Development of a new tool for prediction of hospital length of stay and intensive care needs in trauma patients using Machine Learning.

Abstract

Trauma is a major global health burden leading to significant morbidity, disability, and mortality. Predictive models in trauma care traditionally focus on mortality, but early predictions of hospital length of stay (LOS) and intensive care unit (ICU) needs could greatly enhance hospital planning and resource allocation. Machine learning (ML) offers new possibilities for developing prediction tools for these outcomes but remain underexplored in large, unselected trauma populations. To develop and validate machine learning-based models for early prediction of hospital length of stay and ICU admission among severely injured trauma patients using a large patient cohort from a national trauma registry. Patient data from 9056 adult severely injured trauma patients (NISS >15) registered in the Swedish trauma registry SweTrau between 2015 and 2019 were analyzed. Only variables available at hospital arrival were used as predictors. Outcomes were LOS (1-2, 3-9, or ≥10 days) and ICU admission (yes/no). Patients from 2015 to 2018 (n = 6706) were used for training Generalized Linear Model (GLM), Random Forest (RF), and Extreme Gradient Boosting (XGB) models, and patients from 2019 (n = 2350) were used for temporal internal-external validation. Model performance was assessed with ROC curves, calibration curves and DCA. The XGB models consistently outperformed GLM and RF models for all outcomes. For estimation of ICU admission, the XGB model achieved an AUC of 0.85 (95% CI: 0.84-0.87). For estimations of LOS, the XGB model achieved "one-vs- all" AUCs of 0.69, 0.64, and 0.71 for the three LOS categories, respectively. A clinical prediction tool based on the best-performing models was created and is available online (https://hipfx.shinyapps.io/traumaadvisorapp/). Machine learning models trained on national trauma registry data demonstrated strong performance in predicting ICU admission and moderate accuracy in categorizing hospital length of stay. The XGB model showed the highest overall predictive power and may serve as a useful tool to support early triage, guide clinical decision-making, and optimize resource allocation in trauma care settings.