Regularized win ratio regression for variable selection and risk prediction, with an application to a cardiovascular trial

TL;DR

This paper introduces a new method for analyzing complex health outcomes in trials by combining win ratio analysis with regularization techniques, improving prediction accuracy.

Contribution

The paper proposes an elastic net-type regularization approach for win ratio regression, enabling variable selection and risk prediction in high-dimensional settings.

Findings

The wrnet method outperforms regularized Cox regression in scenarios with differing covariate effects on mortality and nonfatal events.

Application to the HF-ACTION trial showed superior predictive accuracy compared to traditional models.

The method is implemented in the wrnet R-package, offering a user-friendly interface for clinical researchers.

Abstract

The win ratio has been widely used in the analysis of hierarchical composite endpoints, which prioritize critical outcomes such as mortality over nonfatal, secondary events. Although a regression framework exists to incorporate covariates, it is limited to low-dimensional datasets and may struggle with numerous predictors. This gap necessitates a robust variable selection method tailored to the win ratio framework.

We propose an elastic net-type regularization approach for win ratio regression, extending the proportional win-fractions (PW) model in low-dimensional settings. The method addresses key challenges, including adapting pairwise comparisons to penalized regression, optimizing model selection through subject-level cross-validation, and defining performance metrics via a generalized concordance index. The procedures are implemented in the wrnet R-package, publicly available at https://lmaowisc.github.io/wrnet/.

Simulation studies demonstrate that wrnet outperforms traditional (regularized) Cox regression for time-to-first-event analysis, particularly in scenarios with differing covariate effects on mortality and nonfatal events. When applied to data from the HF-ACTION trial, the method identified prognostic variables and achieved superior predictive accuracy compared to regularized Cox models, as measured by overall and component-specific concordance indices.

The wrnet approach combines the interpretability and clinical relevance of the win ratio with the scalability and robustness of elastic net regularization. The accompanying R-package provides a user-friendly interface for routine application of the procedures, whenever appropriate. Future research could explore additional applications or refine the methodology to address non-proportionalities in win-loss risks and nonlinearities in covariate effects.

Figures