Decoding Working Memory Capacity from Gaze and Arousal Dynamics using Machine Learning

Abstract Young adulthood is the peak of cognitive performance, yet there exist individual differences in learning, problem-solving, and decision-making abilities. We recorded eye movements and electrodermal activity (EDA) from 89 young adults performing dual n-back tasks to investigate whether psychophysiological signals can reveal these differences. Machine learning (ML) models reliably classified individuals with high versus low working memory capacity (WMC), with eye-tracking features (fixation duration, pupil diameter, saccadic velocity) emerging as dominant predictors and EDA providing complementary arousal-related cues. By extending permutation-based feature importance, we identified a minimal set of psychophysiological measures enabling robust classification across participants. Our subsequent counterfactual ablation of this psychophysiological metric set further revealed an asymmetry wherein low-WMC individuals could be reclassified as high-WMC through perturbations of gaze and arousal features, whereas high-WMC individuals remained resistant to such manipulation. This suggests high WMC may emerge as a stable trait, while low WMC may be enhanced via interventions. Together, these findings establish multimodal ML as a robust framework for quantifying WMC through physiological markers, namely eye and EDA metrics, and highlight promising pathways for enhancing the cognitive capacity of individuals.