Updating cardiovascular disease risk prediction equations to account for lipid lowering treatment drop-in during follow up: a national data-linkage cohort

Abstract

Abstract Background and Purpose No commonly recommended cardiovascular disease (CVD) risk prediction equations account for the effect of medications initiated during follow-up (treatment drop-in) in the cohort studies used to derive equations. We have previously demonstrated that treatment drop-in is common. In this study we aim to compare equations before and after accounting for lipid-lowering medication drop-in during follow-up. Methods De-identified individual-level linkage of multiple administrative health datasets in Aotearoa New Zealand was undertaken to establish a cohort of almost all New Zealanders without CVD or heart failure, alive and aged 30-74 years on December 31, 2006, with follow-up linkage to hospitalisations and mortality until 31 December 2018. We derived age- and sex-specific Cox regression models to predict 5-year CVD risk, using pre-defined routinely available CVD risk predictors, both with and without a time-dependent lipid-lowering medication drop-in variable. The drop-in effect was set at a 20% CVD risk reduction, based on randomised controlled trial evidence. We compared the observed and predicted risks of the two equations using calibration plots. Results During 19,728,300 person-years of follow-up (mean: 11.3 years), 1,746,695 people had 155,943 first CVD events. Kaplan-Meier adjusted 5-year CVD risk was 2.36% (95% CI: 2.33%, 2.39%) in women and 4.36% (95% CI: 4.31%, 4.40%) in men. 1,586,421 individuals were not taking lipid-lowering medications at baseline, among whom 371,415 initiated lipid-lowering medications during follow-up. Median predicted five-year CVD risk was slightly higher when using equations that accounted for lipid-lowering medication drop-in and identified significantly more individuals as ≥15% 5-year CVD risk than equations that did not (15,355 vs 9,680 in women and 46,603 vs 36,460 in men). As expected, the increased predicted risk in equations accounting for medication drop-in was most marked in the high-risk groups. Conclusions CVD risk prediction equations that do not adjust for treatment drop-in will underestimate risk and lead to undertreatment, particularly among the highest risk patients. Traditional methods to evaluate model performance are unable to identify the impact of treatment drop-in. We demonstrate that incorporating a time-dependent lipid-lowering medication effect from trials is likely to be an effective approach for addressing treatment drop-in in CVD risk prediction equations. Future studies will also need to address the impact of other common drop-in treatments, particularly blood pressure-lowering medications.