High-Accuracy Situation Awareness Prediction From Cognitive State Measures Relies on Individual Training Data

Abstract

Autonomous vehicles promise to improve the safety of driving, but modern implementations do so at the expense of situation awareness (SA) in human drivers—drivers who remain critical for safe operations. To enable autonomous systems capable of monitoring and maintaining human operator/driver SA, this work generates predictive models of SA from embedded measures taken from in-task performance, post-trial measures from cognitive assessments, and pre-experiment measures of operator background information and cognitive performance. Models using pre-experiment measures outperformed models using both post-trial and embedded measures, predicting measures of SA with high accuracy (Q 2 scores up to 0.61 and standardized mean absolute errors as low as 0.47), and provide interpretable models of underlying factors affecting SA. Models suffer from low predictive performance on unseen participants, and analysis of the dataset suggests that individual differences may drive performance in SA prediction. By outperforming accuracy of prior work in predictive models of SA with interpretable models, we highlight the critical role of underlying cognitive states—and other factors affecting cognition—in predictive models of SA. However, future work remains to isolate temporal dynamics of SA from other sources of variance.