Table 3.

Predictive performance of different diagnostic models with the test set^a.

Model	Predictive performance
	Accuracy (95%CI)		Sensitivity	Specificity
Single metabolite model
Arachidonic acid	0.887	(0.732, 0.958)	0.933	0.887
Sebacic acid	0.867	(0.701, 0.943)	0.800	0.933
Indoxyl sulfate	0.850	(0.701, 0.942)	0.920	0.733
PC (14:0/0:0)	0.825	(0.672, 0.926)	0.760	0.933
Deoxycholic acid	0.773	(0.644, 0.910)	0.880	0.773
Trimethylamine N-oxide	0.653	(0.535, 0.834)	0.467	0.840
Machine learning model
PLS b	0.947	(0.830, 0.994)	0.960	0.933
RF c	0.947	(0.831, 0.994)	0.960	0.933
GBM d	0.960	(0.830, 0.994)	0.830	0.994
SVM e	0.980	(0.868, 0.999)	0.960	1.000

^aA total of 40 samples for the test set, including 25 samples from ESCC patients and 15 samples from healthy controls.

^bPartial least-square.

^cRandom forest.

^dGradient boosting machine.

^eSupport vector machine.