Prehospital Trauma Scoring Systems for Discriminating Recorded Secondary Transfer in Multiply Injured Patients: A Comparative Study with Machine Learning
Journal of Clinical Medicine · 2026-11-12 · 15 (22) · 8838
DOI: 10.3390/jcm15228838 ↗Abstract
Background/Objectives: We compared eight prehospital trauma scores for discriminating recorded secondary interhospital transfer in multiply injured patients, identified factors associated with transfer, and compared a random forest with logistic regression. Methods: Retrospective cohort of Kocaeli emergency medical services records (January 2024–June 2025). Scores were evaluated with receiver operating characteristic analysis; logistic regression was adjusted for the initial receiving facility. Models were trained on a patient-level 70% sample and evaluated on the remaining 30% and by repeated grouped cross-validation. Results: Of 1194 patients, 561 (47.0%) underwent secondary transfer. The Circulation, Respiration, Abdomen, Motor, Speech (CRAMS) score had the highest discrimination (area under the curve [AUC] 0.676; 95% confidence interval [CI] 0.647–0.705). Pediatric (odds ratio [OR] 4.46; 95% CI 2.82–7.07) and elderly age (OR 2.44; 1.71–3.48) and thoracic injury (OR 1.68; 1.23–2.29) were independently associated with transfer. In the held-out set (n = 356), the random forest AUC was 0.836 (0.795–0.878) versus 0.765 for logistic regression with the same predictors (p < 0.001); its calibration slope was 1.34, indicating that predicted probabilities were compressed toward the average risk. Conclusions: Individual scores showed limited discrimination for this system-level outcome. The random forest discriminated better than regression with the same information but requires recalibration and external validation before clinical use.
EUVIMED