Abstract
Purpose:
Existing methods for high-myopia risk-assessment involve models with population-specific parameters and many hard-to-obtain variables. These issues cast doubts on the existing methods’ generalization ability and clinical practicality. We hypothesize that follow-up refraction error measurements during childhood allows accurate high-myopia risk-assessment at an individual level. This accordingly gives rise to a novel risk-assessment model with enhanced practicality and generalization ability.
Methods:
Subjects from of the Guangzhou Twins Eye Study (GTES, 2006-2013) and from a retrospective data collected from Optometry Clinic of Zhongshan Ophthalmic Center (OZOC, 2009-2014) were used for analysis. Refraction error was quantified as Spherical Equivalence (SE), defined as sphere plus half cylinder. Subjects with 2 follow-up measurements before age 12.5 year and endpoint SE measured after age 14.5 years were retained for analysis (84 and 84 subjects, respectively for each study). A Change-Point regression (CR) model was fitted using the follow-up SE, and the endpoint refraction error was used to validate the predicted endpoint SE. Standard Linear Mixed-Effects (LME) model was considered as a competing method. A generalization ability analysis was further performed, where the models fitted using the GTES data were directly used to predict the endpoint SE of the subjects in the OZOC data. The methods were compared based on the mean-square-error (MSE) between the predicted and the clinically measured endpoint SE. To determine whether the methods could well-predict the subjects at high-risk of high myopia, the MSE for the subjects with clinically measured endpoint SE < -5 Diopters (MSE5) was also computed.
Results:
For the OZOC data, CR and LME with two follow-ups both achieved MSE5=0.73 (CR MSE = 0.85, LME MSE=1.23). For the GTES data, CR with two follow-ups achieved MSE5=1.29 (MSE=1.32), while the 2nd best method was LME with one follow-up (MSE5=2.26, MSE=2.14). For the generalization analysis, CR achieved superior generalization accuracy over the LME method (with two follow-ups, CR: MSE5=0.89 & MSE=0.74; LME: MSE5=2.58 & MSE=2.10).
Conclusions:
The CR method, which merely requires follow-up SE measurements, enjoys enhanced prediction accuracy and generalization ability over the LME method. The results suggest that SE follow-ups are valuable measures for assessing the risk of high myopia.