Fig. 3

SHAP analysis results. (a) SHAP feature ranking and (b) SHAP summary plot. The top-ranked features are RNFL symmetry, RNFL inferior, RNFL clock-11 (part of the superior quadrant) and RNFL-superior thickness. The blue colour represents lower magnitudes of feature values while the red colour represents higher values. positive SHAP values on the right side push the model to predict a win (in this case the probability of being classified as glaucoma), while negative SHAP values on the left push the model to predict a loss (probability of being classified as normal). (c) SHAP dependence plot (RNFL symmetry). Positive SHAP values on the vertical axis mean that the corresponding feature values have a positive impact on the prediction, while negative SHAP values mean that the feature values have a negative impact. Points close to the zero line can be useful for identifying instances where a feature did not contribute to the model’s prediction. (d) SHAP dependence plot with interaction visualisation (interaction between RNFL inferior and RNFL symmetry). For (c), the data points in red above the horizontal dashed line represent the samples classified as glaucoma, while those in blue below the line represents the samples classified as normal. In (d), the red points corresponding to higher RNFL inferior thickness indicate the presence of higher RNFL symmetry, which generates negative SHAP values and shows a lower chance of being classified as glaucoma. RNFL clock hours-2,6,7,11 represents RNFL thickness measured in clock-hour sectors. Each sector is measured in degrees, usually in increments of 30 degrees, creating a 360-degree circle around the optic nerve head.