FROM PREDICTED RISK TO PREDICTED BURDEN OF CARDIOVASCULAR DISEASE

Purpose: Prediction models for the risk of developing cardiovascular disease (CVD) in the general population may be useful for estimating the burden of disease. However, assessment of the health loss associated with the predicted endpoint may be difficult because prediction models are commonly developed based on composite endpoints. Moreover, the composite endpoints may also differ between different prediction models regarding the types of CVD events included. Therefore, different prediction models might lead to different estimated burden. The purpose of this study is to explore the extent of the differences in definition and constitution of composite endpoints, in four widely used CVD risk prediction models, and to assess how these differences influence estimates of CVD burden.

Method(s): Data from a large Dutch cohort study (n=19484; mean follow up 12.3 years) was used to investigate differences in composite endpoints of four widely used CVD risk prediction models: the Adult Treatment Panel III (ATP), Framingham Global (FRS), Pooled Cohort Equations (PCE) and SCORE-low (SCORE) models. Across these four prediction models, we calculated the 10-year individual CVD risks and the corresponding health loss based on the CVD event types included in the composite endpoint. Subsequently, each prediction model was used to estimate the expected CVD burden in the 25% individuals with highest predicted risks, expressed as Quality-Adjusted Life Years (QALYs) lost.

Result(s): The observed constitution of the composite endpoints varied widely across the four models. For example, the percentage non-fatal MI events was 81%, 19%, 37%, and 0% according to ATP, FRS, PCE, and SCORE respectively, and for fatal MI this was 19%, 5%, 9%, and 57%, respectively. FRS predicted the highest CVD risks and the composite endpoint used in SCORE had the highest health burden. The predicted CVD burden in the 25% individuals with highest predicted risk was 0.19, 0.72, 0.36, and 0.23 QALYs lost per individual when using ATP, FRS, PCE and SCORE, respectively.

Conclusion(s): The investigated CVD risk prediction models showed huge variation in definition and constitution of the composite endpoints. This directly resulted in large differences in estimated CVD burden. When interpreting the estimated CVD burden derived with a risk prediction model, it is therefore crucial to consider which CVD event types are included, and which are excluded, in that prediction model.