Sleep, decoded
without a wire in sight

The Problem With One Good Night

Sleep quality swings from night to night — which means a single night of lab-based polysomnography (PSG), the clinical gold standard, can misjudge a person's real sleep. Multi-night tracking with consumer devices suggests up to 20% of sleep apnoea cases get misdiagnosed, and half get the wrong severity rating, when judged on one night instead of two weeks.

PSG is accurate but impractical to repeat: it's wired, lab-bound, and expensive to score. Wearables solve the repeatability problem but introduce their own — skin contact, compliance, motion artifacts. MEDBUG tries a third path: a completely non-contact strip under the mattress cover, reading ballistocardiographic (BCG) signals — the tiny mechanical push of every heartbeat and breath — alongside body pressure and ambient light.

Where The Model Shines — And Where It Struggles

Tested against expert-scored PSG in 60 independent participants, the four-stage classifier (Wake, Light, Deep, REM) was strongest at telling REM sleep and wakefulness apart from everything else, and comparatively weaker at pinning down light sleep — a transitional stage whose signal genuinely overlaps its neighbours.

The bias runs in one direction: the model tends to read light sleep as REM or deep sleep, overestimating REM by about 40 minutes a night and underestimating light sleep by roughly 68 minutes. Total sleep time and sleep efficiency, by contrast, track PSG closely — the correlation for total sleep time was r = 0.78 with essentially no systematic bias (−2.6 min).

Built On A Pretrained Sleep Brain

Under the hood, an 8-layer Transformer sits on top of a VGG-style convolutional encoder — pretrained on a large internal dataset, then fine-tuned on just 60 recordings for this study. Three-minute input windows let it reason across long stretches of the night rather than judging each 30-second epoch in isolation, which helps it track the slow architecture of NREM/REM cycling.