Objective: To provide a noninvasive type 2 diabetes mellitus (T2DM) prediction model for a rural Chinese population.
Methods: From July to August, 2007 and July to August, 2008, a total of 20 194 participants aged ≥18 years were selected by cluster sampling technique from a rural population in two townships of Henan province, China. Data were collected by questionnaire interview, anthropometric measurement, and fasting plasma glucose and lipid profile examination. A total 17 265 participants were followed up from July to August, 2013 and July to October, 2014. Finally, 12 285 participants were selected for analysis. Data for these participants were randomly divided into a derivation group (derivation dataset, n= 6 143) and validation group (validation dataset, n=6 142) by 1∶1, respectively. Randomization was carried out by the use of computer-generated random numbers. A Cox regression model was used to analyze risk factors of T2DM in the derivation dataset. A T2DM prediction model was established by multiplying β by 10 for each significant variable. After the total score was calculated by the model, analysis of the receiver operating characteristic (ROC) curve was performed. The area under the ROC curve (AUC) was used for evaluating model predictability. Furthermore, the model's predictability was validated in the validation dataset and compared with the Finnish Diabetes Risk Score (FINDRISC) model.
Results: A total 779 of 12 285 participants developed T2DM during the 6-year study period. The incidence rate was 6.12% in the derivation dataset (n=376) and 6.56% in the validation dataset (n=403). The difference was not statistically significant (χ(2)=1.00, P=0.316). A total of four noninvasive T2DM prediction models were established using the Cox regression model. The ROCs of the risk score calculated by the prediction models indicated that the AUCs of these models were similar (0.67-0.70). The AUC and Youden index of model 4 was the highest. The optimal cut-off value, sensitivity, specificity, and Youden index were scores of 25, 65.96%, 66.47%, and 0.32, respectively. Age, sleep time, BMI, waist circumference, and hypertension were selected as predictive variables. Using age<30 years as reference, β values were 1.07, 1.58, and 1.67 and assigned scores were 11, 16, and 17 for age groups 30-44, 45-59, and ≥60 years, respectively. Using sleep time<8.0 h/d as reference, the β value and assigned score were 0.27 and 3, respectively, for sleep time ≥10.0 h/d. Using BMI 18.5-23.9 kg/m(2) as reference, β values were 0.53 and 1.00 and assigned scores 5 and 10, respectively, for BMI 24.0-27.9 kg/m(2), and ≥28.0 kg/m(2). Using waist circumference <85 cm for males/< 80 cm for females as reference, β values were 0.44 and 0.65 and assigned scores 4 and 7, respectively, for 85 cm ≤ waist circumference <90 cm for males/80 cm≤ waist circumference <85 cm for females, and waist circumference ≥90 cm for males/≥85 cm for females. Using nonhypertension as reference, the respective β value and assigned score were 0.34 and 3 for hypertension. The AUC performance of this model and the FINDRISC model was 0.66 and 0.64 (P=0.135), respectively, in the validation dataset.
Conclusion: Based on this cohort study, a noninvasive prediction model that included age, sleep time, BMI, waist circumference, and hypertension was established, which is equivalent to the FINDRISC model and applicable to a rural Chinese population.