Loading…
Automated sleep staging for polysomnography in Parkinson's disease and isolated REM sleep behaviour disorder, with per-epoch confidence estimates.
No providers recorded yet. Browse all providers
The AASM scoring rules that define a sleep stage assume the sleeper has the features those rules key on. N2 is called on spindles and K-complexes; REM is called on rapid eye movements plus muscle atonia. In Parkinson's disease (PD) and isolated REM sleep behaviour disorder (iRBD), spindles thin out, rapid eye movements can be sparse or absent, and atonia is by definition compromised — so raters fall back on rules of absence, and agreement erodes. Interrater agreement is around Cohen's κ = 0.76 in healthy adults but κ ≈ 0.61 in the largest PD study, and automated scorers inherit the gap: one random-forest stager fell from κ = 0.73 in elderly controls to κ = 0.54 in RBD.
This model is U-Sleep, the fully convolutional sleep-staging network, adapted to that population by a group at Aarhus University with the Cologne-Bonn and Copenhagen sleep centres. U-Sleep was first pretrained on PUB, a 19,236-recording multisite corpus of healthy sleepers, children and apnoea patients, where it reached κ = 0.81 — and only κ = 0.66 when that checkpoint was pointed at PD and iRBD cohorts. Fine-tuning jointly on two research cohorts produced the released Generalized Model, which recovers κ = 0.74 there and carries the gain to a clinical centre it never saw.
The second contribution is that the network's per-epoch output distribution is treated as a usable quantity rather than discarded: because confidence tracks staging accuracy, it can be thresholded to trade recall for precision on a stage, or used to route questionable nights to a human.
U-Sleep is an encoder-decoder convolutional network that segments the raw signal at its native 128 Hz sampling rate, with a segment classifier averaging the segmentation into 30-second epochs. This work uses an encoder/decoder depth of 12, a progression factor of 2 and a complexity factor of 1.67, giving roughly 3.1 million trainable parameters. Pretraining on PUB — 19,236 polysomnograms from about 15,000 subjects across 12 sites, with the hold-out centre's recordings removed — gave a validation κ of 0.813. Fine-tuning used 134 PACE and 215 Cologne-Bonn recordings (112 PD, 138 iRBD, 89 controls) with cross-entropy loss, Adam, a fixed learning rate of 10⁻⁵ and 10-fold cross-validation, the folds combining into a majority-vote ensemble for hold-out testing. On the independent Danish Centre for Sleep Medicine cohort (81 PD, 36 iRBD, 87 sleep-clinic controls), mean per-night κ rose from 0.60 to 0.64 and median κ from 0.64 to 0.69 over the pretrained model. Median κ there was 0.75 in controls, 0.65 in iRBD, 0.64 in PD with RBD and 0.59 in PD without RBD; stage-R F1 reached 0.86 in iRBD patients.
Confirmatory video-polysomnography is the bottleneck in RBD screening, where questionnaire and biomarker triage now generate more candidates than sleep laboratories can manually score. A stager that holds up in this population lets those pipelines scale, and lets retrospective multicentre cohorts scored by different raters be restaged consistently. The confidence output serves two further workflows: selecting high-precision REM epochs to feed automated RSWA quantification tools, and flagging low-confidence nights or arousal-affected epochs for expert review.
The result argues that generalisable sleep staging in neurodegenerative disease is achievable without a model per clinic — per-site fine-tuning being exactly what most centres cannot afford. Code and both checkpoints ship with a demo notebook that reads MNE-compatible EDF, SET and VHDR files, though they are restricted to non-commercial research use and the underlying polysomnography cannot be shared. The authors are explicit about the boundaries: the cohorts are almost entirely Northern European, cover early-to-moderate PD, and generalisation may depend on recording equipment. Video was unavailable, so REM misclassified as wake during dream enactment cannot be excluded, and an RBD diagnosis still requires RSWA quantification on top of the staging this model provides.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.