Table 3. Main findings, validation, best models, and clinical relevance for each study.

AUROC, area under the receiver operating characteristic curve; MESS, mangled extremity severity score; ACS, American College of Surgeons; NSQIP, National Surgical Quality Improvement Program; VSGNE, Vascular Study Group of New England; CV, cross-validation; TEG, thromboelastography; AUC, area under the curve; RF, random forest; NB, Naïve Bayes; SVM, support vector machine; ANN, artificial neural network; LR, logistic regression.

Study	Main findings	Calibration/Validation	Best model	Clinical utility/Comparison
Li et al. (2024) [1]	XGBoost AUROC=0.93 (0.92-0.94), LR AUROC=0.72 (0.70-0.74). ~9.0% event rate.	Brier score 0.09 (good calibration). 10-fold CV, 70/30 split.	XGBoost	Outperformed logistic regression (0.72 AUROC). Accurate 30-day outcome prediction; authors note need for prospective validation.
Perkins et al. (2020) [2]	BN model AUROC=0.95 (0.92-0.98) for predicting failed revascularization (vs MESS AUROC 0.74). Accuracy not reported explicitly (high).	Calibration slope 1.96 (dev), 1.72 (val); Brier 0.05. Validation: 10-fold CV + external UK cohort.	Bayesian network	Outperformed Mangled Extremity Severity Score (0.95 vs 0.74 AUROC ). Provides individualized limb-salvage risk to inform decision-making in trauma.
Ghandour et al. (2025) [3]	Logistic regression (with baseline+TEG data) AUC=0.76; accuracy=0.70; sensitivity=0.68; specificity=0.71. XGBoost and tree had similar AUCs (~0.72-0.76).	Five-fold CV with 70/30 split. Logistic model had best combined discrimination and calibration.	Logistic (with TEG)	Combining patient factors and thromboelastography improved prediction of 1-year post-revascularization thrombosis. May help identify high-risk patients for tailored anticoagulation.
Li et al. (2024)[4]	XGBoost AUROC=0.93 (0.92-0.94) (versus RF 0.92, NB 0.87, SVM 0.85, ANN 0.80, LR 0.63 ). Overall accuracy ~0.86.	Brier score 0.08 (good calibration). 10-fold cross-validation (CV) with 70/30 train-test split.	XGBoost	Demonstrated strong discrimination where no clinical risk tool exists. Potential to improve risk stratification beyond traditional ACS-NSQIP/VSGNE scores.
Li et al. (2024) [5]	XGBoost AUROC=0.94 (0.93-0.95); accuracy=0.86; sensitivity=0.87; specificity=0.85. LR AUROC=0.67.	10-fold CV (70/30 train-test). XGBoost performance remained high post-op (AUROC up to 0.98).	XGBoost	Significantly better than logistic. High predictive accuracy could guide perioperative risk mitigation strategies.