We introduce Binacox+, an advanced extension of the Binacox method for prognostic analysis of high-dimensional survival data, enabling the detection of multiple cut-points per feature. The original Binacox method leverages the Cox proportional hazards model, combining one-hot encoding with the binarsity penalty to simultaneously perform feature selection and cut-point detection. In this work, we enhance Binacox by incorporating a novel penalty term based on the L1 norm of coefficients for cumulative binarization, defined over a set of pre-specified, context-dependent cut-point candidates. This new penalty not only improves interpretability but also significantly reduces computational time and enhances prediction performance compared to the original method. We conducted extensive simulation studies to evaluate the statistical and computational properties of Binacox+ in comparison to Binacox. Our simulation results demonstrate that Binacox+ achieves superior performance in important cut-point detection, particularly in high-dimensional settings, while drastically reducing computation time. As a case study, we applied both methods to three real-world genomic cancer datasets from The Cancer Genome Atlas (TCGA). The empirical results confirm that Binacox+ outperforms Binacox+ in risk prediction accuracy and computational efficiency, making it a powerful tool for survival analysis in high-dimensional biomedical data.
翻译:暂无翻译