Comparative analysis and interpretability of survival models for melanoma prognosis

Abstract

The increasing incidence of cutaneous melanoma highlights the critical need for precision medicine to optimize treatment strategies and improve patient outcomes. This study evaluated and compared the performance and interpretability of various survival models for melanoma patients. We analyzed survival outcomes, including overall survival (OS), specific survival (SS), and disease-free survival (DFS), using the Xarxa Melanoma database, which comprises data from over 9,000 patients across 19 hospitals collected over a decade. The performance of Cox Proportional Hazards, Random Survival Forest, XGBoost, DeepSurv, and DeepHit models was assessed using the Concordance Index (C-Index), time-dependent Area Under the Curve (tdAUC), ROC-AUC, and Inverse Brier score. Our analysis showed that the Random Survival Forest outperformed the other methods, achieving C-Index scores of 0.846 for OS, 0.869 for SS, and 0.846 for DFS. Furthermore, by applying an interpretability method specifically designed for survival analysis, we identified histological subtypes associated with poor prognoses. Additionally, this study reaffirms the established importance of the Breslow index as a key prognostic factor in melanoma survival prediction.