A meta-analysis of the diagnostic test accuracy of artificial intelligence predicting emergency department dispositions

Table 9 Summarization of the performance of predictive models for three emergency department dispositions

Disposition	Metric	Characteristic
		Data		Sample	Artificial-intelligence technique
		Source	Feature Type	Group	Approach	Top algorithm	Ensemble	Cross validation
Admission	Sens	Public > Private	S > B > U	A = Y > E > M	DL > ML	CNN	No ensemble > Ensemble	No CV > CV
Admission	Spec	Public > Private	S > B > U	M > A > Y > E	ML > DL	CNN	No ensemble > Ensemble	No CV > CV
Critical care	Sens	Private > Public	S > B	M > A > Y	ML > DL	XGB	Ensemble > No ensemble	CV > No CV
Critical care	Spec	Private > Public	S > B	A > M > Y	ML > DL	LightGBM	Ensemble > No ensemble	No CV > CV
Mortality	Sens	Public > Private	S > B	Y > E > A > M	ML > DL	RF/LightGBM	Ensemble > No ensemble	CV = No CV
Mortality	Spec	Public > Private	S > B	A > M > Y > E	ML > DL	RF	Ensemble > No ensemble	CV > No CV

Notes:
1. Metric: Sens = Sensitivity and Spec = Specificity
2. Source: Public = public data source and Private = private data source
3. Type of feature: S = structured data, B = structured and unstructured data, and U = unstructured data
4. Sample: A = adults, Y = youths, E = elders, and M = mixed samples
5. Approach: ML = machine learning and DL = deep learning
6. CNN = Convolutional neural network, XGB = eXtreme gradient boosting, and RF = Random forest
7. Cross-validation: CV = cross-validation

ISSN: 1472-6947