Predicting Cognitive Workload in Visuospatial Tasks Using Pupillometry: A Machine Learning Approach

Abstract

Monitoring and controlling cognitive workload are critical for maintaining smooth performance in complex visuospatial tasks that require high mental effort. This study utilizes pupillometry in conjunction with machine learning models to predict cognitive workload during a Lego building task. Twenty university students participated in the complex Lego building task, and their pupil diameter data were collected using eye-tracking glasses. After the data were preprocessed and normalized, the K-means clustering method classified baseline-corrected pupil data into moderate and high cognitive workload levels. Random forest, XGBoost, and Long Short-Term Memory (LSTM) models were trained to predict workload levels. The findings reveal that LSTM outperforms the other two models, achieving better evaluation scores and higher accuracy. This study demonstrates the usability of pupillometric measurements and machine learning in predicting cognitive workload, benefiting professionals in various domains with complex visuospatial tasks.