Cross-Dataset Wearable Sleep Staging under Source-Only and Target-Normalized Conditions

Abstract

Cross-dataset evaluation of wearable sleep staging depends on whether normalization uses statistics from the held-out dataset. We examine this issue in four public wearable datasets using a fixed three-class temporal fusion pipeline and leave-one-dataset-out (LODO) evaluation. We compare strict source-only normalization with an unlabeled target-normalized condition that estimates feature medians, means and standard deviations from the held-out inputs. The model, source splits, training budget, three random seeds and inference rule are identical. Target normalization increased macro-F1 by 0.340 on AAUWSS, 0.220 on BIDSleep and 0.182 on DREAMT, but decreased it by 0.018 on SleepAccel. These dataset-dependent changes show that normalization defines the evaluation information boundary. Target-normalized results represent transductive adaptation; strict source-only results represent generalization without target access.