부산대학교 도서관

Artificial intelligence is gaining interest among clinicians, but its results are difficult to be interpreted, especially when dealing with survival outcomes and censored observations. Explainable machine learning (XAI) has been recently extended to this context to improve explainability, interpretability and transparency for modeling results. A cohort of 231 patients undergoing an allogeneic bone marrow transplantation was analyzed by XAI for survival by two different uni- and multi-variate survival models, proportional hazard regression and random survival forest, having as the main outcome the overall survival (OS) and its main determinants, using the survex package for R. Both models’ performances were investigated using the integrated Brier score, the integrated Cumulative/Dynamic AUC and the concordance C-index. Global explanation for the whole cohort was performed using the time-dependent variable importance and the partial dependence survival plot. The local explanation for each single patient was obtained via the SurvSHAP(t) and SurvLIME plots and the ceteris paribus survival profile. The survex package common interface ensured a good feasibility of XAI for survival, and the advanced graphical options allowed us to easily explore, explain and compare OS results coming from the two survival models. Before the modeling results to be suitable for clinical use, understandability, clinical relevance and computational efficiency were the most important criteria ensured by this XAI for survival approach, in adherence to clinical XAI guidelines.