Machine learning classification of emotional videos via autonomic arousal based on pupil diameters, electrodermal activity and heart rate.

We evaluated the ability of Electrodermal Activity (EDA), Heart Rate Variability (HRV), and Pupil Diameter (PD) to detect emotional states via machine learning to determine whether simple arousal metrics can distinguish complex emotional labels: Amusement, Awe, Disgust, Nurturant Love, and Sadness. In a laboratory session, participants (N = 47) selected the most appropriate emotional label out of five immediately after viewing 10 short videos. Classifications relied solely on the three psychophysiological measures and by applying three machine learning models: Support Vector Machine (SVM/SVC), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). Features included baseline-corrected average pupil diameter (PD) and electrodermal activity (EDA) peaks. Noisy heart rate variability (HRV) data were excluded to prevent signal-to-noise ratio degradation. PD reflects sympathetic noradrenergic activation and reciprocal coupling within the Autonomic Nervous System (ANS), providing a stable state estimate. In the core 'affect circumplex', SVC achieved peak accuracy, and PD achieved the highest feature importance score across architectures. Results revealed a many-to-one mapping where discrete valence states produced indistinguishable physiological signatures: high-arousal states (Amusement, Sadness) and low-arousal states (Awe, Disgust) were conflated. Notably, Awe triggered PD constriction, mirroring Disgust's profile. Tactical removal of the cardiovascular features significantly improved RF and XGBoost performance. These findings suggest that simple arousal metrics, and specifically pupillometry, can classify affect without using complex time-series analysis. Non-invasive sensors may facilitate the development of sensitive computerized systems calibrated to objective physiological data.