Event-Driven Predictive Gaze Analytics to Enhance Aviation Safety

Future intelligent aircraft are envisioned to predict human failures and mitigate threats at runtime to optimize safety. Contributing towards realizing this vision, and more specifically, recognizing when timely interventions should be invoked, this paper investigates gaze-based predictions on pilots' success and failure. In a simulated study with 50 licensed pilots performing an Instrument Landing System approach in low-visibility conditions, we examined the use of saccade magnitude and direction, pupil dilation, fixation duration, and blink rate to detect notable events experienced by pilots. We performed classification experiments using established machine learning models demonstrating that pilots' performance can be successfully predicted as early as 3.42 minutes after task initiation with accuracies up to 80.92% depending on the gaze measure used. Our findings also showed improved accuracies compared to predictions generated using established gaze measures such as gaze entropy. Furthermore, accuracies tend to peak during critical flight phases and support vector machines performed well across the tested gaze measures.