Machine learning prediction of hospitalization outcomes in critically Ill emergency department patients transported by ambulance: A Retrospective Single Center Cohort Study

Comprehensive Summary

Durmuş and Eke Kurt investigated whether machine-learning models can predict the need for hospitalization among critically ill emergency department (ED) patients transported by ambulance. In this retrospective single-center cohort study, demographic, clinical and early laboratory parameters in 2,338 adult patients were used to develop machine learning models, These included logistic regression, random forest, and gradient boosting models, and were evaluated using five-fold cross validation. The primary outcome was defined as admission vs. no admission. Model performance was assessed using receiver operating characteristic area under the curve (ROC AUC), accuracy, sensitivity, and specificity. The study found that 37.0% of patients required hospitalization, and the random forest model showed the strongest performance (cross-validation ROC AUC 0.850; test ROC AUC 0.776). Important predictors included troponin, altered mental status, lactate, age (including ≥65 years), creatinine, leukocyte count, and pH. Meanwhile, the random forest vs gradient boosting ROC AUC difference was not statistically significant (mean difference 0.008; p=0.052).

Outcomes and Implications

The study demonstrates that machine learning models have potential in supporting early clinical decision making in ED populations. Notably, the random forest model demonstrated clinically relevant performance in predicting hospitalization using early clinical and laboratory data, positioning such models as early ED decision-support tools rather than prehospital ones. The study is limited, however, by its retrospective single-center design, missing data requiring imputation, and potential selection bias stemming from clinician-directed lab ordering. Prospective multicenter trials are warranted to validate these findings.