Abstract

Cognitive load detection in educational environments remains a critical challenge for adaptive learning systems, with traditional methods being intrusive and disruptive to the learning process. This study addresses this gap by validating a facial emotion recognition (FER) system for detecting potential indicators of cognitive load in controlled educational settings. Using the DeepFace model, the system captured and analyzed 29,834 emotional data points from eight participants in a controlled, simulated environment over a structured learning session. A descriptive analysis revealed a high prevalence of neutral (27%) and sad (26%) states, with significant differences between emotion means confirmed by ANOVA (F(6,42) = 2799.73, p < .001, ηp² = .998). Additionally, a Kruskal-Wallis test (H = 15569.94, p < 0.001) further corroborated these significant differences in emotion distributions, reinforcing the nonparametric nature of the data. Furthermore, time series analysis, including the autocorrelation function (ACF), showed a consistent ascending trend and significant temporal dependency in sadness probabilities. This study makes three key contributions: (1) validates sadness as a reliable proxy for cognitive load in controlled environments (r = -0.33 with neutrality, p < .001), (2) demonstrates temporal persistence of cognitive fatigue through autocorrelation analysis, and (3) establishes a functional prototype at Technology Readiness Level 3 for future educational applications. The results validate the system as a functional laboratory prototype andunderscore a critical challenge for affective computing in education: the need to distinguish between genuine emotional expressions and indicators of cognitive effort. These findings provide the empirical foundation for developing non-intrusive cognitive load monitoring systems in real classroom environments.