Table 1.

Performances of the developed classifiers.

Classifier		Precision	Recall	F1 score	P value
EDa-irrelevant versus other 2 labelsb
	CNNc	0.88	0.89	0.89	N/Ad
	LSTMe	0.86	0.89	0.88	.15
	NBf	0.85	0.73	0.75	<.001
	LNg	0.84	0.78	0.81	<.001
	SVMh	0.87	0.83	0.85	<.001
	RFi	0.86	0.85	0.86	.005
	GBj	0.77	0.75	0.76	<.001
ED-promotional and education versus ED-laypeoplek
	LSTM	0.90	0.89	0.90	N/A
	CNN	0.87	0.87	0.87	.006
	NB	0.80	0.74	0.76	<.001
	LN	0.83	0.80	0.81	<.001
	SVM	0.82	0.79	0.80	<.001
	RF	0.84	0.82	0.83	<.001
	GB	0.84	0.82	0.83	<.001

^aED: eating disorder.

^bED-irrelevant versus other 2 labels: in this task, the performances of CNN and LSTM have no significant difference; they are both significantly higher than the others (P<.01).

^cCNN: convolutional neural network.

^dN/A: not applicable.

^eLSTM: long short-term memory.

^fNB: naïve Bayes.

^gLN: linear regression.

^hSVM: support vector machine.

ⁱRF: random forest.

^jGB: gradient boosting trees.

^kED-promotional and education versus ED-laypeople: in this task, the performance LSTM is significantly higher than the others (P<.01).