Hypertrophic scarring (HS) following severe burns remains a persistent rehabilitative challenge, yet traditional linear prediction models fail to capture the non-linear pathophysiological complexity of fibrosis. This study aimed to engineer an interpretable machine learning framework to stratify HS risk and elucidate its driving mechanisms. Utilizing a retrospective cohort of 520 severe burn patients, we benchmarked four machine learning algorithms, selecting Extreme Gradient Boosting (XGBoost) for model construction. The SHapley Additive exPlanations (SHAP) framework was integrated to decode algorithmic decision-making, specifically analyzing feature contributions and interaction effects. We benchmarked four machine learning algorithms, selecting Extreme Gradient Boosting (XGBoost) for model construction. The XGBoost model demonstrated superior discrimination (AUC: 0.905, 95% CI: 0.865-0.945) and calibration (Brier score: 0.112) compared to conventional logistic regression. Decision Curve Analysis confirmed the model's incremental clinical net benefit (range: 0.01-0.85). We successfully developed an Intelligent Clinical Decision Support System (iCDSS) that translates complex algorithmic computations into visualized, individualized risk attribution profiles. This framework refines the epidemiological understanding of inflammatory drivers in scarring and offers a promising approach for shifting from empirical prognostication to data-driven precision prevention, pending further external validation.
Tian et al. (Mon,) studied this question.