Table 6.

Sleep stage classification accuracy across different features and classifiers in the Apple Watch (PPG, MEMS) dataset

		Wake correct	NREM correct	REM correct	Best accuracy	κ
Logistic regression	Motion	0.6	0.506	0.332	0.71	0.085
	HR	0.6	0.452	0.453	0.698	0.033
	Motion, HR	0.6	0.625	0.625	0.701	0.161
	Motion, HR, Clock	0.6	0.623	0.623	0.699	0.13
k-Nearest neighbors	Motion	0.6	0.294	0.532	0.698	0.072
	HR	0.6	0.402	0.402	0.671	0.108
	Motion, HR	0.6	0.607	0.605	0.711	0.227
	Motion, HR, Clock	0.6	0.648	0.647	0.721	0.243
Random forest	Motion	0.6	0.397	0.441	0.702	0.075
	HR	0.6	0.434	0.434	0.676	0.165
	Motion, HR	0.6	0.615	0.615	0.695	0.293
	Motion, HR, Clock	0.6	0.638	0.638	0.686	0.302
Neural net	Motion	0.6	0.394	0.498	0.713	0.084
	HR	0.6	0.454	0.454	0.698	0.04
	Motion, HR	0.6	0.622	0.622	0.723	0.256
	Motion, HR, Clock	0.6	0.651	0.65	0.723	0.277

Performance metrics for wake/NREM/REM classification across multiple classifiers with use of motion, HR, clock proxy, or combination of features. NREM and REM Correct refer to the fraction of NREM and REM sleep epochs scored correctly when a threshold is chosen so they are as close as possible, while maintaining the fraction of correctly scored wake epochs at 0.6. Best accuracy refers to the highest accuracy found during the threshold search, and κ is the Cohen’s kappa for that accuracy. HR, heart rate.