Machine learning model for predicting corneal stiffness and identifying keratoconus based on ocular structures
Longhui Li, Yifan Xiang, Xi Chen, Duoru Lin, Lanqin Zhao, Jun Xiao, Zhenzhe Lin, Jianyu Pang, Xiaotong Han, Lixue Liu, Yuxuan Wu, Zhenzhen Liu, Jingjing Chen, Jing Zhuang, Keming Yu, Haotian Lin
Abstract
Corneal stiffness abnormalities play an important role in the onset and progression of keratoconus. However, the limited availability of specialty devices for measuring corneal stiffness restricts their application in clinical practice. This study aimed to develop a machine learning (ML) model that can predict corneal stiffness based on ocular structures and investigate its efficacy in diagnosing keratoconus. This retrospective study enrolled healthy individuals and keratoconus patients at the Zhongshan Ophthalmic Center from June 2018 to June 2021. Eleven features, including ocular structural parameters, intraocular pressure (IOP), and age were used to train ML regression models for predicting the stiffness parameter at first applanation (SP-A1) and the Corvis biomechanical index for Chinese populations (cCBI) measured by a Corvis ST device. Mean absolute errors (MAEs) and the area under the receiver operating characteristic curve (AUC) were used to evaluate the performance of the models. The diagnostic efficacy of the predicted SP-A1 and cCBI for keratoconus was evaluated by the AUC, net reclassification index (NRI), and integrated discrimination improvement (IDI). A total of 1,523 eyes were involved, of which 601 were diagnosed with keratoconus. The MAEs of the SP-A1 prediction were similar in cross-validation (8.95 mmHg/mm) and testing (10.65 mmHg/mm). The R 2 value for the SP-A1 prediction exceeded 0.7, indicating that the performance was clinically acceptable. The AUC for the cCBI prediction was 0.935 (95% CI 0.906-0.963). The top three predictors for SP-A1 and cCBI were IOP, keratometry, and central corneal thickness. The addition of the predicted SP-A1 and cCBI significantly improved model performance in diagnosing keratoconus, with NRI of 0.607 (95% CI 0.367-0.812) and 0.188 (95% CI -0.022- 0.398), and IDI of 0.028 (95% CI 0.006-0.048) and 0.045 (95% CI 0.018-0.072), respectively. Our models predicted SP-A1 and cCBI relatively accurately in both keratoconus and normal corneas. Moreover, the predicted SP-A1 and cCBI values significantly contributed to the diagnosis of keratoconus. These models could provide a potential alternative for evaluating corneal stiffness and thus facilitate keratoconus screening.