Modeling the risk of age-related macular degeneration and its predictive comparisons in a population in South India

Abstract

INTRODUCTIONAge-related Macular Degeneration (AMD) is one of the leading causes of irreversible vision loss among the elderly and is becoming a major public health problem in the developing world due to an increased size of the older population resulting from increased life expectancy. An estimated 25 to 30 million people around the world are believed to have been affected by AMD, and this number is expected to triple within 25 years. The etiology of AMD is not fully understood but is assumed to be of multi-factorial origin. Modeling the risks of AMD using the new approach is one way of finding the nonmodifiable and modifiable risk factors that suggests ways of prevention strategies for delaying the onset of AMD or early diagnosis of AMD, thereby preventing the further deterioration of the visual loss due to AMD.Many techniques have been used to derive clinical algorithms to infer the presence or severity of disease. The accuracy of a statistical method is finally reflected by its ability to predict outcome from training as well as test samples. Modeling a validated risk score based on a set of risk factors that cause AMD in the population is necessary for several reasons. The validated risk score yields a threshold score which can potentially identify ABSTRACT Background: Objective of current study was to develop and cross-validate the prediction models for Age-related Macular Degeneration (AMD) by using Logistic Regression (LR) and Artificial Neural Networks (ANN). Methods: A population based cross-sectional epidemiologic study. The data (n=3723) were analyzed on participants aged ≥40 years in Andhra Pradesh, South India. Sub-population data from this sample was drawn by using random under sampling and random over sampling techniques to derive a risk score from the LR model. The models were compared for their predictive abilities by an Area under the Receiver Operating Characteristic Curve (AUROC). Results: The LR risk score was built with a score ranging from 0 to 60 for a sub-population dataset (n=213). A cutoff score of ≥30 had a sensitivity of 79% and a specificity of 69%. The predictive performance of ANN and LR was statistically equivalent (76% vs. 78%; P = 0.624). Both the models were stable and consistently obtained the same predictive accuracies in a 30-fold split-sample cross validation. Conclusions:The sensitivity analysis of the ANN model indicated the relative importance of prioritizing modifiable risk factors for AMD in order to base preventive interventions to reduce the impact of the modifiable factors on AMD.