From CLMI.X to CLMIX-AI: A Machine Learning-Based Upgrade of the Cone Location and Magnitude Index Expanded to Detect...

To enhance the CLMI.X index for detecting keratoconus suspect (KCS) cases by incorporating advanced machine learning (ML) algorithms.

Development and validation of a ML diagnostic algorithm.

This study included 352 eyes divided into 3 categories: normal (NL) (n = 133), KCS (n = 77), and keratoconus (KC) (n = 142). Imaging was obtained using the Galilei dual Scheimpflug-Placido system. Eleven variables, identical to those used in the original CLMI.X, were employed to train and test multiple ML models, including logistic regression (LR) and more advanced algorithms, for classification of corneas into NL, KCS, and KC. SHapley Additive exPlanations (SHAP) analysis was performed to identify influential variables for differentiating KCS from NL eyes, both for the full KCS group and a subgroup of topographically and tomographically borderline cases.

The original CLMI.X demonstrated high sensitivity (100%) and specificity (99%) for detecting KC but low sensitivity (4%) for KCS. Training the CLMI.X-AI using LR on 3 classes improved KCS sensitivity to 68%. Incorporating advanced artificial intelligence algorithms further increased KCS sensitivity to 75%, with an overall accuracy of 89%. SHapley Additive exPlanations analysis identified the posterior axial, posterior instantaneous, and anterior axial curvature maps as the most influential variables, alongside zonal pachymetry and posterior elevation. For topographically and tomographically NL fellow eyes, SHAP analysis revealed significant influence for zonal pachymetry.

The CLMI.X-AI demonstrates substantially improved performance in detecting KCS compared with the original CLMI.X. Its ML enhanced decision-making and 3-class training make it a more clinically relevant and robust tool for early KC detection.

The author has no/the authors have no proprietary or commercial interest in any materials discussed in this article.