Development of a prediction model for advanced age-related macular degeneration using a penalized machine learning approach: the EYE-RISK project

Abstract

Purpose : Prediction models for advanced age-related macular degeneration (AMD) have been proposed in the past, but they were based on a limited set of risk factors. The objective of this study was to develop a more comprehensive prediction model, using a penalized machine learning approach for the selection of the most relevant predictors from a wider set of risk factors.

Methods : The Eye-Risk database is a unique harmonized database of individual data from 16 European population-based studies, including two prospective studies (Rotterdam and Alienor studies) with comprehensive data on genetic, phenotypic and lifestyle risk factors for AMD. The Rotterdam study (training set) included 3843 individuals with a 20 years follow-up and 108 incident cases of advanced AMD. The Alienor study (test set) included 362 individuals with a 5 years follow-up and 33 incident cases of advanced AMD. The prediction model was developed by applying a bootstrap lasso (Bolasso) method in a survival framework to select the most predictive factors of incident advanced AMD in the training set. Predictors were retained in the prediction model according to the elbow criterion. Prediction performance was assessed by internal cross-validation in the training set and by external validation in the test set. Both discrimination and calibration performances were evaluated by the area under receiver operating characteristic curves (AUCs) and the goodness of fit calibration curves respectively.

Results : The prediction model included age, a combination of phenotypic predictors (based on the presence of intermediate drusen, hyper-pigmentation in one or both eyes and age-related eye disease study (AREDS) simplified score), a summary genetic risk score based on 41 SNPs, smoking, diet quality (evaluated by a Mediterranean diet score), education, multivitamin supplementation and pulse pressure. Split sample validation in the training set showed an AUC of 0.95 at 5 years, 0.93 at 10 years and 0.91 at 15 years. In the independent validation set, the AUC reached 0.92 at 5 years.

Conclusions : Our prediction model included a broader set of predictors and reached higher discrimination abilities than previously published prediction models. The proposed model may be very useful to identify subjects at high risk for advanced AMD.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.