Loading…
Sleep staging model that segments polysomnography from any single EEG and EOG channel pair, labelling stages at resolutions finer than 30 s epochs.
No providers recorded yet. Browse all providers
An automatic sleep stager trained at one clinic usually cannot be carried to the next: most systems assume a particular electrode montage, sampling rate and filter chain, so a new site means a retraining run before anyone trusts the output. That barrier has kept automated staging out of routine clinical use.
U-Sleep answers this by refusing to tell the model where a recording came from. It is a fully convolutional encoder–decoder over raw EEG and EOG, and during training every sample was shown a randomly drawn pair of one EEG and one EOG channel from whatever electrodes that recording happened to carry, with no dataset identifier supplied. Electrode placement, hardware and sampling rate become nuisance variables the network must be invariant to rather than assumptions baked into the pipeline, and one trained instance then scores recordings from clinics and protocols absent from its training data.
The second idea is temporal. Because the decoder restores the signal to its input resolution before anything is pooled into epochs, U-Sleep holds an internal sleep-stage representation at the full 128 Hz working rate, and its segment classifier aggregates that representation over whatever window is asked for. Trained against 30 s labels, it can emit stages far more often, and those sub-epoch labels carry diagnostic signal a conventional hypnogram discards. U-Sleep was developed at the University of Copenhagen Department of Computer Science with the Danish Center for Sleep Medicine at Rigshospitalet, and published in npj Digital Medicine in April 2021 as the successor to the same group's U-Time architecture.
Twelve encoder blocks (kernel-size-9 convolution, ELU, batch normalisation, max-pooling of width 2) are mirrored by twelve decoder blocks with skip connections, filter counts growing and shrinking by a factor of √2 per block from an initial 5. A segment classifier mean-pools the decoder output over the requested interval, then applies two point-wise convolutions and a softmax. That comes to roughly 3.1 million trainable parameters against U-Time's 1.1 million. Training used 19,924 PSG records from 15,660 participants across 16 clinical studies (21 datasets, about 19.4 years of annotated signal), sampling cohorts under a half-uniform, half-size-proportional policy so small datasets were not drowned out. Hyperparameters were carried over from U-Time rather than searched, and every reported number comes from the first and only trained instance.
Across the test splits of all 21 datasets the mean F1 was 0.79 ± 0.03, with per-stage scores of 0.90 (Wake), 0.53 (N1), 0.85 (N2), 0.76 (N3) and 0.90 (REM); eight of those datasets were untouched during development. On the consensus-scored DOD-H set of 25 healthy subjects U-Sleep reached 0.79 ± 0.07, statistically indistinguishable from the best of five clinical experts, and on DOD-O's 55 obstructive sleep apnea patients it reached 0.76 ± 0.10, above every individual expert. Separating apnea patients from controls using stage-transition triplets improved from F1 0.60 at two stages per minute to 0.94 at 1280 stages per minute.
Sleep laboratories can pre-score overnight studies for expert review, and cohort studies can score whole archives uniformly instead of pooling hypnograms from different scorers working to different manuals, and legacy recordings whose montages no longer match current practice can be reanalysed without adaptation. Scoring runs through a free web service for non-commercial research, with Python bindings and a REST API for batch work; the service is not approved for clinical decision-making, and commercial use is licensed separately.
U-Sleep became a common reference point for automated staging. Independent groups have applied the released model unchanged to populations it never saw — an assessment of its stability under differing AASM scoring practice, a paediatric cohort evaluation — and its architecture has been reimplemented in third-party EEG libraries and fine-tuned as a pretrained backbone for downstream sleep-disorder tasks. The main constraint is openness: the published checkpoint is not distributed, so reproducing it means assembling 21 datasets, several under data-access agreements, and retraining. N1 remains the weakest stage at F1 0.53, and the model deliberately ignores EMG, which carries information relevant to disorders such as REM sleep behaviour disorder.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.