Table III.

HMM classifier accuracy K¹ for Problem 1 (second night PSG; n = 19).

	Scenario:	I (N1, N2, N3 pooled)	II (N1 and W pooled)	III (W and R pooled)	IV (N1 and N2 pooled)	V (N1 and R pooled)
	Unsupervised	58±7*	59±6	59±6	57±7*	56±7*
N3	Quasi-supervised	59±6	74±5	72±5	69±6	67±5
	Supervised	76±4	76±4†	76±4†	76±4†	76±4†
	Unsupervised	44±5*	46±5	47±5	44±5	46±6
N2	Quasi-supervised	55±4	66±4	66±3	62±4	65±4
	Supervised	72±3	72±3†	72±3†	72±3	72±3†
	Unsupervised	8±3*	7±3*	6±3	6±3*	7±3*
N1	Quasi-supervised	18±5	15±5	26±5	14±4	4±3
	Supervised	23±5†	23±5†	23±5	23±5†	23±5
	Unsupervised	41±7	40±7	39±7	40±7	40±7
R	Quasi-supervised	62±7	67±5	58±6	67±6	59±6
	Supervised	69±6†	69±6†	69±6	69±6†	69±6
	Unsupervised	54±6	51±7*	54±6*	56±5	57±5
W	Quasi-supervised	65±5	63±5	55±7	71±5	71±5
	Supervised	74±5†	74±5	74±5	74±5†	74±5†
	Unsupervised	44±4	44±4	45±4	44±3	45±4
All	Quasi-supervised	56±3	63±3	61±3	62±3	60±3
	Supervised	69±3	69±3†	69±3†	69±3	69±3

¹Cohen's kappa (mean±standard error).

^*Quasi-supervised model is not significantly different (p > 0.01) from unsupervised model according to a Wilcoxon sign rank test (matched samples).

^†Quasi-supervised model is not significantly different (p > 0.01) from supervised model according to a Wilcoxon sign rank test (matched samples).