ABSTRACT/PROJECT SUMMARY Surgical treatments such as transplantation often pose considerable analytic challenges to risk prediction for mortality. For example, disease prognosis and treatment decisions in pediatric acute liver failure (PALF) calls for a reliable tool to predict mortality risk. However, the development of this prediction tool is hampered by the high frequency of liver transplantation (LTx), the occurrence of which modifies the disease course of the patient and dependently censors the death event of interest. Existing competing risks methods are not well suited to risk prediction for PALF. Recognizing the substantial prognostic value in multiple longitudinal biomarkers as well as baseline covariates, we aim to tackle risk prediction in the presence of treatment-induced competing risks by developing, implementing and applying sensible and computationally feasible modeling, validation and inference procedures. In this project, (Aim 1) the team proposes a modeling framework that tackles the dependence be- tween death and LTx through aggregating information from multiple longitudinal and baseline covariates. When compared to existing methods, the proposed modeling strategy can integrate information from more longitudinal biomarkers to better capture patients' dynamic disease status. Next, (Aim 2) we propose a comprehensive set of validation procedures to evaluate prediction performance in the presence of competing risks. The methods assess prediction performances in both cumulative incidence prediction and marginal probability prediction to ascertain and enhance prediction performance from all angles. We also develop formal testing procedures to detect potential predictive heterogeneity among different subtypes of patients. Moreover, we propose (Aim 3) statistical procedures to examine LTx-benefit under a causal inference framework, accommodating subject- specific benefit to inform personalized LTx decisions. All statistical methods will be rigorously justified through extensive simulation studies, sensitivity analysis and theoretical derivations, to ensure their theoretical rigor and practical usefulness. The methods will be systematically applied to a recent PALF registry database. The final prediction tool will be disseminated to practitioners through a user-friendly web-interface (Aim 4), to facilitate PALF prediction and dynamic prediction. We anticipate that our methods will be broadly applicable to other clinical studies and will develop R packages for the broader research community. 1