Decision tree-based modelling for identification of potential interactions between type 2 diabetes risk factors: a decade follow-up in a Middle East prospective cohort study

OBJECTIVE The current study was undertaken for use of the decision tree (DT) method for development of different prediction models for incidence of type 2 diabetes (T2D) and for exploring interactions between predictor variables in those models. DESIGN Prospective cohort study. SETTING Tehran Lipid and Glucose Study (TLGS). METHODS A total of 6647 participants (43.4% men) aged >20 years, without T2D at baselines ((1999-2001) and (2002-2005)), were followed until 2012. 2 series of models (with and without 2-hour postchallenge plasma glucose (2h-PCPG)) were developed using 3 types of DT algorithms. The performances of the models were assessed using sensitivity, specificity, area under the ROC curve (AUC), geometric mean (G-Mean) and F-Measure. PRIMARY OUTCOME MEASURE T2D was primary outcome which defined if fasting plasma glucose (FPG) was ≥7 mmol/L or if the 2h-PCPG was ≥11.1 mmol/L or if the participant was taking antidiabetic medication. RESULTS During a median follow-up of 9.5 years, 729 new cases of T2D were identified. The Quick Unbiased Efficient Statistical Tree (QUEST) algorithm had the highest sensitivity and G-Mean among all the models for men and women. The models that included 2h-PCPG had sensitivity and G-Mean of (78% and 0.75%) and (78% and 0.78%) for men and women, respectively. Both models achieved good discrimination power with AUC above 0.78. FPG, 2h-PCPG, waist-to-height ratio (WHtR) and mean arterial blood pressure (MAP) were the most important factors to incidence of T2D in both genders. Among men, those with an FPG≤4.9 mmol/L and 2h-PCPG≤7.7 mmol/L had the lowest risk, and those with an FPG>5.3 mmol/L and 2h-PCPG>4.4 mmol/L had the highest risk for T2D incidence. In women, those with an FPG≤5.2 mmol/L and WHtR≤0.55 had the lowest risk, and those with an FPG>5.2 mmol/L and WHtR>0.56 had the highest risk for T2D incidence. CONCLUSIONS Our study emphasises the utility of DT for exploring interactions between predictor variables.