Predictive Performance of Cardiovascular Disease Risk Prediction Algorithms in People Living With HIV

Abstract

BACKGROUND: People living with HIV (PLWH) experience a higher cardiovascular disease (CVD) risk. Yet, traditional algorithms are often used to estimate CVD risk. We evaluated the performance of 4 commonly used algorithms. SETTING: The Netherlands. METHODS: We used data from 16,070 PLWH aged ≥18 years, who were in care between 2000 and 2016, had no pre-existing CVD, had initiated first combination antiretroviral therapy >1 year ago, and had available data on CD4 count, smoking status, cholesterol, and blood pressure. Predictive performance of 4 algorithms [Data Collection on Adverse Effects of Anti-HIV Drugs Study (D:A:D); Systematic COronary Risk Evaluation adjusted for national data (SCORE-NL); Framingham CVD Risk Score (FRS); and American College of Cardiology and American Heart Association Pooled Cohort Equations (PCE)] was evaluated using a Kaplan-Meier approach. Model discrimination was assessed using Harrell's C-statistic. Calibration was assessed using observed-versus-expected ratios, calibration plots, and Greenwood-Nam-D'Agostino goodness-of-fit tests. RESULTS: All algorithms showed acceptable discrimination (Harrell's C-statistic 0.73-0.79). On a population level, D:A:D, SCORE-NL, and PCE slightly underestimated, whereas FRS slightly overestimated CVD risk (observed-versus-expected ratios 1.35, 1.38, 1.14, and 0.92, respectively). D:A:D, FRS, and PCE best fitted our data but still yielded a statistically significant lack of fit (Greenwood-Nam-D'Agostino χ ranged from 24.57 to 34.22, P < 0.05). Underestimation of CVD risk was particularly observed in low-predicted CVD risk groups. CONCLUSIONS: All algorithms perform reasonably well in PLWH, with SCORE-NL performing poorest. Prediction algorithms are useful for clinical practice, but clinicians should be aware of their limitations (ie, lack of fit and slight underestimation of CVD risk in low-risk groups).