Table 2.
Prediction performance of the reference, and machine learning models in infants hospitalized for bronchiolitis.
| Outcomes and models | AUC | P-valuea | NRIb | P-valueb | Sensitivity | Specificity | PPV | NPV |
|---|---|---|---|---|---|---|---|---|
| Positive pressure ventilation outcome | ||||||||
| Reference model | 0.62 (0.53–0.70) | Reference | Reference | Reference | 0.62 (0.49–0.75) | 0.57 (0.54–0.60) | 0.075 (0.054–0.097) | 0.96 (0.95–0.97) |
| Logistic regression with Lasso regularization | 0.88 (0.84–0.93) | < 0.001 | 1.09 (0.87–1.32) | < 0.001 | 0.84 (0.73–0.93) | 0.79 (0.77–0.82) | 0.19 (0.14–0.24) | 0.99 (0.99–0.99) |
| Logistic regression with elastic net regularization | 0.89 (0.85–0.92) | < 0.001 | 1.05 (0.82–1.28) | < 0.001 | 0.89 (0.80–0.96) | 0.73 (0.70–0.75) | 0.15 (0.11–0.18) | 0.99 (0.99–0.99) |
| Random forest | 0.89 (0.85–0.92) | < 0.001 | 1.17 (0.96–1.38) | < 0.001 | 0.85 (0.75–0.95) | 0.74 (0.71–0.76) | 0.15 (0.12–0.21) | 0.99 (0.99–0.99) |
| Gradient boosted decision tree | 0.88 (0.84–0.93) | < 0.001 | 1.08 (0.84–1.33) | < 0.001 | 0.89 (0.80–0.96) | 0.77 (0.75–0.80) | 0.17 (0.08–0.21) | 0.99 (0.99–0.99) |
| Intensive treatment outcome | ||||||||
| Reference model | 0.62 (0.57–0.67) | Reference | Reference | Reference | 0.58 (0.55–0.62) | 0.58 (0.50–0.66) | 0.21 (0.18–0.24) | 0.88 (0.86–0.89) |
| Logistic regression with Lasso regularization | 0.79 (0.76–0.83) | < 0.001 | 0.68 (0.52–0.84) | < 0.001 | 0.75 (0.69–0.82) | 0.70 (0.66–0.73) | 0.31 (0.26–0.38) | 0.94 (0.93–0.94) |
| Logistic regression with elastic net regularization | 0.80 (0.76–0.83) | < 0.001 | 0.58 (0.42–0.74) | < 0.001 | 0.72 (0.64–0.79) | 0.74 (0.71–0.77) | 0.33 (0.28–0.41) | 0.93 (0.92–0.94) |
| Random forest | 0.79 (0.75–0.84) | < 0.001 | 0.70 (0.55–0.86) | < 0.001 | 0.70 (0.63–0.77) | 0.78 (0.76–0.81) | 0.37 (0.29–0.45) | 0.93 (0.92–0.94) |
| Gradient boosted decision tree | 0.79 (0.75–0.84) | < 0.001 | 0.72 (0.57–0.87) | < 0.001 | 0.74 (0.67–0.80) | 0.74 (0.71–0.77) | 0.33 (0.26–0.42) | 0.93 (0.92–0.94) |
AUC area under the receiver-operating-characteristic curve, NRI net reclassification improvement, PPV positive predictive value, NPV negative predictive value.
aP-value was calculated to compare area-under-the-curve of the reference model with that of each machine model.
bWe used continuous NRI and its P-value.