Performance of current guidelines for coronary heart disease prevention: optimal use of the Framingham-based risk assessment.

BACKGROUND: There is a strong positive association between Framingham Risk Scores (FRS) in a population and incidence of hard coronary heart disease (hCHD) events. Under current Adult Treatment Panel III guidelines, individuals with FRS that indicate ≥20% 10-year risk of hCHD are recommended to receive intensive medical risk factor modification. We sought to assess the performance of FRS as a predictive tool when used as in current guidelines. METHODS: A retrospective analysis of two prospective cohort studies, the Atherosclerosis Risk in Communities (ARIC) study, and Cardiovascular Health Study (CHS), including 11,436 and 2569 participants, respectively, without known cardiovascular disease or diabetes at baseline, with available FRS variables were analyzed. The FRS was computed according to standard algorithm. The main outcome was hCHD event defined as MI or coronary death. Using Receiver Operating Characteristics (ROC) curves, sensitivity, specificity, accuracy and other test performance characteristics were determined at various 10-year risk thresholds. ROC curves were plotted. RESULTS: During 10-year follow-up, 822 hCHD events occurred. FRS was significantly associated with hCHD with an AUC of 0.77 and 0.68 for ARIC and CHS, respectively (p-values <0.0001). However, at standard "high risk" cut-off (≥20%), the sensitivity of FRS was only 13% and 25%, respectively and Youden's Index was only 0.10 and 0.15. Lowering the 10-year risk threshold to >5% improved prediction sensitivity to 75% and 83%, with specificity of 66% and 40%, respectively. CONCLUSION: When used dichotomously as in current guidelines, sensitivity of the conventional 20% 10-year risk threshold for subsequent hCHD events is quite low. Since the 20% 10-year risk threshold for intensive medical risk factor therapy is on the steep part of the ROC curve, lowering the threshold results in substantial increases in sensitivity with much smaller losses in specificity, even to a threshold as low as 5%.

Duke Scholars

Author Michael J Pencina Biostatistics & Bioinformatics, Division of Biostatistics