Predicting Early Keratoconus Progression Using Biomechanics via Multi-Machine Learning: A Multicenter 2-Year Prospective Cohort Study.

BACKGROUND: This study aimed to develop and validate a biomechanical early progression index (BEPI) based on machine learning algorithms, intended to predict early keratoconus progression at the first visit. METHODS: This multicenter prospective cohort study recruited 247 eyes of 247 participants from 3 centers; 225 completed all visits and were divided into a training and validation cohort. Corvis ST and Pentacam were used to obtain biomechanical and tomographic data. Based on disease progression during the follow-up period, predictive models were constructed using biomechanical data by applying Random Forest, XGBoost, Naïve Bayes classifier, K-Nearest Neighbours and Support Vector Machine. RESULTS: In the training cohort (n = 157), 66 eyes remained stable after a mean period of 25.18 ± 4.17 months, whereas 91 eyes demonstrated progression after 6.74 ± 2.53 months. The random forest provided the highest predictive performance and was designated as BEPI. With a cutoff value of 0.528, BEPI achieved 0.989 sensitivity and 0.985 specificity in distinguishing between stable and progressive cases. In the external validation cohort (n = 68), 31 eyes remained stable after 23.16 ± 6.75 months and 37 eyes progressed after 7.03 ± 3.42 months of follow-up. BEPI demonstrated the highest predictive performance, with an accuracy of 0.941 and an F1-score of 0.947. CONCLUSION: The machine learning-based index BEPI derived from corneal biomechanics enabled the accurate prediction of early keratoconus progression at the first visit, with an accuracy of 0.941. BEPI provides clinicians with a novel index for the accurate evaluation of early keratoconus progression, facilitating personalised clinical decision-making to preserve vision and improve prognosis.