Table 2 Performance of EV classifier with varying feature selection methods, feature set size and train/test split.
From: Radiomics approach for identifying radiation-induced normal tissue toxicity in the lung
Feature selection method | Size of feature set | Predictor variable | Classification model | Train/test split | AUC | Accuracy (%) | Confidence intervals | P-value (*if < 0.05) | Sensitivity | Specificity |
|---|---|---|---|---|---|---|---|---|---|---|
Recursive feature elimination | 14 | EV | Random forest | 50/50 | 0.859 | 85.7 | 69.7–95.2 | * 2.28 × 10−5 | 0.941 | 0.778 |
Feature importance | 14 | EV | Random forest | 50/50 | 0.832 | 82.9 | 66.4–93.4 | * 1.13 × 10−3 | 0.941 | 0.722 |
Principal components | 14 | EV | Random forest | 50/50 | 0.546 | 54.3 | 36.7–71.2 | 4.34 × 10−1 | 0.647 | 0.444 |
Recursive feature elimination | 14 | EV | Random forest | 60/40 | 0.795 | 78.6 | 59.1–91.7 | 5.7 × 10−2 | 0.923 | 0.667 |
Recursive feature elimination | 14 | EV | Random forest | 70/30 | 0.727 | 71.4 | 47.8–88.7 | 6.15 × 10−2 | 1.00 | 0.455 |
Recursive feature elimination | 14 | EV | Random forest | 80/20 | 0.750 | 71.4 | 41.9–91.6 | 2.11 × 10−1 | 1.00 | 0.500 |
Recursive feature elimination | 21 | EV | Random forest | 50/50 | 0.801 | 80.0 | 63.1–91.6 | * 1.4 × 10−3 | 0.812 | 0.790 |
Recursive feature elimination | 34 | EV | Random forest | 50/50 | 63.1 | 60.0 | 42.1–76.1 | 5.73 × 10−1 | 0.476 | 0.786 |
