Emotional valence through pupil: Machine learning classification under controlled visual complexity and emotional...

Pupillometry has long been proposed as a noninvasive physiological measure for emotional valence. However, its empirical effectiveness remains inconclusive due to confounding visual and emotional factors. This study examined whether pupil response patterns alone can reliably distinguish between positive and negative emotional stimuli while explicitly controlling for visual complexity (spatial frequency; SF) and emotional arousal at three standardized levels. Fifty images (25 positive and 25 negative) were presented, and pupil responses were recorded. Dynamic time warping-based clustering captured temporal variations and similarities in pupil size responses across visual conditions. Initial classification without controlling SF and arousal yielded near-chance accuracy (~57%) despite luminance control. However, performance improved substantially when stimuli were segmented by specific arousal-SF combinations. Under a representative low arousal, high spatial-frequency condition (SF level 4), the best-performing configuration (logistic regression) achieved a mean classification accuracy of approximately 79% and an AUC of 0.88, with consistently high precision, recall, and specificity across cross-validation folds. Feature importance analyses highlighted critical pupillary parameters, including the area under the pupil dilation curve, as key predictors. These results suggest that pupillary responses can reliably indicate emotional valence under rigorously controlled visual conditions, emphasizing control of perceptual and emotional factors in pupillometry-based emotion research.