Stability selection machine learning identifies biomechanical markers for subclinical idiopathic corneal endothelial...

The early detection of idiopathic corneal endothelial compromise, a common yet underdiagnosed pre-cataract finding, remains a challenge. This study aimed to investigate the association between corneal biomechanical parameters and endothelial status, identify potential biomechanical markers, and explore the biomechanical manifestations of idiopathic endothelial impairment.

In this cross-sectional study, cataract patients with normal endothelium or idiopathic endothelial abnormality underwent Corneal Visualization Scheimpflug Technology (Corvis ST) tonometry and specular microscopy. Beyond conventional statistics, we employed stability selection-a machine learning method with explicit error control-to identify the most reproducible predictors from a multitude of biomechanical parameters, while rigorously accounting for central corneal thickness (CCT) via stratified analysis and propensity score matching (PSM).

Among 241 patients (35 with endothelial abnormality), five biomechanical parameters (SSI, HCR, A2L, A1T, cTBI) significantly differed between groups, even after PSM for CCT. Stratified analysis revealed the strongest biomechanical-endothelial correlations specifically within the mid-range CCT (530-570 μm). A2L was identified as an independent protective factor (OR: 0.202, 95%CI: 0.051-0.731, p = 0.018). Stability selection confirmed A2L and cTBI as the most robust markers. A novel DASC model (incorporating A2L, DA, SSI, and CCT) achieved an AUC of 0.711 for detecting endothelial abnormalities, comparable to the cTBI score (AUC: 0.683, 95%CI: 0.565-0.800).

Corneal biomechanics reflect early endothelial alterations. A2L, cTBI, and the DASC model serve as promising, noninvasive tools for risk stratification. Our findings suggest that subclinical endothelial impairment manifests as reduced corneal stiffness-related response and impaired deformation behavior.