A1987
Title: Survival trees for competing risks data with time-varying covariates
Authors: Seungjoo Kim - Sookmyung Women University (Korea, South) [presenting]
Yang-Jin Kim - Sookmyung Women University (Korea, South)
Abstract: In real-world clinical data, patients may experience terminal events from various causes. To accommodate this data structure, various methodologies have been developed under the competing risks framework. When time-varying covariates are present during patient follow-up, the traditional linear assumptions applied in the PH model may not be appropriate. However, studies on competing risks models incorporating time-varying covariates remain limited. A new survival tree approach is proposed wherein time-varying covariates are handled by transforming the data into an LTRC (left-truncated and right-censored) structure via the counting process formulation, and a new test statistic is introduced as the splitting criterion suitable for this data structure. The predictive performance of the proposed model is evaluated across various simulation environments using the Integrated Brier Score (IBS) and the C-index. Additionally, the practical applicability of the proposed method is demonstrated by applying it to real-world data from patients in the intensive care unit (ICU) of a university hospital, where the occurrence of hospital-acquired pneumonia (HAP) is treated as a time-varying covariate, and ICU death and discharge alive are defined as competing events.
