Identification of fixations and saccades in eye-tracking data using adaptive threshold-based method

A team of researchers including Charles Orioma and Shailendra Bhandari has developed a new adaptive method to solve a core problem in eye-tracking analysis: accurately distinguishing between fixations (when the eye is still) and saccades (rapid eye movements). Current industry-standard algorithms rely on fixed, ad-hoc thresholds for velocity or dispersion, which fail to account for variability between individuals and experimental tasks, potentially biasing cognitive research. The team's novel approach treats fixations and saccades as states in a Markovian model, with the core innovation being an algorithm that dynamically optimizes the classification threshold by minimizing the ratio of state transitions (K-ratio).

In rigorous testing against a multi-threshold benchmark, the research yielded critical, actionable findings. While a fixed velocity threshold achieved the highest baseline accuracy (90-93%) in clean data, its performance catastrophically degraded with added noise, falling below 20% accuracy. In contrast, their adaptive optimization method applied to a dispersion-based algorithm demonstrated remarkable noise robustness, maintaining accuracy above 81% even at extreme noise levels of 50 pixels. This parsimonious method provides a significant upgrade, though it involves a precision-recall trade-off that favors detecting fixations over saccades.

The paper provides a practical decision framework for researchers and developers. For high-quality data in controlled settings, optimized velocity thresholds remain top performers. However, for noisy, real-world applications—such as usability testing, cognitive load assessment in VR, or assistive technology—the adaptive dispersion method is clearly superior. This work moves the field from one-size-fits-all thresholds to intelligent, context-aware algorithms, enabling more reliable extraction of cognitive proxies from eye-tracking data across diverse and challenging environments.