Transfer learning for high-dimensional win ratio regression

Transfer learning is a crucial machine learning technique that addresses the issue of insufficient target domain data by integrating cross-domain knowledge. Currently, transfer learning has been widely applied in fields such as text classification, image recognition, and disease diagnosis. In statistical model research, its application has primarily focused on classical models like classification and regression, while studies on hierarchical composite endpoints remain scarce. Win ratio regression is a pivotal statistical tool for analyzing hierarchical composite endpoints; paired with regularization, it enables automated high-dimensional covariates selection and holds substantial value in cardiovascular clinical trials. However, in practical scenarios with limited target samples, the conventional regularized win ratio regression models suffer from low prediction accuracy and unstable variable selection. Therefore, this article considers utilizing sample sets that differ from but share similarities with the target dataset to enhance the model’s estimation and predictive performance in the target domain, while employing a source detection mechanism to prevent negative transfer. We propose a two-stage transfer learning-based regularized win ratio regression model (TL-WRnet) for variable selection and risk prediction that effectively incorporates information from auxiliary datasets while preserving target domain specificity, significantly enhancing model performance in small-sample scenarios. Numerical experiments and clinical data applications validate the superiority and practicality of the proposed approach.