Table 5.
TOP-Net performance based on transfer learning in the general ward (2-hour forecast range).
| Model | AUROCa, mean (SD) | Accuracy (%), mean (SD) | Sensitivity (%), mean (SD) | Specificity (%), mean (SD) | F1 score (%), mean (SD) | Precision (%), mean (SD) |
| TOP-Net | 96.5 (1.92) | 93.7 (1.02) | 95.5 (4.85) | 88.1 (4.28) | 79.3 (4.33) | 68.0 (5.99) |
| CNNb | 93.8 (2.02) | 95.3 (1.43) | 90.1 (2.88) | 88.1 (8.4) | 83.8 (5.38) | 78.8 (9.85) |
| LSTMc | 93.2 (1.89) | 92.6 (0.61) | 93.6 (2.76) | 81.5 (5.6) | 73.0 (3.4) | 60.0 (4.89) |
| XGBoostd | 89.9 (2.1) | 92.9 (1.1) | 83.4 (5.2) | 82.6 (7.9) | 73.7 (3.7) | 66.6 (6.8) |
| MLPe | 84.2 (4.1) | 91.0 (0.7) | 75.9 (9.6) | 78.9 (9.1) | 62.6 (2.0) | 54.0 (2.9) |
| Random forest | 87.3 (3.0) | 92.5 (1.0) | 76.6 (5.2) | 86.8 (4.7) | 75.0 (3.7) | 73.8 (4.9) |
aAUROC: area under the receiver operating characteristic curve.
bCNN: convolutional neural network.
cLSTM: long short-term memory.
dXGBoost: extreme gradient boosting.
eMLP: multilayer perceptron.