Response to Rathmann, Kowall, and Schulze R athmann, Kowall, and Schulze (1) suggest that the diabetes risk score (DRS) model is no better than simple clinical models and thus is of limited utility. To support this contention, they compare our area under the receiver operating characteristic curve (AROC) with those reported for different models. However, the AROC of a test is population specific; therefore, comparisons between populations with different baseline risks are problematic because

more »

... sitivity and specificity are subject to alteration by disease prevalence (2). In general, the AROC decreases as prevalence increases as it is increasingly difficult to differentiate outcomes in less healthy populations. The In-ter99 subpopulation with age Ͼ39 years and BMI Ն25 kg/m 2 that we used had a 5-year risk of 5.7%, nearly 2.5-fold higher than the EPIC-Potsdam study used to train the German DRS (GDRS). These differences in baseline risk make discrimination of incident diabetes more difficult in the Inter99 population and may explain apparent differences in AROCs. The sensitivity and specificity are 90 and 57%, respectively, at the moderate risk cut point (DRS Ն4.0) and 38 and 92%, respectively, at the high risk cut point (DRS Ն7.3). However, this result also depends on the population studied, and cut point optimization depends on multiple factors. In our subsequent work on the entire Inter99 cohort (5-year risk ϭ 3.4%), an AROC of 0.84 was reported (3). We cannot compare the DRS directly to the GDRS because most of its dietary and lifestyle risk factors are not available in In-ter99. Instead, we compared the DRS to a noninvasive clinical model trained on the same Inter99 subpopulation and three published diabetes models. The noninvasive clinical model includes age, BMI, waist circumference, and family history, but excludes sex, hypertension, and smoking status because they did not improve performance. The Atherosclerosis Risk in Communities (ARIC) model (4) consists entirely of clinical measurements, while the Framingham and San Antonio models use clinical measurements plus fasting plasma glucose and lipids. In the Inter99 subpopulation, the DRS (AROC ϭ 0.78) is superior to the noninvasive clinical model (AROC ϭ 0.70; P ϭ 0.0025) and the ARIC (AROC ϭ 0.70; P ϭ 0.0004), Framingham (AROC ϭ 0.75; P ϭ 0.02), and San Antonio (AROC ϭ 0.75; P ϭ 0.09) models. Rathmann, Kowall, and Schulze also argue that we do not provide evidence that the DRS shows improvement over existing tools. The AROC does not reflect how a model would perform in clinical practice (2); however, we can estimate clinical utility as net reclassification improvement (NRI) based on tertiles of each risk score. The DRS shows superior classification when compared with the ARIC (NRI ϭ 0.178; P ϭ 0.0026), Framingham (NRI ϭ 0.271; P Ͻ 0.0001), or San Antonio (NRI ϭ 0.036; P ϭ 0.2377) models. Finally, it was our intent to develop a model to predict incident diabetes that could be conveniently adopted in routine clinical practice. It is well established that risk indexes that require questionnaires and calculations have not been widely adopted by physicians (5). The DRS requires a single fasting blood specimen and provides a score of 5-year risk that is simple to interpret. Given its superior performance over other clinical indexes demonstrated in Inter99 and its ease of use, we believe that the DRS provides physicians with an improved tool for assessing diabetes risk in primary care settings.