Performance Comparison of Machine Learning Algorithms for Keratoconus Detection

This paper provides an in-depth examination of eight Machine Learning (ML) algorithms commonly employed in diagnosing keratoconus (KC) over the past decade. The analysis is conducted using a clinical dataset accumulated by the authors from two prominent eye-care centers in Jordan over the same period. By scrutinizing the data collected from these centers, the proposed study aims to shed light on the efficacy and comparative performance of various ML algorithms in KC diagnosis. Through rigorous evaluation and comparison, we seek to identify the algorithms that exhibit superior predictive accuracy and reliability in differentiating between normal, subclinical (or suspect) KC, and clinical KC corneas. The obtained results demonstrated that the tree-based models achieved superior performance compared to other models. Specifically, the Random Forest attained the highest classification accuracy at 99.6%, closely followed by Gradient Boosting at 99.2%, and Decision Tree at 98.4%. These promising findings underscore the effectiveness of these classifying algorithms in KC detection and their promise for practical implementation in clinical settings.