Assessing Workload in Human-Machine Teams from Psychophysiological Data with Sparse Ground Truth

Data-driven approaches to human workload assessment generally attempt to induce models from a collection of available data and a corresponding ground truth comprising self-reported measures of actual workload. However, it is often not feasible to elicit self-assessed workload ratings with great frequency. As part of an ongoing effort to improve the effectiveness of human-machine teams through real-time human workload monitoring, we explore the utility of transfer learning in situations where there is sparse subject-specific ground truth from which to develop accurate predictive models of workload. Our approach induces a workload model from the psychophysiological data collected from subjects operating a remotely piloted aircraft simulation program. Psychophysiological measures were collected from wearable sensors, and workload was self-assessed using the NASA Task Load Index. Our results provide evidence that models learned from psychophysiological data collected from other subjects outperform models trained on a limited amount of data for a given subject.