In this community-based cohort study, we developed a quantifiable and simple nomogram to predict the 3-year risk of T2DM in Central Chinese residents. After an internal validation, high degrees of predictive accuracy were found in both training and validation cohorts. Decision curve analysis also demonstrated the clinical value of this nomogram. We also estimated the optimal cutoff values of each risk factor for T2DM prediction. To the best of our knowledge, this study is the first to develop a nomogram by using continuous values instead of segmented values in estimating the T2DM risk in China. Moreover, the nomogram will also be of considerable practical value for its readily obtained parameters.
Different diabetes prediction models on the basis of demographic information and clinical measurements have been developed in European, North American [9, 10], and Asian populations [15, 16]. Given the genetic and environmental differences (i.e., economic level, diet, lifestyle, climate), risk factors for T2DM vary in terms of intensity or distribution across different populations, which suggested that a predictive model may not perform well in other ethnic groups [7, 17], even in individuals of the same ethnic group living in different cultural settings . For the Chinese population, a limited number of T2DM risk prediction models was developed. In 2009, Chien et al.  constructed a simple point model for the prediction of diabetes incidence in Taiwan, with the estimated AUC value of 0.702 (95% CI, 0.676–0.727). In 2010, Ko et al.  developed a simple risk score to identify the young-to-middle-aged Chinese population at high risk for diabetes in Hong Kong, with the AUC value of 0.735 (95% CI, 0.705–0.765). In 2011, Chuang et al.  provided multiple diabetic prediction equations, which were derived from large-scale health check-up data that estimated the risk of diabetes in Taiwan. In 2014, Xu et al.  demonstrated that the Framingham diabetes score underestimated diabetes incidence in older Chinese populations and constructed a prediction model for this population in Guangzhou (AUC = 0.779, 95% CI, 0.756–0.801). In 2017, Chen et al.  developed a noninvasive T2DM risk score model for rural adults in Deqing.
All of the studies conducted in China mentioned above were carried out in Southern and Eastern China. Considering the large diversity of the Chinese population, T2DM prediction models that reflect regional characteristics will be needed in other parts of China. Our study filled this gap. This nomogram will be applicable to most individuals in Central China. In addition, all of the previously conducted T2DM risk prediction studies in China established T2DM risk scores with integer point or segment values, whereas our nomogram can provide more accurate and individualized risk predictions because of the use of continuous values, which is in line with the EPMA’s view that individualization should be a general societal trend in medicine and healthcare [3, 4].
In our study, good degrees of discrimination and prediction ability were found both in the training cohort (AUC = 0.863 for females, AUC = 0.751 for males) and validation cohort (AUC = 0.847 for females, AUC = 0.755 for males), which indicated a relatively good predictive capability to discriminate individuals who are at risk to develop T2DM from those who are not. The AUC values of females were larger than those of males in this study; this result was consistent with some previous similar studies [20, 21]. Decision curve analysis demonstrated that the nomogram can avoid the need to perform OGTT in individuals with a low risk of developing T2DM in 3 years, which relieved both burden and costs. In addition, the analysis of a model by using all risk factors (i.e., full model) showed that including SBP and DBP did not improve prediction. The MFP model showed slightly better accuracy than the stepwise model, but the complicated formula restricted its use. (Details are described in the Supplemental Appendix.) Therefore, the stepwise model is the simplest model under the premise of guaranteeing accuracy.
The parameters included in this nomogram for females and males were age, BMI, FBG, LDLc, HDLc, and TG, which were also included in other prediction models [8, 10]. Considerable research has proved that advanced age is a nonmodifiable risk factor for diabetes manifestation . Impaired FBG is one of the diagnostic criteria for diabetes; studies have shown that FBG, hemoglobin A1c, and 2hPG all predict diabetes mellitus, yet test reliability is better for FBG and hemoglobin A1c than for 2hPG . Moreover, compared with hemoglobin A1c, the feasibility and applicability of FBG testing in low-resource settings are more pronounced. Multiple studies have found that obesity , dyslipidemia , and T2DM conditions typically co-exist in an individual and share common pathological mechanisms [23, 25] (insulin resistance, metabolic disorders, inflammation and alteration of gut microbiota, etc.). Therefore, the application of these parameters for this model is well-founded.
Some known risk factors of T2DM, such as dietary habits and physical inactivity, were not included in this study because they are difficult to assess precisely. Similarly, smoking status was excluded because of the wide presence of passive smoking. Given the insufficiency in the health system of China in the last century, many people possessed undiagnosed diabetes. Therefore, many participants are unsure of the medical history of the previous generation of their family. Therefore, we did not collect data on family history. In addition, Chien et al.  demonstrated that lifestyle factors were the insignificant predictors of T2DM, and the parameter of family history cannot improve risk prediction. Chung et al.  and Poltavskiy et al.  also suggested that physical activity was unrelated to T2DM risk or does not significantly affect the prediction of undiagnosed prediabetes.
The current study did not collect information on 2hPG, insulin resistance, and genetic markers because these tests are expensive, time-consuming, and not routinely measured in clinical practice. Moreover, the addition of 2hPG and insulin resistance was found to not improve the risk prediction accuracy of a simple clinical model [9, 27]. Noble et al.  also suggested that sociodemographic and clinical data are much better predictors for the risk of diabetes than genetic markers.
Developing a nomogram for T2DM by using parameters that can be collected in general health care settings results in considerable clinical and social implications, especially for the residents in mainland China, where OGTT is not easily accessible. Our nomogram presents a quantitative approach to distinguish the high-risk groups of T2DM who must focus on their physical conditions and follow advanced intervention strategies (e.g., lifestyle interventions, appropriate drug interventions, and/or surgical interventions) [5, 6] to prevent or at least delay the onset of T2DM. Meanwhile, further T2DM screening (OGTT) is unnecessary for those with low risk of T2DM at present, which may increase the cost-effectiveness of T2DM screening. This simple risk assessment tool can be accepted by nonprofessional personnel and healthcare workers.
Besides, we used a relatively large community-based sample to estimate the optimal cutoff values of each risk factor to predict T2DM incidence, which may provide references in defining the best thresholds of age, BMI, FBG, and serum lipid parameters for the Chinese population.
The strength of our study was the fact that the nomogram was constructed from a large physical examination data of community-dwelling residents, which can provide good generalizability to the rest of the population. The limitation of this study was the fact that we did not collect information on waist circumference, which was included in some previous similar studies. However, the AUCs of our nomogram without waist circumference remained satisfactory. In addition, the studies of Framingham , along with the study of Chien  and Xu , all found that including waist circumference instead of BMI in the model cannot improve the prediction performance.