PRM65 - An Evaluation Of Competing Models For Predicting Cv Event Rates From Ldl-C Levels In Secondary Prevention

Abstract

This study aims to evaluate four alternative models that each describe the relationship between LDL-C and CV event rates in secondary prevention. A secondary aim of this analysis is to promote an approach to develop more reliable disease simulation models in cardiovascular disease. Data describing LDL-C and nfMI+CV Death outcomes were abstracted from several landmark secondary prevention statin trials in clinically stable patient populations (4S, CARE, LIPID, HPS, TNT, IDEAL, GREACE, AVERT, and LIPS). Linear (L), quadratic (Q), one-knot linear spline (S1L) and two-knot quadratic spline (S2Q) models were fit. RMSE, leave-one-out cross validated (LOOCV) and Monte Carlo cross validated (MCCV) RMSE (90% training, 1,000 replicates) were used to evaluate predictive performance. Predicted event rates based on LDL-C = 50mg/dL were estimated to illustrate the clinical implications of each model. To encourage full reproducibility and transparency, all raw data and code (R) will be made freely available for download online via authors Git repository. A total of eight models were fit and evaluated, including four distinct functional forms (L, Q, S1L, and S2Q) across two sets of data (all data (A) and data censured for high leverage studies(C)). Of all the models evaluated, S2Q-C exhibited the lowest RMSE (0.317) while L-C produced both the lowest LOOCVRMSE (0.463) and MCCVRMSE (0.389). Model selection had an impact on predicted percentage of events/yr at 50mg/dL (range= 0.769 to 2.09). This analysis suggests that a linear model may be among the most appropriate when describing the relationship between LDL and nfMI+CV Death in secondary prevention; however, other models (linear spline and quadratic) warrant further investigation. These findings along with additional research will help inform more reliable and accurate models of long-term outcomes and economic benefit of LDL-C lowering treatment modalities.