Beyond predictive performance: Interpretability challenges and feature importance bias in XGBoost-based readmission models

Song et al. report a machine-learning framework based on the eXtreme Gradient Boosting (XGBoost) algorithm for predicting 1-year unplanned readmissions among elderly patients with coronary heart disease (CHD). This commentary examines critical limitations in the interpretability and methodological robustness of such models. Extensive prior work has demonstrated that tree‑based algorithms can exhibit structural biases in feature‑importance estimates, particularly when predictors display substantial collinearity or heterogeneous measurement scales. SHAP explanations, by inheriting these model‑embedded biases, may overstate the relevance of variables whose prominence arises from algorithmic artifacts rather than clinically coherent patterns. To enhance reliability, model‑agnostic validation, non‑parametric association analyses, and unsupervised strategies that mitigate multicollinearity should complement predictive modeling efforts. Strengthening interpretability frameworks is essential to ensure that machine‑learning-derived insights meaningfully inform cardiovascular care and support reproducible clinical translation.