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

Figures

• 

  Figure 1. AOI layouts for ECG interpretation. Left: Grid-based AOI layout showing 24 equal spatial zones over a typical 12-lead ECG image, designed for fine-grained visual attention analysis. Each zone represents a unique rectangular region encompassing one or more ECG leads. Right: Long vs. short lead AOI layout separating the diagnostic leads (12 short leads) from the rhythm strip (long lead). The shaded green area represents the short leads (diagnostic AOI), while the red region highlights the rhythm strip, typically used to detect arrhythmias and timing abnormalities. This dual-AOI design supports both spatial and functional analyses of gaze behavior.

• 

  Figure 2. Cross-validation classification performance across models. Mean accuracy (± SD) for the random forest, SVM, and KNN classifiers using five-fold stratified cross-validation. The random forest model showed the highest performance (84% ± 3%), indicating strong internal generalizability.

• 

  Figure 3. Fixation count and mean fixation duration as a function of expertise. Mean fixation count and mean fixation duration for novices (n = 21), intermediates (n = 22), and experts (n = 19) in grid-based AOIs during ECG reading. Experts exhibited higher total fixation counts per ECG, although they used fewer fixations within specific diagnostic AOIs, reflecting efficient but more comprehensive visual scanning.

• 

  Figure 4. Fixation count, total gaze duration, and revisit count for expertise group. The mean fixation count, gaze duration, and revisit count for the novice (n = 21), intermediate (n = 22), and expert (n = 19) categories. One-way ANOVA for fixation count [F(2, 59) = 12.9, p < 0.001, η2 = 0.30], gaze duration [F(2, 59) = 10.3, p < 0.001, ε2 = 0.26], and revisit count [F(2, 59) = 8.4, p = 0.001, ε2 = 0.22] revealed significant differences among groups. Post hoc Tukey revealed that the difference between novices and experts was significant. Error bars indicate ±1 SD. AOI revisits:

• 

  Figure 5. Mean AOI revisit count across expertise levels. Mean number of AOI revisits per ECG for novice, intermediate, and expert groups. A Kruskal–Wallis test indicated significant differences (H = 36.26, p < 0.001), with experts showing the fewest revisits. Error bars represent ±1 SD. Time to first fixation (TTFF): Experts located diagnostic regions faster (Median = 812 ms, IQR 675–930) compared to novices (Median = 1,692 ms, IQR 1470–1830; H = 36.26, p < 0.001 (Figure 6).

• 

  Figure 6. Time to first fixation (TTFF) across expertise levels. Median TTFF on diagnostic AOIs for novice, intermediate, and expert participants. A Kruskal–Wallis test showed statistically significant differences between groups (H = 43.15, p < 0.001), with experts orienting to key diagnostic regions fastest. Error bars reflect interquartile ranges.

• 

  Figure 7. Linear relationship between fixation count and gaze duration. Scatterplot showing the correlation between fixation count and total gaze duration across all participants. Linear regression analysis yielded a strong positive relationship (R2 = 0.76, p < 0.001), indicating that increased visual scanning is associated with longer attention on ECGs. This figure illustrates the performance of an earlier model configuration with class imbalance.

• 

  Figure 8. Model comparison for expertise classification based on gaze-derived features. Supervised models classification performance [random forest, support vector machine (SVM), K-nearest neighbors (KNN)] accuracy. Random forest outperformed with 84% accuracy, followed by SVM (78%), and KNN (74%).

• 

  Figure 9. Confusion matrix for the three-class random forest expertise classifier. Confusion matrix showing classification performance across novice, intermediate, and expert categories. The model correctly classified intermediate participants most frequently, with some overlap observed between adjacent expertise levels.

• 

  Figure 10. K-means clustering of gaze-based features. K-means clustering results of fixation count, gaze duration, TTFF, and revisits. Three clusters were identified that closely mirrored novice, intermediate, and expert participants. Silhouette score = 0.64.

• 

  Figure 11. Feature importance of ECG AOIs in random forest classification. Relative importance of gaze features for each ECG AOI derived from the random forest model. Leads V1, V2, and the rhythm strip showed the highest importance in distinguishing expertise levels. Feature importance was calculated using mean decrease in impurity.

• 

  Figure 12. Permutation test null distribution and observed accuracy. Null distribution of classification accuracy generated from 1,000 label permutations compared with the observed random forest accuracy (84%). The observed performance exceeded all permutations (p < 0.001), confirming that classification accuracy was significantly above chance.