Predicting Hospital Length of Stay in Orthopedic Trauma Patients Using Fracture-Specific Machine Learning Models: A Multicenter Retrospective Prediction-Modeling Study.

BACKGROUND AND AIMS: Hospital length of stay (LOS) is a key indicator for resource allocation, bed management, and discharge planning in orthopedic trauma care. Given the substantial heterogeneity of LOS across fracture types and the limitations of conventional approaches, this study aimed to develop and validate fracture-specific machine learning models for predicting short vs. long hospital stay. METHODS: In this multicenter retrospective prediction-modeling study, routinely collected hospital information system data from three trauma centers were analyzed. From 326,323 diagnosis-level records, a cleaned visit-level cohort was constructed, including 8935 (S42), 16,906 (S52), 8091 (S72), and 15,293 (S82) admissions after applying eligibility criteria and fracture-specific outlier removal. Predictors included age, sex, insurance status, hospital, COVID-19 period, and season. LOS was dichotomized using fracture-specific medians. Logistic regression and XGBoost models were developed using stratified 80/20 splits and further evaluated using temporal validation (earlier vs. later years). Performance was assessed by AUC (with 95% bootstrap confidence intervals), accuracy, sensitivity, positive predictive value (PPV), F1-score, Brier score, calibration (intercept and slope), and decision-curve analysis. RESULTS: LOS distributions were right-skewed and differed substantially across fracture groups, with the longest median LOS in S72 (6 days) and the shortest in S52 (1 day). In internal validation, XGBoost showed superior discrimination (AUC: 0.703 [S52], 0.673 [S42], 0.659 [S82], and 0.618 [S72]) compared with logistic regression. Calibration was acceptable (slopes near 1), and Brier scores ranged from 0.218 to 0.238. Temporal validation demonstrated modest performance decline (AUC range: 0.602-0.677), indicating limited transportability. Decision-curve analysis showed consistent net benefit across clinically relevant thresholds. Age was the dominant predictor in most groups, while hospital-related variables were prominent in S52. CONCLUSION: Fracture-specific modeling provides a more appropriate framework for LOS prediction than pooled approaches. XGBoost demonstrated consistent improvements over logistic regression with acceptable discrimination, calibration, and clinical utility. However, moderate performance and temporal degradation highlight the need for richer clinical predictors and external validation before routine implementation.