The American Diabetes Association's 57th Annual Advanced Postgraduate Course

Risk prediction in diabetes Michael Stern (San Antonio, TX) discussed risk prediction in diabetes, addressing the use of biomarkers and approaches with risk calculation engines. In general, he noted, the state of wellness is inherently asymptomatic without functional impairments, and many individual with diagnosed diabetes, hypertension, and even with cardiovascular disease (CVD) are “well.” For these individuals, treatment provides no current benefit but only the promise of a future benefit, which may never come to pass. If a well person in this sense cannot be made better, Stern pointed out that we must be particularly cognizant of the dictum ascribed to Niels Bohr: “Prediction is very difficult, especially about the future.” A tool used in analysis of tests is the calculation of the area under the curve (AUC) of the receiver-operating characteristic graph of sensitivity versus (1specificity). The AUC can be interpreted to equal the likelihood that a person destined to develop the disease or characteristic being tested for has a higher score, comparing one person who is with another who is not going to develop the given outcome (1). The AUC should be contrasted with an odds ratio (OR) or relative risk, which may be calculated as equaling (sensitivity) (1 false positive rate)/[(1 sensitivity) (false positive rate)]. This may better be thought of as pertaining to populations, with an OR of 1.5–3.0 giving rise to AUCs of 0.57– 0.68, which Stern termed not terribly impressive levels, as can be observed in the overlap of distribution curves of those developing versus not developing the disease or characteristic. An OR of 10 is needed to give a reasonable AUC of 0.83 and, hence, to distinguish individuals, whereas an OR of 1.5 is useful in distinguishing populations. Stern reviewed the San Antonio study diabetes prediction model, which is based on age, sex, ethnicity, fasting glucose, systolic blood pressure, BMI, and family history (2). The AUC is 0.85 for this model, a level higher than that seen with the fasting or the 2-h postglucose load blood glucose alone. Stern commented also that impaired fasting glucose (IFG) and impaired glucose tolerance (IGT) are given as fixed points, while this (or any) predictive equation gives a range of values whose cut point can be set at whatever level seems appropriate for a given population. A recently proposed approach is the commercially available Tethys assay, which gives a diabetes risk score (DRS) using fasting glucose, 2-h insulin, adiponectin, ferritin, interleukin-2 receptor , C-reactive protein, and, in some applications, A1C. The DRS has been validated based on the Inter99 study of 6,794 Danes (3) and the Botnia study of 1968 Finnish and Swedish subjects (4), with 5and 15-year follow-up, respectively. The score ranges from 1–10 and low risk is considered 4.5, moderate 4.5–8, and high 8. Stern illustrated the use of the DRS by reviewing its contribution to risk assessment in those with IFG, whose 9% risk of developing diabetes exceeds by up to 2% the risk of those with NFG. Among the IFG group, those with low DRS (comprising 26% of the group) have 2% risk, those with moderate risk (53%) have 7%, and the 21% with high-risk DRS scores have 24% risk of converting to diabetes over 5 years, suggesting that clinically useful additional information can be gained from this approach (3). In a similar analysis, the 25% of the population with metabolic syndrome had an overall 8% risk, while those remaining had 2% risk of developing diabetes. Stratifying those with metabolic syndrome, the 23% with low DRS had 2% diabetes risk, the 54% with moderate DRS had 6% diabetes risk, and the 23% with high DRS developed diabetes at a 23% rate. Thus, risk models are more potent than dichotomous measures such as IFG and metabolic syndrome and, hence, “contain more information about the future.” The AUC for the DRS is similar to that of the Stern risk model at a bit over 0.8, suggesting that a clinical score (inexpensive, in the sense that ascertainment of data by clinicians is considered to be “free”) would have similar benefit to the Tethys test. Why, Stern asked, should we bother with risk prediction scores? How is the information to be used? Certainly, one can more aggressively treat blood pressure and lipids, but he commented, “you don’t really need a diabetes risk score” for this. The reason, rather, is precisely to prevent diabetes and, more importantly, by doing so to prevent its complications. “We know that we can prevent diabetes,” as the Finnish Diabetes Prevention Study, the Da Qing Diabetes Study, the Diabetes prevention program, Tripod, DREAM, and Act-Now have shown. “I am of the opinion,” Stern continued, “that we really need studies that show an impact on hard end points.” Surrogate measures such as glucose, blood pressure, coronary Ca , and carotid IMT “can have abnormal values [while the person is] ‘well.’” Actual morbidities to be prevented are death, heart attack, stroke, end-stage renal dis● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●