Paper details

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

Author: Emmanuel Pio Pastore

Abstract: Obtained from CrossRef

Surv-TCAV is introduced as a concept-based interpretability framework for gradient-boosted accelerated failure time (AFT) survival models on clinical tabular data. Using the Primary Biliary Cholangitis dataset (historical pbc; PBC-276 complete cases over 17 covariates), an XGBoost model with AFT objective is trained and assessed for discrimination with Harrell’s concordance index (C-index) and for calibration with the integrated Brier score (IBS) using inverse-probability-of-censoring weights. Interpretability is delivered at two levels: feature attributions via TreeSHAP and directional concept sensitivity via Surv-TCAV, which quantifies how clinician-defined concepts perturb the predicted log-time location parameter μ. Across 25 independent 80/20 train/validation splits with fixed boosting rounds (no early stopping), the boosted AFT attains C-index = 0.835 ± 0.039 and IBS = 0.357 ± 0.018. A one-sample t-test of split-wise C-indices against 0.85 yields t = −1.863 (df = 24), two-sided p = 0.075, thus statistically indistinguishable from strong mid–0.8 discrimination at α = 0.05. Surv-TCAV reveals clinically coherent negative directional effects for low albumin, older age, and clinical complications, with more modest but still negative effects for cholestasis and coagulopathy. Feature-level attributions further suggest a small residual protective association for female sex after adjustment. The protocol delivers concept-based interpretability for boosted survival models on structured clinical data with full reproducibility.

CODECHECK details

Certificate identifier: 2025-028

Codechecker name: Daniel Nüst

Time of check: 2025-11-04 14:36:30

Repository: https://github.com/codecheckers/surv-tcav-pbc

Full certificate: https://doi.org/10.5281/zenodo.17533508

Summary:

This code was straightforward to check and the reproduction is partially successful. Figure 1, 2, and 3 could be recreated without errors in the code but in part with differences beyond the variations due to re-computation. Table 2 could be reproduced within computational variation, but Table 3 could not be linked to any of the output files.

https://codecheck.org.uk/ | GitHub codecheckers

© team@cdchck.science | Published under CC BY-SA 4.0

DOI of Zenodo Deposit

CODECHECK is a process for independent execution of computations underlying scholarly research articles.