Somewhere in a sleep laboratory at Sunnybrook Health Sciences Centre in Toronto, an adult is wired for a clinical polysomnogram: electrodes on the scalp, bands on the chest, the whole apparatus that lets a technologist say, epoch by epoch, thirty seconds at a time, this person is awake and this person is asleep. On the same wrist, quietly, sits a small plastic accelerometer. It knows nothing about brain waves. It knows only that the arm moved, or did not. The gap between those two kinds of knowing is where a great deal of dementia sleep research lives.
That gap is the subject of a preprint posted to arXiv on December 1, 2025 by Nasim Montazeri, Andrew Lim and colleagues. They collected wrist accelerometry simultaneous with polysomnography in 453 adults aged 40 and over who were referred for clinical sleep testing, using three commonly used research devices: the Axivity, the GENEActiv and the ActiGraph GT9X. They cut the signal into 30-second epochs, trained a model to sort each epoch into three classes (wake, sleep, and sleep with arousals), then collapsed that into the familiar binary of sleep versus wake with a decision tree. To force the model to get better at the hard part, they trained it deliberately on recordings from people who slept badly: short total sleep time, disrupted nights, high arousal burden. Across the test sets, the model reached an F1 score of 0.86, sensitivity for sleep of 0.87 and specificity for wake of 0.78, with moderate correlation to polysomnography for total sleep time (R=0.69) and sleep efficiency (R=0.63).
The number that matters most is the specificity one, and it is worth explaining why. Movement-based sleep scoring has always been good at calling sleep and bad at calling wakefulness, because lying still in the dark, awake, looks to an accelerometer exactly like sleeping. In the Axivity test set, covering 621 hours of recording, the model posted an F1 of 0.90, sensitivity of 0.91, specificity of 0.79 and a Cohen's kappa of 0.71. Against GGIR, the widely used open algorithm the authors chose as comparator, sensitivity was essentially a wash (-0.01) while specificity improved by 0.11 and kappa by 0.09. Similar differentials appeared in the GENEActiv and GT9X sets. Across 381 individual recordings from the three devices, the deep learning model's mean F1, accuracy, precision, kappa and specificity were all higher. Performance held up in people with sleep apnea and with periodic limb movements in sleep, and was stable across age and sex for two of the three accelerometers.
For families reading this section, none of that is a treatment or a test anyone can ask for. What it is, is plumbing. The reason actigraphy keeps appearing in dementia research is that it can sit on a wrist for weeks in someone's own home, which polysomnography cannot. In 2013 Lim and colleagues, working in a large cohort of senior adults, quantified sleep fragmentation from up to 10 consecutive days of actigraphy and followed participants for a mean of 3.3 years, during which 97 developed Alzheimer's disease. Greater fragmentation was associated with higher risk (hazard ratio 1.22 per standard deviation increase), with roughly a 1.5-fold difference between the most and least fragmented. That was an observational finding; it showed association, not cause. But everything it reported rested on a definition of "fragmented" produced by one device and one algorithm.
That is the standardization problem this preprint is aimed at. If one cohort's wristband and scoring routine overcalls sleep and another's does not, two studies can disagree about whether broken sleep tracks with later dementia for reasons that have nothing to do with brains. A model that reads raw triaxial accelerometry the same way across brands is, in principle, the thing that lets separate cohorts be pooled and separate findings be replicated. The same Toronto group published a related deep learning sleep staging model for a non-EEG wearable system in the journal SLEEP in February 2026, which suggests a sustained push in this direction rather than a one-off.
The authors are candid about what they have not done. The recordings were overnight, with only a brief stretch of wakefulness before lights out, so sleep heavily outnumbered wake in the data, and they state plainly that it is unclear how well the model would detect daytime naps or perform across periods dominated by wake. Daytime napping and day-night reversal are precisely what families living with dementia watch most closely. All recordings came from a single site. And this was a validation against a sleep technologist's scoring, not a study of memory, diagnosis or outcome.