Use of natural language processing to improve predictive models for imaging utilization in children presenting to the emergency department

Table 3 Predictive performance of logistic regression models with 10-fold classification in identifying patients with various medical imaging use during emergency department triage, NHAMCS 2012–2016

	Probability cut-off	Sensitivity	Specificity	Accuracy	AUC (95% CI)
Any Imaging use
Unstructured variables	0.28	0.72	0.74	0.73	0.810 (0.807–0.813)
Structured variables	0.31	0.62	0.67	0.66	0.706 (0.698–0.714)
Unstructured + Structured variables	0.27	0.75	0.73	0.74	0.824 (0.818–0.829)
Xray
Unstructured variables	0.22	0.73	0.74	0.74	0.824 (0.822–0.826)
Structured variables	0.26	0.61	0.67	0.65	0.694(0.685–0.704)
Unstructured + Structured variables	0.22	0.75	0.74	0.74	0.834 (0.830–0.839)
CT Scan
Unstructured variables	0.04	0.79	0.77	0.78	0.845 (0.832–0.858)
Structured variables	0.05	0.71	0.69	0.69	0.771 (0.759–0.783)
Unstructured + Structured variables	0.04	0.80	0.79	0.79	0.868(0.858–0.878)

Note: The best cutoff of the probabilities was determined by using the point on the ROC curve with the shortest distance to the upper left corner (where sensitivity = 1 and specificity = 1)

ISSN: 1472-6947