Eye-tracking biomarkers of clinical expertise in ECG interpretation: statistical and machine learning evidence

Abstract

Introduction: Interpreting the electrocardiogram (ECG) is a fundamental clinical skill, and mistakes are still prevalent in the workforce, especially among trainees and non-specialist clinicians. Eye-tracking technology has recently become a popular method for investigating visual expertise. However, few studies have integrated visual behavior metrics with machine learning to accurately classify expertise levels. Methods: The original dataset included 62 participants from 10 healthcare roles (students, nurses, technicians, residents, fellows, consultants) who interpreted standardized ECGs. Eye movements were recorded using a Tobii Pro X2-60 tracker. ECGs were segmented into grid-based and functional Areas of Interest (AOIs). Certain eye-tracking metrics, such as Fixation Count, Time to First Fixation (TTFF), Gaze Duration, and Revisit Count, were evaluated via statistical analyses (ANOVA, Kruskal–Wallis, t-tests). Gaze features were used to train machine learning models (Random Forest, Support Vector Machine, K-Nearest Neighbors), and clustering was performed with K-means. Results: Experts demonstrated faster TTFF, fewer revisits, and shorter fixation durations compared to novices. Experts exhibited more efficient gaze behavior, with fewer fixations within each diagnostic AOI but a higher overall fixation count per ECG due to broader systematic scanning. The correlation between fixation count and gaze duration was high (R2 = 0.76). Random Forest achieved the best classification accuracy (84%), outperforming SVM (78%) and KNN (74%). A Random Forest classifier achieved an accuracy of 84% using five-fold cross-validation, and performance significantly exceeded chance based on a 1,000-permutation test (p < 0.001), demonstrating robust discriminative ability. These findings indicate that gaze-based features can reliably differentiate expertise levels. The groups identified by K-means clustering corresponded (for the most part) to novice, intermediate, and expert. Feature importance showed that leads V1, V2, and the rhythm strip were the top predictors of expertise. Conclusion: Eye-tracking parameters differentiated levels of ECG interpretation expertise. These results suggest that gaze-derived metrics may serve as potential surrogate indicators that support assessment and training in medical education.