Assessing and predicting human workload is an important consideration in the design process of new systems, the modification of existing systems, or for the purposes of alleviating or avoiding task overload in real time through task reallocation or adaptive automation. Previous research has established that truly adaptive systems require information on the human operator's workload levels in real time, as it is difficult to reliably predict actual workload based upon a priori modelled estimates alone. The combination of several eye activity measures may provide a psychophysiological estimate of workload for some tasks. Pupil diameter has been shown to increase with higher cognitive workload, while blink rate and duration decline as a function of greater visual demands imposed by a task. Likewise, fixation parameters (dwell time and frequency) also show workload-related changes. Here, we demonstrate signal processing methods that make possible the identification of eye measures most sensitive to task demands. We then assess the sensitivity of non-linear and artificial neural network (NN) models using combined eye measures to predict moment-to-moment fluctuations in task workload. Eleven subjects completed four 2 h blocks of a visuo-spatial memory task in which they were required to examine and remember the classification (friend/enemy) of targets moving inward towards two ship icons presented on a computer display. Subjects were required to prosecute each target (fire upon/allow to pass) when targets passed between two range rings surrounding each ship. Between one and nine targets could be simultaneously present on the display. Eye activity measures were recorded at 60 Hz from a near infrared eye tracking system. For each participant, moving estimates of blink frequency and duration, fixation dwell time and frequency, and pupil diameter, integrated over periods of 20 s or less, were obtained every 2 s.

Eye activity measures demonstrated significant and predictable changes as a function of target density. Blink frequency, fixation frequency, and pupil diameter most characteristically defined individualised and general non-linear regression models relating eye activity to target density. A general regression model derived from all sessions produced estimates of target density that correlated well with actual workload levels (R = 0.55). Individual non-linear regression models based upon all sessions from every subject produced individual correlations ranging from 0.39 to 0.79. Next, individualised NN models were developed through training on several sessions and subsequently tested on a different session within each subject. Cross-session validation of NN-derived estimated to actual target density yielded correlations of 0.33 to 0.85 (mean R = 0.68).

The study demonstrates that information from multiple eye measures may be combined to produce reliable and near real-time estimates of cognitive and visual workload for some visuo-spatial tasks. The use of moving-mean estimates of pupil, blink and fixation measures with relatively brief integration times and individualised NN models represents a significant progression of eye-activity based psychophysiological assessment of workload.