A Novel Deep Learning Approach for Recognizing Stereotypical Motor Movements within and across Subjects on the Autism Spectrum Disorder

Table 3

Results of our models and previous works [7, 8, 21] on 6 subjects of Study1 and Study2 datasets. For each study, two to three video sessions were performed per subject except for Subject 6 (in Study2) who had only one session recorded; therefore, experiments could not be performed, which is indicated by “_”. Labels “CNN-Rad” and “RF-RQA” refer to the CNN performed by Rad et al. and the Random Forest with Recurrence Quantification Analysis. Abbreviations “Acc.” and “F1-sc.” refer to “accuracy” and “F1-score”, respectively.

	Experiments	S1		S2		S3		S4		S5		S6		Mean
		Acc.	F1-sc.	Acc.	F1-sc.	Acc.	F1-sc.	Acc.	F1-sc.	Acc.	F1-sc.	Acc.	F1-sc.	Acc.	F1-sc.
Study1	SVM-C [7]	85.90	73.00	85.30	36.00	94.00	50.00	66.50	73.00	75.10	44.00	87.30	46.00	82.35	53.67
	CNN-Rad [21]	_	71.00	_	73.00	_	70.00	_	92.00	_	68.00	_	94.00	_	78.00
	RF-RQA [8]	83.00	_	89.00	_	93.00	_	91.00	_	80.00	_	88.00	_	87.33	_
	Time-domain CNN	96.55	91.23	88.51	76.76	97.19	84.95	93.34	93.38	92.51	86.41	94.20	95.11	93.71	87.97
	Frequency-domain CNN	98.80	96.54	88.93	78.41	98.89	93.62	96.56	96.46	97.77	95.74	98.33	98.58	96.55	93.23
	Frequency-domain DBN	91.50	82.41	87.10	78.06	93.73	71.50	93.55	93.63	89.31	81.69	93.72	94.65	91.49	83.66
Study2	SVM-C [7]	71.00	43.00	79.00	26.00	99.00	3.00	90.00	86.00	73.00	72.00	_	_	82.40	46.00
	CNN-Rad [21]	_	68.00	_	22.00	_	2.00	_	77.00	_	75.00	_	_	_	48.80
	RF-RQA [8]	80.00	_	69.00	_	99.00	_	95.00	_	85.00	_	_	_	85.60	_
	Time-domain CNN	96.88	95.97	89.53	75.67	99.10	60.17	96.88	91.68	91.69	82.55	_	_	94.81	81.21
	Frequency-domain CNN	96.95	96.07	98.28	95.27	99.79	85.03	99.31	98.03	97.11	93.88	_	_	98.29	93.66
	Frequency-domain DBN	88.60	85.57	88.84	74.09	99.11	62.70	96.91	91.65	94.04	87.92	_	_	93.50	80.39