Aim: We assessed the performance of non-invasive risk prediction models to predict 10-year incident type 2 diabetes in a large Swedish cohort.
Methods: Using the Västerbotten Intervention Programme (VIP) cohort, which includes serial oral glucose tolerance tests (OGTTs), we assessed discrimination (concordance (c)-statistic) and calibration (expected-to-observed probability ratio, integrated calibration index, calibration slope and plots) before and after recalibration in twelve non-invasive models. Incident diabetes cases were determined by an OGTT at a 10-year follow-up visit or through previously validated register-based cases.
Results: Among 91708 VIP participants, the 10-year diabetes incidence was 2.8 %. Most models had acceptable discrimination (c-statistic ≥0.70 and < 0.80). Discrimination was better in women and persons <50 years old. Eight models overestimated and four models underestimated mean absolute risk. Recalibration improved miscalibration in all models. Overall, the Framingham Personal and QDScore models’ predictions were most accurate but the Framingham model included more easily obtainable variables. Most models overestimated risk in older people while no consistent pattern was observed across sexes.
Conclusion: All models required recalibration to improve prediction accuracy. The Framingham personal model is recommended for risk predictions and will be the easiest to implement.