Predicting Keratoconus Progression From a Single Visit: Is Machine Learning Successful?

PURPOSE: To develop and evaluate a machine learning (ML)-based model for predicting keratoconus (KCN) progression in an Iranian cohort.

METHODS

This retrospective study included 1,000 eyes of 529 patients with KCN (mean age: 31.1±8.0 years; 63.3% male) with two ocular examinations at least six months apart (mean interval: 71.1±41.7 months) and no prior corneal surgery. Progression was defined by a composite criterion: ≥1.00 D increase in KmaxF or anterior astigmatism, ≥25 μm corneal thinning, or ≥0.42 increase in Belin/Ambrosio D-index. Three XGBoost-based algorithms were developed: (1) using baseline data plus interexamination change rates, (2) using only baseline data for three-class prediction (progressive/stable/regressive), and (3) a refined binary model (progressive vs. nonprogressive).

RESULTS

Of the 1,000 eyes, 32.3% were progressive, 59.4% stable, and 8.3% regressive. The type 1 algorithm achieved near-perfect performance (AUC=0.999, accuracy=99%). However, the single-visit type 2 model showed limited accuracy (65%) and low sensitivity for progression (46%). The optimized type 3 binary model improved sensitivity to 69.1% and AUC to 0.72. Feature importance analysis identified combination of anterior maximum curvature more than 48.0 D and thinnest pachymetry less than 470 μm is the most identifier parameter. The Clinical Risk Score enabled stratification into low, moderate, and high-risk groups for progression.

CONCLUSION

While ML models excel when longitudinal data are available, predicting KCN progression from a single visit remains challenging. Integrating engineered features and a clinical risk score enhances performance, but current accuracy is insufficient for standalone clinical use. Prospective validation and population-specific thresholds are needed before real-world implementation.