Machine learning methods to predict 30-day hospital readmission outcome among US adults with pneumonia: analysis of the national readmission database

Table 1 Evaluation of AUROC and AUPRC for all machine learning algorithms

Algorithm	AUROC (95% CI)		AUPRC (95% CI)	Net difference^b
Testing set
Rule-based model	0.6591 (0.6556–0.6627)	[Reference]	0.2146	[Reference]
Decision tree	0.5783 (0.5751–0.5815)	P < 0.001	0.156	− 0.0586
Random forest	0.6509 (0.6473–0.6545)	P < 0.01	0.2052	− 0.0094
XGBoost	0.6606 (0.657–0.6641)	0.015	0.2147	0.0001
LASSO	0.6087 (0.6053–0.612)	P < 0.001	0.2042	− 0.0104
Training set
Rule-based model	0.669 (0.6654–0.6725)	[Reference]	0.219	[Reference]
Decision tree	0.5773 (0.5741–0.5805)	P < 0.001	0.1556	− 0.0634
Random forest	0.6558 (0.6522–0.6594)	P < 0.001	0.2109	− 0.0081
XGBoost	0.6725 (0.669–0.6761)	P < 0.001	0.2279	0.0089
LASSO	0.6062 (0.6029–0.6095)	P < 0.001	0.2007	− 0.0183

The best performance model is in bold
ML, machine learning; XGBoost: Extreme Gradient Boosting; AUROC, area under receiver operating curve; LASSO, least absolute shrinkage, and selection operator; AUPRC: area under the precision-recall curve a p value is based on the DeLong test for comparison of area under the receiver operating characteristic curves for different models with reference to the rule-based model. b Calculated based on the net difference between all baseline models with reference to the rule-based model

ISSN: 1472-6947