Comparison of NN and LR classifiers in the context of screening native American elders with diabetes

Classification is a frequently used decision making tool, however there are many classification methods and these seldom provide adequate and consistent results. In this paper we compare the classification efficiency of neural networks (NN) to more traditional methods such as LR (LR), in the context of identifying American Indian/Alaskan Native (AI/AN) elders who are at risk of developing diabetes. Feature selection is an important first step in building these classification models. We used both stepwise selection and genetic algorithm (GA) to identify features related to diabetes. Each LR and NN models were built twice, once based features identified by stepwise regression and second using features identified using genetic algorithm. Analysis of results from this approach lead to several conclusions: (a) although both LR and NN models exhibit similar classification ability, NN models were marginally better compared to LR models. (b) While the ROC value of these two models were the same (ROC = 1), the type of feature selection methodology had no impact on the sensitivity and specificity of these models. In conclusion results from our study shows that although both these models are equally capable of identifying AI/AN elders at risk of developing diabetes, NN models are marginally better.