Table 12.

Binary classification scores using 56 depressed users and 3 of their matched control users and 6 temporal post subsets^a.

			Positive predictive value, mean (SD)		Sensitivity, mean (SD)		F1-score, mean (SD)
Last 4 weeks
	BERTb LMc	0.480 (0.027)		0.538 (0.019)		0.489 (0.010)
	MentalBERT LM	0.494 (0.019)		0.577 (0.009)		0.525 (0.007)
Last 8 weeks
	BERT LM	0.446 (0.032)		0.538 (0.036)		0.472 (0.035)
	MentalBERT LM	0.427 (0.027)		0.524 (0.029)		0.461 (0.023)
Last 12 weeks
	BERT LM	0.498 (0.031)		0.619 (0.037)		0.543 (0.035)
	MentalBERT LM	0.448 (0.007)		0.569 (0.017)		0.494 (0.009)
Last 16 weeks
	BERT LM	0.471 (0.010)		0.565 (0.021)		0.504 (0.011)
	MentalBERT LM	0.481 (0.023)		0.643 (0.037)		0.541 (0.028)
Last 20 weeks
	BERT LM	0.475 (0.039)		0.577 (0.037)		0.510 (0.034)
	MentalBERT LM	0.487 (0.018)		0.595 (0.011)		0.524 (0.009)
Last 24 weeks
	BERT LM	0.470 (0.033)		0.591 (0.036)		0.518 (0.033)
	MentalBERT LM	0.501 (0.022)		0.591 (0.018)		0.536 (0.022)
All posts
	BERT LM	0.625 (0.021)		0.519 (0.032)		0.562 (0.015)
	MentalBERT LM	0.588 (0.005)		0.508 (0.010)		0.540 (0.003)
Naive baseline			0.250 (N/Ad)		1.000 (N/A)		0.400 (N/A)

^aThe classifiers used are BERT LM and MentalBERT LM, both of whose experiments were run 3 times each, therefore both mean and SD scores are provided..

^bBERT: Bidirectional Encoder Representations From Transformer.

^cLM: language model.

^dN/A: not applicable.