Performance of the Framingham models and Pooled Cohort Equations for prediction of cardiovascular disease in the general population: a meta-analysis

Background: The implementation of the Framingham risk models and Pooled Cohort Equations (PCE) is currently recommended in the United States for predicting ten-year risk of developing cardiovascular disease (CVD). These prediction models have been extensively validated in other individuals and settings.
Objectives: To systematically review and summarize the discrimination and calibration of three CVD prediction models, and to determine heterogeneity in performance of these models across subpopulations or geographical regions.
Methods: In December 2015, Medline, Embase, Web of Science, and Scopus were searched for studies investigating the external validation of three CVD prediction models (Framingham Wilson 1998, Framingham ATP III 2002 and PCE 2013). Studies published before June 2013 were identified from a previous review. Studies were eligible for inclusion if they validated the original prediction model without updating, in a general population setting. Critical appraisal was performed based on the CHARMS checklist. Data were extracted on case-mix, essential study design characteristics, and model performance (quantified by the c-statistic and observed/expected ratio). Performance estimates were summarized using random effects meta-analysis models that accounted for differences in case-mix to explore sources of heterogeneity.
Results: The search identified 10,687 references, of which 1,501 were screened in full text and 47 met our eligibility criteria. These articles described the external validation of Framingham Wilson (27 articles), Framingham ATP III (16 articles) or the PCE (10 articles). Discriminative performance (c-statistic) varied between 0.56 and 0.92. At the Cochrane colloquium, we will present how case-mix differences (e.g. age, comorbidities, treatment) influence the performance of these models.
Conclusions: The results of this study can help in identifying which of these three CVD models can reliably be used, whether there is heterogeneity in their performances, and whether there are subpopulations for which further research is necessary to improve CVD risk prediction.