Comparison of machine learning and deep learning-based survival models: An explainable artificial intelligence approach

Aim: This study aimed to compare the predictive performance of ten models representing statistical, tree-based, ensemble, kernel, and deep learning approaches on colon cancer survival data and to evaluate their decision mechanisms using explainability methods. Material and method: The colon dataset from the R survival package (n=888, events=430, event rate=48.4%) was used. Model performance was evaluated using 5×3 nested cross-validation with C-index, Integrated Brier Score, and time-dependent AUC. Model reliability was assessed through 1000-iteration bootstrap with 95% confidence intervals. Results: In nested cross-validation the highest mean C-index values were very close across DeepSurv (0.6569), RSF (0.6568), and DeepHit (0.6567); RSF showed the most balanced performance with the lowest IBS (0.1743) and one of the highest TD-AUC values (0.7161). In bootstrap analysis ExtraSurvTrees achieved the highest C-index (0.7350), IBS (0.1503), and TD-AUC (0.8098), although its low nested-CV ranking indicates the optimistic bias of this estimate. SHAP analysis identified the number of positive lymph nodes as the strongest prognostic factor across all models. SurvSHAP(t) revealed time-dependent variations in feature importance, while SurvLIME provided patient-level interpretable risk profiles. Conclusion: Ensemble methods demonstrated consistent superiority while deep learning models carried instability risk on small-to-medium datasets. Explainability analyses confirmed consistent prognostic patterns across models, laying the groundwork for clinical decision support applications.