Machine learning-based on model for explain risk of 24-hour death in critically ill patients in the prehospital setting: A retrospective cohort study

This study aimed to develop and validate a machine learning-based model for predicting 24-hour mortality in critically ill patients using prehospital and admission clinical data. We conducted a retrospective cohort study leveraging data from the prehospital emergency electronic medical record, in-hospital triage, and hospital information systems of a tertiary hospital in Changsha between August 2023 and April 2025. A total of 892 adult patients classified as critically ill were included. Nine machine learning algorithms were trained to predict 24-hour mortality, and model performance was assessed using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, and F1 score. SHapley Additive exPlanations (SHAP) analysis was employed to interpret feature contributions. Among the nine algorithms, the Random Forest (RF) model exhibited the most stable and robust performance. Using nine selected features-prehospital heart rate, prehospital and admission systolic and diastolic blood pressure, prehospital and admission oxygen saturation, admission respiratory rate, and level of consciousness, the RF model achieved an AUC of 0.985(95%CI:0.976–0.993) in the training set and 0.863 (95%CI:0.766–0.961) in the testing set, demonstrating high accuracy and potential clinical applicability. SHAP analysis revealed that prehospital heart rate, admission respiratory rate, and blood pressure are the strongest predictors of mortality. Finally, the model was deployed as an interactive web-based tool for real-time clinical application. In summary, this study developed a simple, interpretable, and accurate machine learning model for predicting 24-hour mortality in critically ill prehospital patients. The RF-based model can be intended as an exploratory, hypothesis-generating tool and should supplement, not replace, clinical judgment. Further validation in larger, multi-center prospective cohorts with higher event rates is essential to confirm the robustness and real-world applicability of our findings.