Using machine learning and natural language processing in triage for prediction of clinical disposition in the emergency department

Table 5 Performance comparison of Random Forest using different feature sets in the CMUH dataset

	AUC	p-value*	F1 score	Accuracy	Sensitivity	PPV	Specificity	NPV
Primary outcome
All features	0.926	-	0.500	0.955	0.591	0.433	0.969	0.983
Structured set	0.906	< 0.05	0.468	0.951	0.561	0.402	0.967	0.982
Unstructured set	0.861	< 0.05	0.406	0.941	0.526	0.331	0.958	0.981
LR-TTAS	0.858	< 0.05	0.000	0.962	0.000	0.000	1.000	0.962
Secondary outcome
All features	0.847	-	0.580	0.782	0.655	0.519	0.820	0.889
Structured set	0.812	< 0.05	0.534	0.756	0.611	0.475	0.799	0.874
Unstructured set	0.761	< 0.05	0.478	0.745	0.509	0.450	0.815	0.848
LR-TTAS	0.666	< 0.05	0.171	0.771	0.103	0.499	0.969	0.784

AUROC Area under receiver-operator curve, CMUH China Medical University Hospital, LR-TTAS Logistic Regression-Taiwan Triage Acuity scale, NPV Negative predictive value, PPV Positive predictive value
*The p-values in the table were derived using the DeLong test, which involved pairwise comparisons of each model’s AUROC with that of the model ranked directly above it in performance

ISSN: 1471-227X