Table 2.

Outcomes of the three prediction models.

ML model	Prediction model
Sample size (patients)	355 for training, 89 for validation, 30 for test, total 474
Sample zero ratio	Train 40.8%, validation 40.5%, test 40.0%
DNN	- 4 layers with 256–512-1024–512 neurons, RMSProp optimizer, relu activation - Training accuracy: 79.7% - Validation accuracy: 87.6% - Test accuracy: 80.0% - Validation AUC 0.887 with CI [0.824–0.951] Test AUC 0.819 with CI [0.685–0.954]
Logistic regression	- Training accuracy: 80.6% - Validation accuracy: 83.2% - Test accuracy: 63.3% - Validation AUC 0.845 with CI [0.772–0.918] Test AUC 0.667 with CI [0.505–0.829]
Random forest	- 500 estimators - Out-of-bag score estimate: 77.8% - Mean validation accuracy score: 84.3% - Mean test accuracy score: 76.7% - Validation AUC 0.855 with CI [0.783–0.926] - Test AUC 0.792 with CI [0.653–0.930]

ML machine learning; DNN deep neural network; SGD stochastic gradient descent; AUC area under the curve; CI confidence interval.