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

Area of Science:

• Biostatistics
• Clinical Trials
• Statistical Modeling

Background:

• Hierarchical composite endpoints are crucial in clinical trials, prioritizing critical outcomes like mortality.
• Existing regression frameworks for win ratio analysis struggle with high-dimensional datasets and numerous predictors.
• A robust variable selection method is needed for win ratio regression.

Purpose of the Study:

• To propose an elastic net-type regularization approach for win ratio regression.
• To extend the proportional win-fractions (PW) model to handle high-dimensional data.
• To provide a scalable and robust variable selection method for win ratio analysis.

Main Methods:

• Developed an elastic net-type regularization for win ratio regression, extending the PW model.
• Adapted pairwise comparisons for penalized regression and used subject-level cross-validation for model selection.
• Defined performance metrics using a generalized concordance index and implemented the method in the wrnet R-package.

Main Results:

• Simulation studies showed wrnet outperformed traditional Cox regression in time-to-first-event analyses.
• wrnet demonstrated superior performance in scenarios with differing covariate effects on mortality and nonfatal events.
• Application to HF-ACTION trial data identified prognostic variables and improved predictive accuracy over regularized Cox models.

Conclusions:

• The wrnet approach integrates the interpretability of the win ratio with the scalability of elastic net regularization.
• The wrnet R-package offers a user-friendly tool for applying these advanced statistical procedures.
• Future work may address non-proportionalities and nonlinearities in covariate effects within the win ratio framework.