Gaze Revisits as Cognitive Signals for Human-Centered AI

Eye movements reveal how visual information is selected over time, but it is unclear whether returning to previously inspected information marks continued cognitive processing a distinction that current AI systems mostly ignore, since gaze is typically reduced to a static heat map that discards its temporal structure. Gaze revisits to abnormality regions were tested for association with diagnostic certainty during chest-radiograph interpretation, with the aim of identifying a process-level behavioral signal that could inform more human-centered AI. The analysis used ۴,۰۶۸ radiologist-drawn abnormality ellipses from ۱,۷۰۴ phase-۳ readings in the REFLACX dataset. A revisit was defined as a return to an abnormality ellipse after gaze had left it; consecutive fixations inside the ellipse formed one visit. Revisit count was analyzed with a negative-binomial mixed model adjusting for relative ellipse area, reading duration, number of abnormalities, abnormality type, and clustering within a reading. Certainty improved model fit, and expected revisit counts were higher for certainty levels ۲ through ۴ relative to the highest-certainty reference category. These results indicate that gaze revisits are a behavioral marker of repeated evidence sampling and unresolved processing, and point toward AI systems that learn from the temporal structure of expert gaze rather than a single spatial summary for instance, models trained to anticipate when a region will draw renewed inspection, or decision-support interfaces that time their explanations to moments of continued human uncertainty rather than presenting them uniformly.

کلیدواژه ها:

Eye tracking ، Gaze revisits ، Chest radiographs ، Human-centered AI ، Cognitive processing