfect performance is specific to binary classification of extreme affective quadrants and does not extend to standard 4-class tasks (81.7%). These findings demonstrate the physiological relevance of nonlinear geometric analysis for detecting prototypical joy versus sadness, with potential clinical utility in automated depression screening.