Predictive performance of established cardiovascular risk scores in the prediabetic population: external validation using the UK Biobank data set

Abstract

Abstract Aims Prediabetes is a highly heterogenous metabolic state with increased risk of cardiovascular disease (CVD). Current guidelines raised the necessity of CVD risk scoring for prediabetes without clear recommendations. Thus, this study aimed to systematically assess the performance of 11 models, including five general population-based and six diabetes-specific CVD risk scores, in prediabetes. Methods and results A cohort of individuals aged 40–69 years with prediabetes (HbA1c ≥ 5.7 and <6.5%) and without baseline CVD or known diabetes was identified from the UK Biobank, which was used to validate 11 prediction models for estimating 10- or 5-year risk of CVD. Model discrimination and calibration were evaluated by Harrell's C-statistic and calibration plots, respectively. We further performed decision curve analyses to assess the clinical usefulness. Overall, 56 831 prediabetic individuals were included, of which 4303 incident CVD events occurred within a median follow-up of 8.9 years. All the 11 risk scores assessed had modest C-statistics for discrimination ranging from 0.647 to 0.680 in prediabetes. Scores developed in the general population did not outperform those diabetes-specific models (C-statistics, 0.647–0.675 vs. 0.647–0.680), while the PREDICT-1° Diabetes equation developed for Type 2 diabetes performed best [0.680 (95% confidence interval, 0.672–0.689)]. The calibration plots suggested overall poor calibration except that the PREDICT-1° Diabetes equation calibrated well after recalibration. The decision curves generally indicated moderate clinical usefulness of each model, especially worse within high threshold probabilities. Conclusion Neither risk stratification schemes for the general population nor those specific for Type 2 diabetes performed well in the prediabetic population. The PREDICT-1° Diabetes equation could be a substitute in the absence of better alternatives, rather than the general population-based scores. More precise and targeted risk assessment tools for this population remain to be established.