The independent test set comprised 60 eyes, of which 14 had forme fruste keratoconus. The area under the receiver operating characteristic curve of the XGBoost model was 0.932, with a 95% CI of 0.829 to 1.000, spanning 17 percentage points that encompassed the boundary between acceptable and outstanding discrimination. At this test set size, the misclassification of 2 additional cases would materially alter both the point estimate and clinical interpretation of sensitivity. The reported absence of statistically significant differences between classifiers similarly reflects limited statistical power rather than confirmed equivalence. These CI widths should temper comparative claims against previous OCT-based indices and multimodal diagnostic systems.

Forme fruste keratoconus was defined as the fellow eye of a patient with manifest keratoconus that met the specified topographic normality criteria. This pragmatic definition enriches the case group for eyes with higher underlying structural and genetic risks than those encountered in a general refractive surgery screening population. A model trained on this distribution will face a systematically different case mix during its deployment. The authors’ supplementary prevalence adjustment illustrates this directly: At a screening prevalence of 0.5%, the positive predictive value falls to 16.4%, meaning fewer than 1 in 6 positive classifications would represent the true disease. This figure is operationally the most relevant performance metric for the stated application of early keratoconus screening and deserves prominence in the main text, rather than the supplementary material. Presenting it alongside the headline accuracy figures would provide readers with a more complete picture of the tool’s real-world clinical utility.

The discriminative features were concentrated along meridians M1 to M3, corresponding to the inferotemporal corneal region, a distribution consistent with the established keratoconus topographic predilection. This spatial alignment provides a plausible biological basis. However, the specific texture features selected through recursive feature elimination remain unclear. Without correlation to histologic or biomechanical markers of early stromal collagen disorganization, it cannot be determined whether these features capture genuine microstructural pathology or imaging acquisition patterns that cosegregate with the forme fruste keratoconus label in this single-center, single-device data set.

Luo and associates have produced a technically sound proof of concept with genuine translational promise. Prospective external validation across multiple OCT platforms, clinical sites, and refractive surgery populations with prespecified prevalence assumptions is the necessary next step before this framework can be used to inform screening decisions.