Table 2.

AD Classification results using audio or text and with or without pre-training. AD: Alzheimer’s Disease. Accuracy: mean and standard deviation of results of 5 rounds. Best: highest accuracy of all epochs in 5 rounds.

Model	Pre-training dataset	Classes	Precision %	Recall %	Fl %	Accuracy %	Best %
Audio (Luz et al., 2020)	-	non-AD	67	50	57	62
		AD	60	75	67
MobileNet	-	non-AD	60.40 ± 7.86	58.40 ± 22.76	56.20 ± 13.79	59.00 ± 5.66	72.91
		AD	61.40 ± 6.89	59.80 ± 21.07	57.60 ± 10.54
	ImageNet	non-AD	72.80 ± 6.97	28.00 ± 8.15	40.40 ± 9.85	58.80 ± 3.49	77.08
		AD	55.80 ± 2.48	90.40 ± 1.96	69.00 ± 1.67
YAMNet	-	non-AD	52.20 ± 11.74	19.80 ± 22.61	24.60 ± 22.81	53.80 ± 6.88	79.17
		AD	53.40 ± 5.95	87.60 ± 9.56	65.80 ± 1.33
	AudioSet	non-AD	69.60 ± 6.80	59.20 ± 7.73	63.40 ± 5.57	66.20 ± 4.79	83.33
		AD	64.40 ± 3.93	73.40 ± 8.82	68.60 ± 4.84
Speech BERT	-	non-AD	67.74 ± 3.69	64.17 ± 3.34	65.82 ± 2.68	66.67 ± 2.95	77.08
		AD	65.84 ± 2.43	69.16 ± 5.65	67.39 ± 3.71
	LibriSpeech	non-AD	66.13 ± 4.12	55.00 ± 4.86	59.94 ± 3.78	63.33 ± 3.12	79.17
		AD	61.48 ± 2.76	71.67 ± 4.86	66.12 ± 3.08
Text (Luz et al., 2020)	-	non-AD	70	87	78	75	-
		AD	83	62	71
BERT base	-	non-AD	78.12 ± 1.98	74.17 ± 3.12	76.05 ± 1.82	76.67 ± 1.56	81.25
		AD	75.47 ± 2.08	79.17 ± 2.63	77.23 ± 1.50
	BooksCorpus/Wiki	non-AD	78.46 ± 1.89	85.00 ± 2.04	81.60 ± 1.96	80.83 ± 2.04	85.42
		AD	83.64 ± 2.22	76.67 ± 2.04	80.00 ± 2.13
BERT large	BooksCorpus/Wiki	non-AD	83.05 ± 5.12	80.00 ± 3.12	81.40 ± 3.09	81.67 ± 3.34	87.50
		AD	80.65 ± 2.66	83.33 ± 5.89	81.89 ± 3.64
Longformer	BooksCorpus/Wiki/ Realnews/Stories	non-AD	77.87 ± 3.75	90.00 ± 2.04	83.44 ± 2.33	82.08 ± 2.83	89.58
		AD	88.14 ± 2.09	74.17 ± 5.53	80.44 ± 3.55