Ocular Structures Predict Mechanical Function Properties Via Machine Learning

Li Longhui,Yifan Xiang,Xi Chen,Lanqin Zhao,Jun Xiao,Zhenzhe Lin,Jianyu Pang,Duoru Lin,Xiaotong Han,Lixue Liu,Yuxuan Wu,Zhenzhen Liu,Jingjing Chen,Jing Zhuang,Keming Yu,Haotian Lin

DOI: https://doi.org/10.2139/ssrn.4112118

2022-01-01

SSRN Electronic Journal

Abstract:Purpose: To develop and validate machine learning models for ocular mechanical function parameters prediction, which could be large-scale corneal biomechanical evaluation tools, based on structural parameters. Design: A retrospective, cross-sectional study. Method: A total of 1558 eyes of 1452 patients from different departments were enrolled at Zhongshan Ophthalmic Center. All patients were examined with an IOLMaster-700 and corneal visualization Scheimpflug technology tonometer (CorVis). To predict ocular mechanical function parameters, including the distance between the cornea and lens after corneal deformation (A-A), intraocular pressure (IOP), and deformation amplitude ratios between the corneal apex and paracentral part (DAr-1 and DAr-2), 13 features were extracted to train intelligent models and build linear formulas. We first established full models with all features, based on which, we further built simplified models (I and II). The model evaluation used the mean absolute errors (MAEs), root mean square errors (RMSEs), intraclass correlation coefficients (ICCs), and Bland-Altman plots. Results: The top 5 predictors for IOP, DAr-1, and DAr-2 predictions were age, central corneal thickness, keratometry, anterior chamber depth, and axial length. For A-A predictions, corneal diameter replaced keratometry as one of the top 5 predictors. The MAEs of the A-A, IOP, DAr-1, and DAr-2 predictions yielded by simplified model II were 0.0830 mm, 1.6992 mmHg, 0.0539, and 0.4933, respectively, under 10-fold cross-validation. Moreover, on the test dataset, the MAEs of simplified model II were 0.0875 mm, 1.5289 mmHg, 0.0571, and 0.5576, respectively. The prediction accuracy is clinically acceptable since the MAEs were lower than the measurement variations of CorVis. The linear regression model for A-A prediction performed well (R2 =0.940). Conclusions: Our intelligent models enable mechanical functional predictions with optical biometry, which is more cost-effective, patient-friendly, and repeatable than air-puff tonometry. This study showed that it is feasible to predict mechanical functions through structural parameters for the first time and could benefit preclinical keratoconus screening.