Real-time prediction of short-timescale fluctuations in cognitive workload
DOI: 10.1186/s41235-021-00289-y
archive: archived pipeline: cataloged
Summary
Boehm et al. (Cognitive Research: Principles and Implications, 2021) test whether short-timescale (several-second) fluctuations in cognitive workload can be predicted from primary-task state and performance, with the goal of supporting adaptive automation that does not require ongoing intrusive workload measurement. Forty-six participants completed a simulated UAV fleet refuelling task across two one-hour sessions on different days (24 two-minute blocks per session) while the ISO 17488 Detection Response Task (vibrotactile stimulus every 3-5 s, foot-pedal response) provided a fine-grained workload criterion. The authors fit linear mixed two-part models of DRT response time and omissions jointly, plus separate RT and omission models for cross-validated prediction. The DRT was sensitive to manipulations of UAV fleet size (validating it as a workload measure), and predictive models built on operator situational-awareness composites (e.g., fuel-level monitoring) substantially outperformed models based on the simple occurrence of preceding task events.
Key finding
Cognitive workload, measured by the ISO 17488 DRT, varies rapidly (over seconds) as a function of recent primary-task events, and composite predictors capturing operators' situational awareness about fuel levels predict DRT RT and omissions in held-out data far better than predictors based only on immediately preceding task-event occurrences, supporting real-time predictive workload models for adaptive automation.
Methodology
Within-subjects experiment. N=46 adults from the University of Utah participant pool, two one-hour sessions on separate days, each consisting of 24 two-minute blocks of a simulated UAV fleet refuelling task on Windows 7 desktop systems running at 35 fps. The ISO 17488 Detection Response Task was active throughout, using a shoulder-mounted vibrotactile motor (3-5 s ISI, 1 s stimulus) and a foot-pedal response on millisecond-accurate ARM-based DRT hardware. Analyses combined linear mixed two-part models of DRT RT and omissions with separate RT and omission models, using cross-validation to evaluate predictors derived from preceding task events, near-future task difficulty, and operator situational awareness with respect to UAV fuel levels.
Sample size: N=46 adults (University of Utah participant pool); two one-hour sessions per participant on different days.
Quality score: 5 / 5