Mastoid ExG U-Sleep

Loss of muscle atonia during REM sleep is the defining sign of REM sleep behaviour disorder (RBD), and by convention it is measured at the chin: a technician mounts submental EMG electrodes as part of a full polysomnography montage, and an expert scores the night epoch by epoch alongside infrared video. That workflow is why RBD — the strongest prodromal marker of Parkinson's disease and dementia with Lewy bodies — is established in sleep laboratories rather than at the scale community screening would need. This system replaces the montage with two self-adhesive electrodes behind the ears. The bipolar derivation between the left and right mastoid carries both the EEG needed to stage sleep and residual muscle activity that tracks chin atonia closely enough to quantify REM sleep without atonia (RWA) from that one channel.

The learned component is a sleep-staging network built on the U-Sleep architecture and trained from scratch on mastoid-and-EOG derivations by researchers at Aarhus University's Lundbeck Foundation Parkinson's Disease Research Center (PACE) and Department of Electrical and Computer Engineering, with collaborators in Cologne, Bonn and Jülich. Its hypnograms feed a rule-based RWA detector and a logistic-regression classifier, so the network supplies sleep architecture and the pipeline as a whole makes the diagnostic call. It is a companion to the same group's U-Sleep for RBD/PD, which stages a conventional clinical montage; here the montage itself shrinks.

#Key Features

• Two electrodes, three signal types: a single bipolar mastoid channel yields EEG for staging and an EMG surrogate for atonia, so no chin, scalp or reference electrodes are needed. A pair of EOG electrodes adds a second channel when sleep architecture matters more than a screening decision.
• Fully automated end to end: nothing in the pathway — staging, artefact rejection, RWA quantification, the RBD call — involves a human scorer, removing inter-rater variability along with the technician visit.
• Channel-pair ensembling: inference runs the model over all 73 valid derivation pairs and averages their per-epoch softmax distributions, dropping pairs whose channels are largely missing, so a detached electrode shifts weight to the survivors rather than ruining the night.
• Confidence-gated REM: only stage R epochs the model scores above 80% confidence are passed to the RWA estimator. This discards about 35% of REM epochs per night and raises stage R precision from 78.3% to 87.3%.
• A continuous atonia measure: RWA is a fraction of artefact-free REM duration, not binary 3-second mini-epochs, preserving graded signal that mini-epoch scoring discards.

#Technical Details

A Base Model was trained from scratch on an openly available multi-site corpus, filtered from almost 20,000 polysomnograms down to the 6,542 nights from 6,055 adults that carried bilateral mastoid and both EOG electrodes. Training enforced invariance to the electrode combination by sampling each 17.5-minute example from a pool of nine derivations. The Base Model was then fine-tuned on the PACE-CBC dataset (394 participants, 443 nights of iRBD, Parkinson's disease and controls) and again on 84 concurrent polysomnography-plus-home-device nights, adapting it to the take-home montage's lower, displaced placements. Against expert scoring on PACE-CBC, agreement rose from κ = 0.657 to 0.725 for the two-channel model and from 0.533 to 0.654 for the mastoid-only model. Automated mastoid RWA correlated with expert SINBAR chin-EMG scoring at r = 0.82 (n = 136). On 76 participants recording themselves at home, the mastoid-only pipeline detected RBD with an AUC of 0.95 (94% sensitivity, 86% specificity); classifier coefficients fitted on in-lab polysomnography transferred to the home cohort unchanged at AUC 0.96, and the same approach reached AUC 0.93 across 378 in-lab participants (420 recordings).

#Applications

The intended use is second-stage screening: after a questionnaire or genetic risk assessment flags someone, a mailed electrode kit and one night at home can prioritise them for specialist assessment instead of sending everyone to video-polysomnography. Beyond the binary call, the continuous RWA estimate gives observational cohorts and neuroprotective trials a repeatable quantity to track, and per-epoch confidence lets a study reject unreliable nights automatically. Sleep researchers also gain hypnograms from a montage patients apply themselves.

#Impact

Prodromal cohorts are the population disease-modifying trials most need and the hardest to assemble, and this is the first system to derive both sleep staging and quantitative RWA from self-applied electrodes. The evidence is early: the work is a preprint, no code or weights have been released, and Aarhus University and Aarhus University Hospital have a patent application pending on the mastoid-based system. Fine-tuning and validation cohorts were predominantly European and male, and the system has not been tested as a first-line screening tool, since controls were pre-selected by expert interview and community mimics such as untreated apnoea are under-represented. Night-to-night RWA varied about twice as much in RBD as in controls, so a single night suffices for a diagnostic decision while progression monitoring needs several.