Non-invasive risk scores for prediction of type 2 diabetes (EPIC-InterAct): A validation of existing models
Kengne AP., Beulens JWJ., Peelen LM., Moons KGM., van der Schouw YT., Schulze MB., Spijkerman AMW., Griffin SJ., Grobbee DE., Palla L., Tormo MJ., Arriola L., Barengo NC., Barricarte A., Boeing H., Bonet C., Clavel-Chapelon F., Dartois L., Fagherazzi G., Franks PW., Huerta JM., Kaaks R., Key TJ., Khaw KT., Li K., Mühlenbruch K., Nilsson PM., Overvad K., Overvad TF., Palli D., Panico S., Quirós JR., Rolandsson O., Roswall N., Sacerdote C., Sánchez MJ., Slimani N., Tagliabue G., Tjønneland A., Tumino R., van der A DL., Forouhi NG., Sharp SJ., Langenberg C., Riboli E., Wareham NJ.
Background: The comparative performance of existing models for prediction of type 2 diabetes across populations has not been investigated. We validated existing non-laboratory-based models and assessed variability in predictive performance in European populations. Methods: We selected non-invasive prediction models for incident diabetes developed in populations of European ancestry and validated them using data from the EPIC-InterAct case-cohort sample (27 779 individuals from eight European countries, of whom 12 403 had incident diabetes). We assessed model discrimination and calibration for the first 10 years of follow-up. The models were first adjusted to the country-specific diabetes incidence. We did the main analyses for each country and for subgroups defined by sex, age ( < 60 years vs ≥60 years), BMI ( < 25 kg/m 2 vs ≥25 kg/m 2 ), and waist circumference (men < 102 cm vs ≥102 cm; women < 88 cm vs ≥88 cm). Findings: We validated 12 prediction models. Discrimination was acceptable to good: C statistics ranged from 0·76 (95% CI 0·72-0·80) to 0·81 (0·77-0·84) overall, from 0·73 (0·70-0·76) to 0·79 (0·74-0·83) in men, and from 0·78 (0·74-0·82) to 0·81 (0·80-0·82) in women. We noted significant heterogeneity in discrimination (p heterogeneity < 0·0001) in all but one model. Calibration was good for most models, and consistent across countries (p heterogeneity > 0·05) except for three models. However, two models overestimated risk, DPoRT by 34% (95% CI 29-39%) and Cambridge by 40% (28-52%). Discrimination was always better in individuals younger than 60 years or with a low waist circumference than in those aged at least 60 years or with a large waist circumference. Patterns were inconsistent for BMI. All models overestimated risks for individuals with a BMI of < 25 kg/m 2 . Calibration patterns were inconsistent for age and waist-circumference subgroups. Interpretation: Existing diabetes prediction models can be used to identify individuals at high risk of type 2 diabetes in the general population. However, the performance of each model varies with country, age, sex, and adiposity. Funding: The European Union. © 2014 Elsevier Ltd.