A machine learning approach to classify mental workload based on eye tracking data

Purpose: The primary objective of this study was to develop machine learning algorithms for classifying mental workload using eye tracking data.Theory and Methods: This study proposed machine learning algorithms for classifying mental workload using eye tracking data. Four participants performed the N-Back memory task and National Aeronautics and Space Administration -task load index (NASA-TLX) to induce different levels of mental workload. In total, 792 samples were collected from four participants. Twenty-seven eye tracking metrics were selected as independent variables. One output variable reflecting the difficulty level of N-Back memory was classified.Results: As a result of these experiments, it was revealed that almost all eye tracking metrics considered in this study were significantly correlated to both weighted NASA-TLX total score and N-Back memory task difficulty level. As the task difficulty increased, pupil diameter, number of saccades, number of blinks, and blink duration increased, while fixation duration decreased. The results obtained for the two classes of classification problem reached the accuracy of 68% with 14 eye-tracking features due to problem complexity. The results obtained for the two classes of classification problem reached the accuracy of 84% with 27 eye-tracking features as input and the LightGBM algorithm. To determine the degree to which the input variables contribute to the determination of the output variable, a sensitivity analysis was conducted using the gradient boosting machines (GBM) algorithm. The left eye pupil diameter was found to be the most effective metric in the classification of the task difficulty level. The results from the analysis indicate that eye tracking metrics play an important role in the classification of mental workload.Conclusion: The results from the analysis indicate that eye tracking metrics play an important role in the classification of mental workload. Among the classification models developed with machine learning algorithms, the most successful results were obtained with tree-based algorithms. Algorithms such as GBM, LightGBM, XGBoost, which have been developed recently and therefore are not frequently used in the literature on this subject, have achieved better results compared to others.