Table 3.
Summary of different AI-based methods for TBI prognostication.
| Ref. | ML/DL | Algorithm/Method | Dataset Size | TBI-Related Clinical Assessment |
Performance | Main Contribution |
|---|---|---|---|---|---|---|
| [32] | ML | Linear Regression, Binary logistic regression | 106 | 30-day mortality | AUC: Hematoma shape: 0.692 Hematoma size: 0.715–0.786 ICH score: 0.877–0.882 GCS: 0.912–0.922 |
Hematoma shape and size, ICH score, GCS score, age, IVH, presence of infratentorial location were used to estimate 30-day mortality |
| [43] | DL | 2D U-net based CNN, 3D U-net based CNN | 144 | Hematoma segmentation, volume estimation, GOSE prediction | Segmentation Dice: 0.697 Volume estimation correlation coefficient: 0.966 6-month GOSE prediction AUC: 0.85 |
Proposing a novel Multi-view CNN with a mixed loss for hematoma segmentation, quantification, and 6 months mortality prediction |
| [102] | DL/ML | ANN, LR | 785 | early mortality | LR Accuracy: 0.87 AUROC: 0.905 ANN ACC: 0.809 AUROC: 0.875 |
Trauma registry data including Injury, CT findings and demographic characteristics |
| [103] | DL/ML | LR, 22 ML models * | 117 | Survival prediction | AUC: LR: 0.83 ML models: 0.30–0.94 |
Cubic SVM, Quadratic SVM and Linear SVM outperformed LR |
| [60] | ML | RF classifier | 828 | 6 months mortality prediction | AUC: 0.853 AUPRC: 0.559 |
Integrating volumetric characteristics and shape features extracted from the proposed model with IMPACT without CT features to predict six-months mortality |
| [106] | DL/ML | LR, lasso regression, and ridge regression, SVM, RF, GBM, ANN |
11022 (IMPACT-II), 1554 (CENTER-TBI) | GOS < 3, GOSE < 5 |
Mean AUC (external validation) Mortality: 0.82 Unfavorable outcome: 0.77 |
Motor GCS score, CT class, SDH, EDH, hypoxia, hypotension, demographic, and some laboratory test data were used to compare various model performance in predicting patient outcome. |
| [104] | ML | XGB, LR |
368 TBI patients with GSC<13 | Early mortality | XGB: Acc: 0.955 AUROC: 0.955 LR: Acc: 0.70 AUROC: 0.805 |
Data from electronic medical record system Including laboratory test data, injury, and demographic information, CT finding |
| [105] | ML | DNN, SNN, LR-net |
2164 | GOS 1–3 vs. GOS 4–5 |
AUROC DNN: 0.941 SNN: 0.931 LRnet: 0.919 |
Using demographic information, GCS, injury mechanism, heart rate, blood pressure and other clinical data |
| [108] | ML | XGB classifier, SHAP values selection | 831 (ProTECT III data set) | GOSE 1–4 vs. GOSE 5–8 |
AUC: 0.80 Acc: 0.74 F1-score: 0.70 |
Developing an intelligible prognostic model using 2 rounds of variable selection by SHAP values as well as clinical domain knowledge |
| [109] | DL | TFNN vs. RF, XGB, SVM |
833 (ProTECT III data set) | GOSE 1–4 vs. GOSE 5–8 |
AUC TFNN: 0.79 RF: 0.80 SVM: 0.79 XGB: 0.74 |
Developing human interpretable neural network model based on tropical geometry to predict GOSE 6 months after hospitalization |
* Quadratic discriminant, Linear discriminant, Fine KNN, Subspace KNN, Coarse KNN, Coarse Gaussian SVM, Medium Gaussian SVM, Fine Gaussian SVM, Cubic KNN, Boosted trees, Subspace discriminant, Simple tree, Medium tree, Complex tree, Cosine KNN, Medium KNN, RUSBoosted trees, Bagged trees, Weighted KNN, Cubic SVM, Quadratic SVM, Linear SVM.