Table 3.
Random Forest (RF) versus Gaussian process (GP) regression.
| 0 units transfused | 1–3 units transfused | 4 + units transfused | |
|---|---|---|---|
| Model development | |||
| Random forest | 0.829 (0.043) | 1.191 (0.047) | 5.799 (0.612) |
| Gaussian process regression | 0.064 (0.102) | 1.758 (0.033) | 7.613 (0.635) |
| Gaussian process regression for less severe cases | 0.766 (0.016) | ||
| Model validation | |||
| Random forest | 7.007 | 1.624 | 56.568 |
| Gaussian process regression | 0.117 | 1.705 | 56.941 |
| Gaussian process regression for less severe cases | 0.985 | ||
This table shows our GP regression model compared with the RF regression model. GP regression performed best as demonstrated by the low root mean square error (RMSE) in the 0 units transfused and 1–3 units transfused categories. In contrast, performance of both models suffered when predicting 4 + RBCs transfused. Therefore, in the final model we restricted the GP regression prediction to cases where < 4 RBCs transfused was predicted by GP classification.