An explainable machine learning model for comorbidity risk stratification in patients with fractures admitted to the intensive care unit: a multicenter study.

BACKGROUND: Among traumatic-fracture patients admitted to intensive care units (ICUs), those with substantial chronic comorbidities recover more slowly and die more often than their counterparts without such conditions. The age-adjusted Charlson Comorbidity Index (aCCI) quantifies this burden, yet clinicians still lack a tool that can identify-at the point of ICU admission-which fracture patients are likely to have a high aCCI. To fill this gap, we used a large electronic health-record repository to develop and externally validate an interpretable machine-learning model that predicts severe comorbidity burden in this population. METHODS: We extracted 3 763 adult fracture cases from MIMIC-IV (2008-2019) and split them 3:1 into training and internal validation sets. High comorbidity (aCCI ≥ 7) was defined as the optimal cut-off derived from one-year survival analysis. Nine key predictors emerged from the intersection of LASSO, SVM-RFE, and random-forest importance. Eleven candidate algorithms underwent grid-search hyperparameter tuning with 10-fold cross-validation, and their performance was compared to identify the optimal model, while SHAP clarified model logic. External validation in two Chinese tertiary centres (n = 558) confirmed generalisability, and the final model was deployed as a bedside Shiny calculator. RESULTS: XGBoost achieved the best internal discrimination (AUROC = 0.84; AUPRC = 0.76) and exhibited excellent calibration and net benefit across clinically relevant thresholds. In external validation, AUROC values were 0.88 (Hainan) and 0.83 (Guangdong). The interactive calculator delivers patient-specific risk explanations in real time. CONCLUSIONS: An XGBoost-based, SHAP-interpretable model accurately predicts high aCCI in ICU fracture patients and generalises across institutions. The readily accessible web tool can help clinicians identify high-risk individuals early, personalise management, and allocate resources more efficiently.