Sleep staging stops picking electrodes: any EEG-EOG mix scores a night, and resolution drops to seconds
Synopsis
Sleep staging has long depended on fixed electrode sets and 30-second expert-scored epochs, which makes multi-center studies hard to harmonize. AnySleep fuses any number of EEG and EOG channels through channel attention, trained and validated on 20,589 overnight recordings from 28 datasets, matching or slightly exceeding U-Sleep at 30-second epochs and staying strong with no EOG or a single derivation. At sub-30-second resolution, the share of expert-annotated arousal time overlapping predicted Wake rises from 7.7% at 30 seconds to 57.7% near 2 seconds.
Interpretation
A sleep staging model can now take any number and combination of EEG and EOG channels instead of a fixed electrode configuration. U-Sleep previously required exactly one EEG and one EOG channel, and its performance dropped markedly when channel order or type changed. Across two-channel permutations, AnySleep macro F1 stayed in a narrow 0.726 to 0.760 range; U-Sleep lost an average 0.160 to 0.203 macro F1 when EEG and EOG order was swapped.
The model extracts features per channel and merges them with learnable weights, so changing the channel count needs no change to the network structure. These channel-attention modules sit at multiple depths between encoder and decoder rather than fusing once at the input or output. On in-distribution test sets, AnySleep averaged 0.812 macro F1, late fusion 0.804, U-Sleep 0.799 and early fusion 0.787; on hold-out sets the four scored 0.777, 0.774, 0.768 and 0.752.
Refining the same network's temporal output below 30 seconds represents brief wake events that epoch scoring averages away. In conventional 30-second staging, only 7.7% of expert-annotated arousal duration fell on predicted Wake. That overlap peaked at 57.7% near 2 seconds and fell to 53.1% near 0.05 seconds; treating 3 to 15 second Wake runs as candidate arousals gave a maximum IoU F1 of 0.442 against expert annotations.
Second-scale stage sequences also support subject-level disorder discrimination, with transition patterns feeding random forest classifiers. Comparable analyses built on 30-second stages have coarser resolution and poorly capture short-term sleep fragmentation. At timescales between 0.1 and 1.0 seconds, macro F1 was 0.91 for obstructive sleep apnea, 0.86 for narcolepsy type 1 and 0.82 for insomnia against healthy controls.
Perspective
The model was trained and validated on 20,589 overnight recordings from 28 datasets, with 13 datasets used for training and validation and 15 hold-out datasets reserved for testing; 30-second staging performance on hold-out data is slightly below in-distribution data. High-resolution output was assessed indirectly, through overlap with expert-annotated arousals and its usefulness for disorder classification, and the authors call for expert-annotated, multi-rater datasets at high temporal resolution for direct evaluation.
The absolute accuracy of high-frequency sleep stages remains uncertain, because training and primary evaluation use 30-second labels and second-scale output was checked only through proxy measures. Performance drops slightly from in-distribution to hold-out datasets, and modest variability appears across training runs. The authors also note the model is about 12 MB but may still strain low-power consumer hardware for on-device staging.