Artifact Removal of Eye Tracking Data for the Assessment of Cognitive Vigilance Levels

Abstract

In this paper, we present a preprocessing pipeline of Eye tracking data to assess cognitive vigilance levels. We introduced two different levels of vigilance state; alertness and vigilance decrement while subjects were performing Stroop Color-Word Task (SCWT) for approximately 45 minutes. We assessed the levels of vigilance by utilizing Eye tracking data and five machine learning (ML) classifiers. Our preprocessing pipeline consists of baseline correction, and artifacts, and noise removal. We extracted six features namely: fixation duration, pupil size, saccade duration, saccade amplitude, saccade velocity, and blink duration. These features were then used as an input to the five ML classifiers for vigilance level classification. We achieved the highest classification accuracy of 76.8% in differentiating between the two vigilance levels using all features with a selected Support vector machine classifier. Other classifiers have also achieved comparable accuracy.