Classifying mental workload of esports players using machine learning

Abstract

Electrodermal activity (EDA) peak counts, derived from both tonic and phasic components, are widely used as physiological proxies for mental workload in cognitively demanding tasks, such as esports. However, their specificity remains uncertain, particularly given potential confounding effect of time-on-task. This study analyzes 92 competitive gameplay sessions from a multimodal esports dataset using three decomposition techniques: convex decomposition (cvxEDA), sparse deconvolution (sparseEDA), and time varying sympathetic activity (TVSymp). From each method, phasic, and tonic peak counts (TPC), as well as their normalized rates, were extracted. We examined their relationship with self-reported workload through correlation analyses, partial correlations controlling for session duration, and linear mixed-effects models (LMMs). While both peak types exhibited strong positive correlations with gameplay duration (r=0.915 for phasic and r=0.856 for tonic), their association with perceived workload vanished once time was accounted for. Across methods, TVSymp yielded the highest discriminative validity with an area under curve (AUC) of 0.880 in classifying high versus low workload. Machine learning (ML) classifiers trained solely on EDA-based features under a leave-one-subject-out (LOSO) scheme outperformed multimodal models that incorporated heart rate variability (HRV). These results underscore need to disentangle temporal structure from cognitive signals when interpreting EDA and call into question the assumption that EDA peak counts alone reliably encode mental workload across individuals.