Table 2 Study details, model development and performance characteristics of studies predicting mortality following moderate to severe TBI
First author and year of publication | TBI definition | Time of mortality assessment | Sample size (adult, mixed, or pediatric) | Model architecture(s) | AI performance (top model performance reported) | Non-AI performance (top model; if existing nomogram or human experts included, all reported) |
|---|---|---|---|---|---|---|
Yin 2024 | GCS ≤ 12, admitted <24 h from injury; patients that died within 24 h excluded) | In-hospital (at discharge) | 169 (adult) | Transductive support vector machine (SVM), light gradient boosting machine (GBM), feature tokenizer (FT) transformer, self-supervised learning | Self-supervised learning: AUC: 0.994 Accuracy: 99.4% Sens: 100% Spec: 99.4% | Logistic Regression (LR): AUC: 0.762 Accuracy: 88.2% Sens: 54.5% Spec: 90.5% |
Zhang 2023 | GCS ≤ 12, admitted to ICU; excluded patients that died within 24 h | In-hospital (at discharge) | 482 (adult) | eXtreme Gradient Boosting (XGBoost), light GBM, and FT transformer | Light GBM: AUC: 0.953 Accuracy: 94.8% Sens: 86.8% Spec: 95.8% | LR: AUC: 0.813 Accuracy: 86.7% Sens: 58.5% Spec: 90.2% |
Matsuo 2020 | Blunt injury, abnormal CT head requiring admission (not strict msTBI, but median GCS 10) | In-hospital (at discharge) | 232 (mixed; age >10 included) | SVM (multiple kernels), Gaussian/ multinomial naïve Bayes, ridge regression, extra trees, random forest, least absolute shrinkage and selection operator (LASSO) regression, GBM, decision tree | Random forest: AUC: 0.818 Accuracy: 95.5% Sens: 63.6% Spec: 100% | Not included |
Jung 2023 | Undergoing ICP monitoring in ICU | 6 months | 166 (adult) | LightGBM | AUC: 0.893 | Not included |
Gravesteijn 2020 | GCS ≤ 12 | 6 months | 12397 (adult) | SVM, artificial neural network (ANN), random forest and GBM | GBM: AUC: 0.83 | LR and LASSO regression: AUC: 0.82 |
Hanko 2021 | Undergoing primary decompressive craniectomy | 6 months | 40 (adult) | Random forest | AUC: 0.811 | Not included |
Lu 2015 | Moderate to severe TBI surviving ≥14 days | 6 months | 115 (adult) | ANN, naïve Bayes, and decision tree | Naïve Bayes: AUC: 0.901 Sens: 81.2% Spec: 90.7% | LR: AUC: 0.873 Sens: 68.4% Spec: 91.0% |
Pease 2022 | GCS ≤ 8 admitted to hospital | 6 months | 757 (adult) | Convolution neural network (CNN) | Fusion model (clinical and imaging model) AUC: 0.80 Accuracy: 86% Sens: 78% Spec: Varied to match each human (mean 81.7%) | International Mission for Prognosis and Analysis of Clinical Trials in TBI (IMPACT) score: AUC: 0.83 Expert humans: Accuracy: 71.3% Sens: 55% Spec: 81.7% |
Arefan 2023 | Blunt injury with post-resuscitation GCS ≤ 8 | 24 months | 395 (with complete 24-month outcome data) | SVM, ANN, decision tree, naïve Bayes | SVM: AUC: 0.85 Accuracy: 79% | LR: AUC: 0.85 Accuracy: 79% |
Daley 2022 | GCS ≤ 8 and Abbreviated injury severity score (AIS) ≥ 4 admitted to ICU | 1 month (in-hospital) | 196 (pediatric) | Random forest | AUC: 0.90 | Not included |
Cui 2021 | Undergoing decompressive craniectomy, surviving hospitalization | 12 months | 230 (adult) | Random forest and LR with synthetic minority oversampling for training data | Random forest: AUC: 0.83 Accuracy: 81% Sens: 83.3% Spec: 80% | LR: AUC: 0.765 Accuracy: 67.2% Sens: 100% Spec: 52.5% |
Wang 2022 | GCS ≤ 12 arriving <6 h from injury | In-hospital | 368 (adult) | XGBoost | AUC: 0.955 Accuracy: 95.5% Sens: 94.5% Spec: 96.4% | LR: AUC: 0.805 Accuracy: 70.3% Sens: 73.8% Spec: 75.4% |
Feng 2019 | Severe TBI without further specification | Not well defined (timing poorly defined) | 117 (adult) | 22 models total including SVM, KNN, tree-based models, GBM | Linear SVM: AUC: 0.94 Accuracy: 92.5% Sens: 97% Spec: 58% | LR: AUC: 0.84 Accuracy: 87.7% Sens: 91% Spec: 65% |
Wu 2023 | Clinical indication of TBI undergoing CT with GCS ≤ 8 | In-hospital | 3917 (adult) | XGBoost, SVM | XGBoost (26-feature model): AUC: 0.87 Sens: 78% Spec: 82% | LASSO LR: AUC: 0.86 Sens: 72% Spec: 85% IMPACT and Corticoid Randomization after Significant Head Injury (CRASH) models no AUC/clinical utility included |
Raj 2019 | ICP monitoring for ≥24 h | 30 days | 472 (adult) | Dynamic LR with rolling time windows (coefficient C tuned as a hyperparameter) | Dynamic LR features selected were ICP, mean arterial pressure (MAP), cerebral perfusion pressure (CPP), GCS: AUC: 0.72 day 1; 0.84 day 5 Sens: 79% Spec: 94% | LR using IMPACT score variables: AUC: 0.78 Sens: 100% Spec: 50% |
vanderPloeg 2016 | Moderate to severe TBI not further specified | 6 months | 11026 (adult) | Classification and regression trees, random forests, SVM and ANN | Random forest with IMPACT core, laboratory and imaging features: AUC: 0.735 | LR using IMPACT score variables: AUC: 0.757 |
Mekkodathil 2023 | Blunt TBI direct to trauma center; not strict GCS definition but >90% msTBI | In-hospital | 922 (mixed) | SVM, random forest, XGBoost | SVM: AUC 0.86 Accuracy: 81% Sens: 70% Spec: 81% | LR: AUC: 0.84 Accuracy: 80% Sens: 71% Spec: 80% |
Yao 2020 | GCS 4-12 according to the Progesterone for Traumatic Brain Injury, Experimental Clinical Treatment (PROTECT) trial72 criteria | 6 months | 828 (adult) | CNN for automated hematoma segmentation, random forest for outcome prediction | Optimal model used IMPACT variables without computed tomography (CT) features & automated CT volume: AUC: 0.853 Sens: 70% Spec: 87.2% | LR using IMPACT score variables: AUC: 0.795 Sens: 71.2% Spec: 74.7% |
Lang 1997 | GCS ≤ 8 for 24 h post admission | 6 months | 1066 (adult) | ANN | Accuracy: 85.8% Sens: 90.6% Spec: 77.1% | LR: Accuracy: 83.1% Sens: 87.7% Spec: 75% |
Fonseca 2022 | GCS ≤ 12 admitted to ICU for ≥24 h or had neurosurgical intervention | In-hospital | 300 (pediatric) | XGBoost, ANN, k-nearest neighbors (KNN), random forest | XGBoost: AUC: 0.91 | Not included |
Raj 2022 | Undergoing ICP monitoring | 1 month | 1324 (adult) | Dynamic LR with rolling time windows (coefficient C tuned as a hyperparameter) | External Dataset 1: AUC 0.66 first 24 h; AUC 0.79 at 120 h Accuracy 120 h: 90.3% Sens 24/120 h: 93%/93% Spec 24/120 h: 79.7%/ 97.6% External dataset 2: 0.67 first 24 h; AUC 0.73 at 120 h Accuracy 120 h: 75.9% Sens 120 h: 77%/ 77% Spec 24/ 120 h: 89.7%/ 98.9% | Not included |
Cao 2023 | Head region AIS ≥ 3 and AIS ≤ 1 in other regions (excluding AIS = 6) | In-hospital | 545,388 (adult) | XGBoost Cox proportional hazard regression | AUC: 0.896 Mean time-dependent AUC: 0.713 | Not included |
Stein 2012 | Severe TBI undergoing ICP monitoring, excluding severe polytrauma | In-hospital | 52 (adult) | Compound covariate predictor, linear discriminant analysis, KNN classifiers, and SVM (differing kernels) | Top model KNN: Accuracy: 88% | Not included |
Bark 2024 | TBI all severities admitted to ICU (83% model development dataset moderate to severe TBI) | 6–12 months | 1808 (adult) | Random forest & ANN | Top model random forest with IMPACT features for External Dataset 1: AUC: 0.86 Top model ANN with IMPACT features for External Dataset 2: AUC: 0.69 | Multivariable proportional odds LR with IMPACT features: External dataset 1: AUC: 0.88 External dataset 2: AUC: 0.63 |
