Performance of the SCORE and Globorisk cardiovascular risk prediction models: a prospective cohort study in Dutch general practice

Abstract

Background GPs frequently use 10-year-risk estimations of cardiovascular disease (CVD) to identify high-risk patients. Aim To assess the performance of four models for predicting the 10-year risk of CVD in Dutch general practice. Design and setting Prospective cohort study. Routine data (2009–2019) was used from 46 Dutch general practices linked to cause of death statistics. Method The outcome measures were fatal CVD for SCORE and first diagnosis of fatal or non-fatal CVD for SCORE fatal and non-fatal (SCORE-FNF), Globorisk-laboratory, and Globorisk-office. Model performance was assessed by examining discrimination and calibration. Results The final number of patients for risk prediction was 1981 for SCORE and SCORE-FNF, 3588 for Globorisk-laboratory, and 4399 for Globorisk-office. The observed percentage of events was 18.6% (n = 353) for SCORE-FNF, 6.9% (n = 230) for Globorisk-laboratory, 7.9% (n = 323) for Globorisk-office, and 0.3% (n = 5) for SCORE. The models showed poor discrimination and calibration. The performance of SCORE could not be examined because of the limited number of fatal CVD events. SCORE-FNF, the model that is currently used for risk prediction of fatal plus non-fatal CVD in Dutch general practice, was found to underestimate the risk in all deciles of predicted risks. Conclusion Wide eligibility criteria and a broad outcome measure contribute to the model applicability in daily practice. The restriction to fatal CVD outcomes of SCORE renders it less usable in routine Dutch general practice. The models seriously underestimate the 10-year risk of fatal plus non-fatal CVD in Dutch general practice. The poor model performance is possibly because of differences between patients that are eligible for risk prediction and the population that was used for model development. In addition, selection of higher-risk patients for CVD risk assessment by GPs may also contribute to the poor model performance.