DEVELOPMENT OF AN INTERPRETABLE MACHINE LEARNING MODEL FOR EARLY DETECTION OF SUBCLINICAL KERATOCONUS

To develop an innovative predictive model based on structural and topographic corneal parameters for the early identification of subclinical keratoconus.

Clinica Oculistica, IRCCS Ospedale Policlinico San Martino and University of Genoa, Genoa, Italy.

A total of 200 eyes (100 keratoconic, 100 healthy) were retrospectively analyzed. Data from anterior segment optical coherence tomography and corneal topography were processed using an automated machine learning platform (AutoML). Model performance was evaluated on an independent test set using the area under the Receiver Operating Characteristic (ROC) curve. Variable interpretability was assessed through explainable artificial intelligence (SHAP – SHapley Additive exPlanations) analysis.

The model achieved an area under the ROC curve of 0.997 and an overall accuracy of 98%. Central and peripheral pachymetric parameters, epithelial metrics (standard deviation and minimum thickness), and the corneal asymmetry index were significantly discriminative (p < 0.001). SHAP analysis confirmed their relevance, demonstrating a clear separation between keratoconic and control eyes.

The integration of structural and topographic parameters within an interpretable AutoML framework enables highly accurate early detection of subclinical keratoconus. The model’s strong performance and consistency of key predictors support its potential clinical application in screening and follow-up of at-risk patients.