Presentation
Unveiling the Mechanisms of Diagnostic Aid Use: Cognitive Modeling Reveals Suboptimal Strategies
DescriptionHuman operators engaged in high-stakes detection tasks (e.g., medical diagnosis) can be assisted by diagnostic aids to improve sensitivity relative to unaided levels. Nevertheless, aided sensitivity often falls short of best-achievable levels, sacrificing safety and productivity. To improve aid use, we need to understand operators’ automation interaction strategies. Here, we employ Bayesian modeling on data from a numeric signal-detection task where diagnostic aid assistance was either 77% or 93% reliable. We compare the fits of three models of aid-use: a contingent cutoff model (Robinson & Sorkin, 1985), an all-or-nothing discrete deference model (Bartlett & McCarley, 2017), and a mixture model of the two. Only a contingent cutoff strategy can lead to optimal aided performance in these contexts. However, consistent with recent literature, comparisons decisively favored the mixture model, which provided good empirical fits to data. Results offer guidance for interventions aimed at increasing the quality of aid use.
Event Type
Lecture
TimeWednesday, September 11th2:10pm - 2:30pm MST
LocationFLW Salon J
Human Performance Modeling