Abstract Purpose To characterize structural heterogeneity within suspect keratoconus (SKC) using an unsupervised phenotyping strategy integrating posterior tomographic and epithelial remodeling parameters. Methods In this cross-sectional study, 55 eyes with clinically defined SKC underwent Scheimpflug–Placido tomography and anterior segment optical coherence tomography. Posterior asymmetry indices (Symmetry Index Back SIb and Keratoconus Vertex Back KVb) and epithelial redistribution metrics (Minimum–Maximum epithelial thickness difference Min–Max ET and Superonasal–Inferotemporal epithelial thickness difference SN–IT ET, 2–5 mm zone) were entered into k-means clustering without predefined diagnostic thresholds. Demographic and refractive variables were compared between derived phenotypes. Independent anterior surface elevation and keratometric parameters were analyzed for structural validation. Cluster robustness and multidimensional separation within the SKC cohort were assessed using silhouette analysis, principal component analysis (PCA), posterior-only sensitivity analysis, and hierarchical clustering sensitivity analysis. Results Two distinct structural phenotypes emerged. A posterior-asymmetric phenotype (26 eyes) demonstrated significantly higher SIb and KVb (both p 0.05). Simulated keratometric indices (K1, K2, Kavg) did not distinguish between phenotypes, whereas anterior elevation parameters differed significantly. PCA loading analysis and posterior-only sensitivity analysis supported posterior tomographic abnormalities as the principal axis of phenotype separation. The two-cluster solution showed a silhouette coefficient of 0.29, and hierarchical clustering sensitivity analysis supported its robustness (agreement = 87.5%, Cohen’s κ = 0.729). Conclusions Suspect keratoconus is not a structurally uniform entity but comprises distinct structural phenotypes driven principally by posterior tomographic asymmetry and accompanied by epithelial remodeling. Epithelial remodeling provided complementary structural information rather than serving as the dominant driver of cluster separation. Because longitudinal and biomechanical data were unavailable, the future clinical significance and progression risk of these phenotypes remain unknown and require further investigation.
Abdelrahman Salman (Sat,) studied this question.