An interpretable early intensive care unit model for predicting in-hospital mortality in trauma patients combined with rib fractures: development and external validation.

Background: Patients with rib fractures admitted to the intensive care unit (ICU) are clinically heterogeneous, and early identification of those at high risk of in-hospital death remains challenging using anatomy-based descriptors alone. We aimed to develop an objective, readily implementable prediction model using routinely available variables within the first 24 hours of ICU admission and to test its generalizability in an independent institutional cohort. Methods: We retrospectively analyzed ICU patients combined with rib fractures from the Medical Information Mart for Intensive Care IV (MIMIC-IV) and used an independent institutional cohort for external validation. Routinely available variables recorded within the first 24 hours of ICU admission were screened to identify the most informative predictors, after which several prediction models were developed and compared. To avoid overly optimistic estimates and information leakage, the derivation cohort was first divided into a training set and an internal validation set, and all outcome-informed feature-selection and model-development procedures were restricted to the training set. Candidate predictors retained by the intersection of least absolute shrinkage and selection operator (LASSO) regression and support vector machine-recursive feature elimination (SVM-RFE) were entered directly into multivariable logistic regression (LR) for final predictor determination. Class balancing was performed only within the training process. In addition, as supplementary analyses, we explored ensemble strategies that combined the predictions of the base models and performed local institutional analyses to examine whether available thoracic trauma and overall injury severity measures materially altered model performance. Model performance was evaluated in terms of discrimination, calibration, and clinical utility. Results: The final multivariable model retained four independent predictors: Simplified Acute Physiology Score II (SAPS II), systolic blood pressure (SBP), platelet count, and any vasopressor use. In internal validation, LR achieved the highest discrimination [area under the curve (AUC) =0.863], with sensitivity (0.800) and specificity (0.658) at the cross-validated operating threshold. In external validation, the LR model maintained good discrimination (AUC =0.873) with acceptable calibration and clinically meaningful net benefit on decision curve analysis. Supplementary ensemble analyses did not outperform LR in internal validation. In supplementary institutional analyses, Thoracic Trauma Severity Score (TTSS) alone and RibScore alone underperformed the early physiologic benchmark, whereas adding TTSS, RibScore, or Injury Severity Score (ISS) produced only limited incremental changes. Conclusions: A parsimonious, interpretable four-variable model using routinely available early ICU data provided stable internal performance and preserved discrimination in an independent external cohort. This tool may support early mortality risk stratification and triage in ICU trauma patients combined with rib fractures, with future work needed for broader multicenter validation and prospective impact evaluation.