Multi-model comparison and nomogram development for predicting in-hospital mortality in Trauma patients based on NLR and other indicators.

Abstract

Objective: This study aimed to systematically compare the predictive performance of logistic regression (LR), random forest (RF), and extreme gradient boosting (XGBoost) for predicting in-hospital mortality among trauma patients, evaluate the prognostic value of key indicators, including the neutrophil-to-lymphocyte ratio (NLR), and develop a visual clinical nomogram.

Methods: A retrospective analysis was conducted on 318 severe trauma patients [Injury Severity Score (ISS) ≥ 16] admitted between January 2022 and October 2024 as the derivation cohort; an independent temporal validation cohort of 161 patients admitted between January 2025 and May 2026 from the same institution was also included. LR, RF, and XGBoost models were trained and evaluated for AUC, sensitivity, specificity, and calibration. LASSO regression was used for variable selection, and a multivariable LR model was constructed to develop a nomogram.

Results: On the internal test set, RF achieved the highest AUC (0.919) with a sensitivity of 0.963. LR had the highest specificity (0.868), with a calibration plot indicating good agreement. In the temporal validation cohort, LR demonstrated robust performance with an AUC of 0.906. Multivariable LR identified age [odds ratios (OR) = 1.071], ISS (OR = 1.234), and shock index (OR = 3.963) as independent risk factors, while NLR was a paradoxical protective factor (OR = 0.872). Variable importance analysis ranked NLR as the top contributor. A nomogram incorporating these variables demonstrated favorable discrimination and clinical utility, supported by DCA.

Conclusion: The developed nomogram integrates key predictors into a visual tool for rapid mortality risk assessment in trauma patients, thereby facilitating early risk stratification and optimized resource allocation. The RF model is superior for identifying high-risk patients (high sensitivity), whereas the LR model excels at ruling out low-risk patients (high specificity) and demonstrates better probability calibration. NLR is a significant predictor of early post-traumatic mortality, although its direction of effect warrants further investigation.