Table 4.

Ablation study comparing HomeSleepNet with its multiple variants.^a

Model variants	Cohen κ	Macro F1-score	Mean per-class sensitivity	Accuracy, %
SoundSleepNet	0.454	0.632	0.650	69.2
SoundSleepNet + transfer learning only	0.477	0.648	0.679	69.3
SoundSleepNet + consistency training only	0.501	0.670	0.660	73.5
HomeSleepNet	0.557	0.714	0.706	76.2

^aAll results were based on the 3-stage classification.