Prediction models for the risk of cardiovascular disease in the general population: a systematic review

Aim: To give an overview of all prognostic models that predict risk of developing cardiovascular disease in the general population, and to describe clinical heterogeneity in predicted outcomes, study populations and included predictors.Data source: In June 2013 a systematic search was performed in Medline and Embase.Study selection: Studies were eligible if they described the development, validation or incremental value of a prognostic model predicting cardiovascular disease in the general population.Data extraction: Separate extraction forms were used for scoring development, validation and incremental value papers and identifying corresponding models. For each model, items such as study design, geographical location, prediction horizon, study population, predicted outcome, predictors in the model, and performance were extracted.Results: A total of 313 studies were included, consisting of 367 developed prognostic models, 519 external validations and 278 incremental value assessments. Only 70 models (19%) were externally validated by independent researchers. The Framingham and SCORE prediction models were the most often validated models (n=291 and n= 63). Most models excluded participants diseased at baseline (n=223), and predicted risk of cardiovascular disease or coronary heart disease (n=105 and n=117 respectively) over a 5 or 10-year period (n=47 and n=205). Furthermore, one-third of models were not stratified for gender, and included age and/or smoking as a predictor (>85%). Other common predictors were systolic blood pressure, diabetes and total cholesterol. The c-statistic was reported for 62% of the external validations (median 0.750, range 0.530-0.993). Crucial clinical and methodological information was often missing, and if reported, typically prone to substantial heterogeneity. For example, 71 different age ranges were reported, and over 35 main categories of predictors were identified. In addition, discordant outcome definitions and different composite outcomes were used. Only 81 models were internally validated and for 40 internal validations the c-statistic was reported (median 0.766, range 0.600-0.872). For 53 models, no prediction horizon was reported and for 82 models the intercepts or baseline hazards were not specified making them unable to use for individual risk predictions. Conclusion: There is an excess of published prognostic models with a large number of overlapping variables predicting cardiovascular disease in the general population. Despite this overabundance, the usefulness of most models remains unclear due to incomplete presentation of models, a lack of external validations, and heterogeneity in predicted outcomes and study populations.