Résumé
Detection of population at risk of type II diabetes, as a multi-factorial disease, is an important issue because of its individual and social impacts. To date, several studies have been conducted to predict the incidence of diabetes, using different statistical methods. However, despite its clinical importance, it is highly difficult to consider all interactions among risk factors, in ordinary statistical models. This study aimed to extract appropriate logic combination of type 2 diabetes risk factors employing the recently introduced method, Logic regression. The study population was selected from a cohort of the Tehran Lipid and Glucose Study [TLGS]. Data for 3523 participants, aged 20 years and over [57.8% female and 42.2% male] were entered into analysis, for which logistic logic regression method was used. The model parameters were estimated using the Annealing algorithm. To avoid overestimation, the optimal number of logic combinations was determined by the cross-validation method. Deviance, sensitivity and specificity measures were computed to evaluate the logic model and its comparison to ordinary logistic regression; the latter accommodated only the main effects. The prediction power of the two models was compared by Area under ROC curve. R software version 2.8.1 was employed for analyses. Logistic logic regression with the 4 Boolean combination including 5 variables was fitted using the Annealing algorithm and resulted in in deviance of 1203.30. This model had better fit compared to other logic models and also ordinary logistic regression with forward procedure [deviance=1206.88]. The Boolean combination of the above model included impaired fasting glucose [OR=5.53, 95%CI: 4.03-7.59], IGT [OR=5.54, 95%CI: 3.96-7.49], family history of diabetes [OR=1.89, 95%CI: 1.38-2.63], and interaction of high triglycerides or abnormal waist circumference [OR=2.4, 95%CI: 1.73-3.32]; all p-values <0.001. The area under ROC curve for the model was 0.843 [95%CI: 0.813-0.874]. This study showed that the logic regression as a newly introduced method has the ability of recognizing and modelling the interactions between different risk factors. Therefore, it is recommended as an appropriate tool for screening of the multi-factorial diseases such as diabetes