Clinically Interpretable Survival Prediction in Primary Biliary Cholangitis with TreeSHAP and Gradient-Boosted Model

Survival prediction tools aid management of chronic liver diseases by informing decisions on monitoring and transplantation. In primary biliary cholangitis (PBC), classical risk scores discriminate well but are hard to interpret and treat variables in isolation, whereas clinicians think in terms of broader constructs such as cholestasis or coagulopathy. We introduce Surv-TCAV, a concept-based interpretability method for accelerated failure-time (AFT) survival models, and apply it to gradient-boosted trees on the Mayo Clinic PBC cohort (276 complete cases, 17 baseline variables). An XGBoost-AFT model was trained on 25 repeated 80/20 splits; performance was assessed using Harrell’s C-index and the integrated Brier score, with Cox and Weibull AFT models as transparent baselines. TreeSHAP provided feature-level attributions, while Surv-TCAV probed model sensitivity along clinician-defined directions (cholestasis, coagulopathy, low albumin, older age and clinical complications). The boosted model achieved high discrimination (C-index ≈ 0.835) but weaker calibration (IBS ≈ 0.36), reflecting the usual flexibility–calibration trade-off. Both feature- and concept-level explanations aligned with hepatology practice: bilirubin, alkaline phosphatase, copper, low albumin, prolonged prothrombin time, age and clinical complications all reduced predicted survival. These results show how boosted AFT models can be interpreted in familiar clinical terms. The PBC analysis is a methodological illustration; all code, figures and data are provided for reproducibility.