Abstract 14191: New Approaches to Factual and Counterfactual Risk Prediction for Cardiovascular Disease Using the Parametric G-formula

Abstract

Introduction: Over the last few decades, clinical prediction models of the risk of cardiovascular disease have proliferated and are now used extensively in research and clinical care. However, most models have several limitations: (1) they often use only the baseline examination cycle despite the fact that the underlying data sources often include longitudinal assessments of risk factors over time, (2) they make predictions under the "natural course" for the population from which they were derived without making explicit how risk factors varied over time, and (3) they are not appropriate for predictions under alternative counterfactual scenarios such as if an individual were to initiate treatment or adopt healthy lifestyle changes. In this study, we consider an alternative approach based on the parametric g-formula which resolves these limitations. Methods: We compare a g-formula-based model to standard approaches to prediction modeling via Monte Carlo simulation as well as in empirical data. For the latter, we use data from 14,222 longitudinal assessments of 2,771 participants in the Framingham Offspring Study (FOS) to predict 10-year risk of ASCVD. Model discrimination and calibration are assessed via C statistic and calibration slope. We also investigate improvements in risk estimates from our g-formula model as compared to those from the Pooled Cohort Equations (PCE) using the integrated discrimination index and net reclassification index. Results: In simulations, the g-formula model performs better than standard approaches when longitudinal observations are added and makes correct counterfactual predictions when assumptions are met. When applied to FOS data, the g-formula model had better discrimination (C statistic, 0.77 vs. 0.73) and better calibration (slope, 1.07 vs. 0.83) than the PCE. Improvements in the net re-classification index at the 7.5% risk threshold (0.14; 95% CI: 0.05 - 0.26) as well as the integrated discrimination index (2.6%; 95% CI: 1.5% - 3.7%) were significant, suggesting the g-formula better classifies those at increased risk of ASCVD. Conclusion: A longitudinal model based on the g-formula could better use available data from cohort studies and provides a principled framework for predicting risk under interventions.