Abstract
Cognitively challenging tasks require complex coordination of information beyond visual input. Predicting accuracy on such tasks has potential applications in education and industry. Task difficulty is associated with increases in reaction time and variation in eye tracking indices. Critically, machine learning has not yet been used to predict accuracy on cognitive tasks with multiple difficulty levels. We report data on 57 (34 females; 20-30 years) participants who completed visuospatial tasks of mental attentional capacity with six levels of difficulty while their eye movements were recorded using EyeLink Portable Duo SR Research eye-tracker with 1ms temporal resolution (at 1000 Hz frequency) in remote head-free-to-move mode. Results show that task accuracy scores can be robustly predicted when all variables (e.g., eye-tracking, difficulty level and reaction time) are considered together (R2 = .80). Reaction time, difficulty level and eye tracking metrics are also effective independent predictors with R2 equaling .73, .58, and .36, respectively. Analyses for feature importance suggest eye-tracking indices with the most importance for the models include the number of fixations, number of saccades, duration of the current fixation and pupil size. Notably, our machine learning algorithms target a prediction question, rather than a classification one, and the current algorithm can be useful for future research and applications in other contexts where visuospatial processing is required. Theoretically, findings show common and distinct metrics that can inform theories of cognition and vision science.