Comment on Vassy et al. Polygenic Type 2 Diabetes Prediction at the Limit of Common Variant Detection. Diabetes 2014;63:2172–2182

Vassy et al. (1) present data from three young adult cohorts to investigate additive predictive values of recently discovered genetic variants for type 2 diabetes (T2D). In the past 20 years, many prediction models have been developed for T2D risk over a wide range of populations (2). More accurate estimation of T2D risk is critical to allocating resources and planning health policies. So far, there is little evidence that any of epidemiologic models with or without genetic markers can accurately estimate T2D risk in different populations.

To bridge the gap between the development and the utility of prediction models, recent studies have validated and compared a variety of models in prospective population-based cohorts (2,3). The evidence shows that models are generally acceptable to identify high-risk individuals for T2D over 10 years in white middle-aged adults (2–4). Moreover, all prediction models were demonstrated to provide better prediction in people younger than 60 years (4). The worldwide epidemic is shifting prevalence of T2D to younger ages (5); therefore, the burden of T2D will rise as the duration of diabetes is elongated. Current modeling approaches do not account for potential variation of the relationship between risk factors and the disease with age. Therefore, younger individuals are commonly classified as low-risk subgroups in the short-term horizon, but of course several of these subjects would be at high-risk if one considers it from a long-term perspective. Vassy et al. (1) analyzed this important aspect in which the use of genetic data led to a greater improvement in prediction among young adults. Likewise, all prediction models are based on incorporation of baseline data and do not include more information than that of one single time point from which prediction started. Consequently, the contribution of risk factors to diabetes risk and potential interactions with genetic factors has been assumed to be constant over time. It merits investigation of what extent existing models in combination with genetic markers predict risk of T2D across the life span.

Yet another aspect is that the clinical utility of prediction models depends on a number of factors and has some challenges (4). For instance, if a physician would like to predict an individual’s T2D risk, gathering personal health data such as age, sex, smoking history, and family history of diabetes only requires a small interview. Assessment of (central) obesity, blood pressure, and routine biochemical testing (e.g., glucose and cholesterol) already requires more effort. Testing of additional biologic and genetic markers is even further away. In clinical prediction, one would typically first come to an impression based on the interview and the simple physical measurements. If the picture still was not clear, one would proceed with an assessment of additional biochemical tests, typically from blood or urine collected during the same or a next visit. This differs from how prediction models are currently being envisioned for clinical utility (2,4). Current evaluation is being done as if all data are available at the same time (6). Better mimicking of actual practice in primary/secondary care would be an important next step in generating models.