The comparison of the predictive precision of artificial neural networks and bivariate logistic regression in diagnosis of patients with hypertension

Diabetes and hypertension are from important non-communicable diseases in the world and their prevalence are very important for health authorities. The objective of this study was to compare the predictive precision of joint logistic regression (LR) and artificial neutral network (ANN) in concurrent diagnosis of diabetes and hypertension. This cross-sectional study was performed on 12000 Iranian people in 2013. The study questionnaire included some items on hypertension and diabetes and their risk factors. A perceptron ANN with two hidden layers was applied to data. The variables in the study were diabetes, hypertension, gender, type of cooking oil, physical activity, family history, age, obesity and passive smokers. To build a joint LR model, and ANN, SAS 9.2 and Matlab software were used. The ROC curve was used to find the higher accuracy model for predicting diabetes and hypertension. The variables of gender, type of cooking oil, physical activity, family history, age, passive smokers and obesity entered to the LR model and ANN. The odds ratio of affliction to both diabetes and hypertension is high in females, user of solid oil, people with no physical activity, with positive family history, age of equal to or higher than 55, passive smokers and obesity. The area under ROC curve for LR model and ANN were 0.78 (p=0.039) and 0.86 (p=0.046) respectively. The best model for concurrent affliction to hypertension and diabetes is ANN which has higher accuracy than the joint LR model.