A comparison of the PROCAM and Framingham point-scoring systems for estimation of individual risk of coronary heart disease in the Second Northwick Park Heart Study

Abstract

We have compared the predictive value of the PROCAM and Framingham risk algorithms in healthy UK men from the Second Northwick Park Heart Study (NPHS-II) (50-64 years at entry), followed for a median of 10.8 years for coronary heart disease (CHD) events. For PROCAM, the area under the receiver operating characteristic (ROC) curve was 0.63 (95% CI, 0.59-0.67), and not significantly different (p = 0.46) from the Framingham score, 0.62 (0.58-0.66). Sensitivities for a 5% false-positive rate (DR(5)) were 13.8 and 12.4%, respectively. Calibration analysis for PROCAM gave a ratio of observed to expected events of 0.46 (Hosmer-Lemeshow test, p < 0.0001) and 0.47 for Framingham (p < 0.0001). Using measures taken at 5 years of high-density lipoprotein cholesterol and (estimated) low-density lipoprotein cholesterol levels increased the ROC by only 1%. An NPHS-II risk algorithm, developed using a 50% random subset, and including age, triglyceride, total cholesterol, smoking status, and systolic blood pressure at recruitment, gave an ROC of 0.64 (0.58-0.70) with a DR(5) of 10.7% when applied to the second half of the data. Adding family history and diabetes increased the DR(5) to 18.4% (p = 0.28). Adding lipoprotein(a) >26.3 mg/dL (relative risk 1.6, 1.1-2.4) gave a DR(5) of 15.5% (p = 0.55), while adding fibrinogen levels (relative risk for 1S.D. increase = 1.5, 1.1-2.0) had essentially no additional impact (DR(5) = 16.9%, p = 0.95). Thus, the PROCAM algorithm is marginally better as a risk predictor in UK men than the Framingham score, but both significantly overestimate risk in UK men. The algorithm based on NPHS-II data performs similarly to those for PROCAM and Framingham with respect to discrimination, but gave an improved ratio of observed to expected events of 0.80 (p = 0.01), although no score had a high sensitivity. Any novel factors added to these algorithms will need to have a major impact on risk to increase sensitivity above that given by classical risk factors.