Table 2.
Summary of applied machine learning programs and corresponding receiver operating characteristic analyzes.
| Machine Learning Tool | Number of patients | Application | Sensitivity | Specificity | AUC | G | Reference |
|---|---|---|---|---|---|---|---|
| LR based ML | 472 | Consecutive TBI in adults | 0.810 | (226) | |||
| 0.840 | |||||||
| LR | 117 | Severe TBI in adults | 0.830 | (227) | |||
| ML average (outliers removed) | 0.880 | ||||||
| 14 input SVM | 94 | Pediatric severe TBI | 63% | 100% | NA | 37% | (232) |
| 3 input SVM | 80% | 99% | NA | 21% | |||
| 14 input LR | 75% | 99% | 0.900 | 26% | |||
| 3 input LR | 71% | 99% | 0.830 | 30% | |||
| RF | 232 | Non-penetrating TBI with abnormal CT scan | 97.20% | 49.20% | 0.860 | 53.60% | |
| Gaussian Naive Bayesian | 68.70% | 82.80% | 0.842 | 48.50% | (233) | ||
| Gradient Boosting Model | 93.70% | 62.80% | 0.857 | 43.50% | |||
| SVM morbidity | 94.50% | 58.60% | 0.894 | 46.90% | |||
| SVM mortality | 77.60% | 97.00% | 0.942 | 25.40% | |||
| LR | 325 | Isolated moderate and severe TBI in adults | 59.38% | 93.54% | 0.942 | 47.08% | (230) |
| SVM | 65.63% | 95.22% | 0.935 | 39.15% | |||
| DT | 44% | 98% | 0.872 | 58% | |||
| NB | 59% | 86.15% | 0.908 | 54% | |||
| ANN | 84.38% | 92.83% | 0.968 | 22.79% | |||
| RF | 51 | Acute kidney injury in burned and unburned adults | 82% | 68% | 0.750 | 50% | (236) |
| k-NN | 91% | 82% | 0.870 | 27% | |||
| DNN & LR | 91% | 93% | 0.920 | 16% | |||
| LR | 195 | Moderate to severe pediatric TBI | 82.10% | 92.30% | 0.930 | 25.60% | (224) |
| ANN | 94.90% | 97.40% | 0.980 | 7.70% | |||
| ESS | 564 | Chest pain patients | 78.90% | 76.50% | 0.837 | 44.60% | (225) |
| DIST | 63.20% | 82.90% | 0.720 | 53.90% | |||
| MEWS | 42.10% | 78.50% | 0.672 | 79.40% | |||
| TIMI | 78.90% | 36.70% | 0.621 | 84.40% | |||
| LR | 674 | Prolonged mechanical ventilation following TBI | 80% | 68% | 0.830 | 52% | PMV > 7 days (223) |
| SVM | 83% | 67% | 0.800 | 50% | |||
| RF | 76% | 70% | 0.770 | 54% | |||
| ANN | 77% | 60% | 0.780 | 63% | |||
| C.5 DT | 70% | 61% | 0.650 | 69% | |||
| LR | 69% | 79% | 0.820 | 52% | |||
| SVM | 643 | Prolonged mechanical ventilation following TBI | 76% | 82% | 0.840 | 42% | Set B PMV > 10 days (223) |
| RF | 81% | 71% | 0.800 | 48% | |||
| ANN | 76% | 77% | 0.770 | 47% | |||
| C.5 DT | 65% | 75% | 0.770 | 60% | |||
| LR | 29% | 90% | 0.750 | 81% | |||
| SVM | 622 | Prolonged mechanical ventilation following TBI | 29% | 91% | 0.740 | 80% | Set C PMV > 14 days (223) |
| RF | 46% | 80% | 0.710 | 74% | |||
| ANN | 27.00% | 94% | 0.720 | 79.00% | |||
| C.5 DT | 25% | 88% | 0.650 | 87% | |||
| rsFNC | 100 | Mild TBI imaging | 89.40% | 78.80% | 0.841 | 31.80% | (231) |
| FA | 76.60% | 74.50% | 0.755 | 48.90% | |||
| FA + rsFNC | 76.60% | 72.30% | 0.745 | 51.10% | |||
| FA | 70.20% | 61.70% | 0.660 | 68.10% | |||
| 61.70% | 66% | 0.648 | 72.30% | ||||
| Marshall CT | 565 | Pediatric TBI CT scans | 0.663 | 6-months unfavorable outcome (GOS < or = 3) (235) |
|||
| Rotterdam CT | 0.748 | ||||||
| Helsinki CT | 0.717 | ||||||
| GCS Score | 0.855 | ||||||
| Marshall CT | 565 | Pediatric TBI CT scans | 0.781 | 6-month mortality (235) |
|||
| Rotterdam CT | 0.838 | ||||||
| Helsinki CT | 0.814 | ||||||
| GCS score | 0.920 | ||||||
| Linear SVM | 632 | Blood plasma predictors in mild TBI college athlete patients | 81.70% | 71.50% | 0.830 | 46.80% | (237) |
| LASSO | 77.80% | 68.60% | 0.811 | 53.60% | |||
| MS/MS | 69.50% | 64.40% | 0.738 | 66.10% |
ML, Machine Learning; LR, Logistic Regression, SVM, Support Vector Machine; ANN, Artificial Neural Network; DNN, Deep Neural Network; RF, Random Forest; (G)NB, (Gaussian) Naive Bayesian; GBM, Gradient Boosting Model; DT, Decision Tree; KNN, k-Nearest Neighbor; ESS, Ensemble-Based Scoring System; DIST, Euclidean Distance-Based Scoring System; MEWS, Modified Early Warning System; TIMI, Thrombolysis in Myocardial Infarction; GPLS, Generalized Partial Least Squares; RR, Ridge Regression; LASSO, Least Absolute Shrinkage and Selection Operator; MS/MS, Tandem Mass Spectrometry.