Download files
Abstract
The objective of this research was to determine whether an artificial intelligence methodology such as artificial neural network (ANN), a new type of predictive model offers an increased performance over a conventional logistic regression model (LR) in predicting the ranking of risk factors for irreversible age-related chronic eye diseases age-related macular degeneration (AMD), diabetic retinopathy (DR), primary open-angle glaucoma (POAG) and primary angle-closure glaucoma (PACG) in a South Indian population. The LR and ANN models were derived and validated for their respective models predictive accuracy based on a sample (n=3,723) aged >=40 years old by using a large scale population-based epidemiologic study. Sub-population data were drawn from this sample by appropriate standard techniques that used for modeling. The LR based risk score models (RS) were derived and the model fit was assessed in a standard manner including the bootstrap method for internal validity. The ANN model was built by using the multi-layer feed-forward back propagation network. The ANN models predictive ability was compared with that of traditional model with respect to the Area under the Receiver Operating Characteristic Curve (AUROC). The sensitivity and specificity of the fitted models with a threshold criterion ranged from 70% to nearly 99% overall for all models. The ANN model outperformed the traditional LR model in a sub-population analysis in predicting AMD and DR. The predictive accuracy of ANN and LR model in predicting AMD was statistically significant (AUROC=89% vs 79%; p<0.0001). Both the models revealed that the modifiable risk factors such as heavy smoking (RS ranged from 10 to 18), lower intake of antioxidants (RS ranged from 5 to 10), hypertension (RS ranged from 2 to 10) were in order of priority predictors for AMD and longer duration of diabetes >=10 year (RS ranged from 29 to 42) was a highest priority predictor for DR. The modifiable risk factor intraocular pressure was in order of highest priority predictor for POAG and PACG. Population attributable risk percentage and population attributable fractions revealed that there is an urgent need of prioritizing modifying the modifiable factors as a public health approach. This was supported by a sensitivity analysis of the ANN model which indicated the relative importance of prioritizing modifiable risk factors on which to base preventive interventions to reduce the impact of onset or progression of these diseases.