Combined eye activity measures accurately estimate changes in sustained visual task performance

Abstract

Five concurrent eye activity measures were used to model fatigue-related changes in performance during a visual compensatory tracking task. Nine participants demonstrated considerable variations in performance level during two 53-min testing sessions in which continuous video-based eye activity measures were obtained. Using a trackball, participants were required to maneuver a target disk (destabilized by pseudorandom wind forces) within the center of an annulus on a CRT display. Mean tracking performance as a function of time across 18 sessions demonstrated a monotonic increase in error from 0 to 11 min, and a performance plateau thereafter. Individual performance fluctuated widely around this trend - with an average root mean square (RMS) error of 2.3 disk radii. For each participant, moving estimates of blink duration and frequency, fixation dwell time and frequency, and mean pupil diameter were analyzed using non-linear regression and artificial neural network techniques. Individual models were derived using eye and performance data from one session and cross-validated on data from a second session run on a different day. A general regression model (based only on fixation dwell time and frequency) trained on data from both sessions from all participants produced a correlation of estimated to actual tracking performance of R=0.68 and an RMS error of 1.55 (S. D.=0.26) disk radii. Individual non-linear regression models containing a general linear model term produced the cross-session correlations of estimated to actual tracking performance of R=0.67. Individualized neural network models derived from the data of both experimental sessions produced the lowest RMS error (mean=1.23 disk radii, S.D.=0.13) and highest correlation (R=0.82) between eye activity-based estimates and actual tracking performance. Results suggest that information from multiple eye measures may be combined to produce accurate individualized real-time estimates of sub-minute scale performance changes during sustained tasks.