Under-Mattress Sleep Sensor Matched a Blood Test at Flagging Advanced Tau Pathology, Small Study Finds

A ballistocardiographic mat installed under participants' mattresses picked up advanced tau pathology as well as a blood test did in a presymptomatic cohort born in 1946, researchers reported at a scientific meeting. The finding comes from a conference abstract, not a validated diagnostic, and performed worse at earlier disease stages.

An under-mattress sleep sensor, deployed unobtrusively in the homes of members of Britain's 1946 birth cohort, identified advanced tau pathology on brain scans with an accuracy statistically comparable to a blood test for phosphorylated tau, according to a conference abstract from the InSleep46 study, nested within the long-running Insight 46 project and published in a supplement to Alzheimer's & Dementia in December 2025. The device, the Withings Sleep Analyzer, uses ballistocardiography, meaning it senses the tiny mechanical vibrations of the heart and breathing through the mattress, rather than a wrist-worn accelerometer.

Researchers led by teams at University College London and Imperial College London installed the sensor under the mattresses of Insight 46 participants at around age 78, after those same volunteers had already undergone 18F-Florbetapir amyloid PET scans around age 73, 18F-MK-6240 tau PET scans around age 77, and serial blood draws for plasma pTau217. In total the mats collected 63,720 nights of data, an average of about 240 nights per person. Among the 161 participants who had both sensor data and a tau PET scan, 12.4 percent showed early-stage (Braak1+) tau buildup and 6.2 percent showed more advanced (Braak3+) buildup; of 153 with amyloid PET, 25 percent were amyloid-positive. A model built from the sleep and circadian data, tested using a leave-one-out cross-validation method within the same sample, distinguished Braak3+ participants from others with an area under the curve of 0.75, which the researchers describe as comparable to plasma pTau217 in the same group.

The same model performed less well at picking out earlier-stage tau buildup (Braak1+) or amyloid positivity alone, the abstract states, meaning the signal strengthens only once pathology is more advanced. The result has not yet been published as a peer-reviewed paper; it is a meeting abstract validated only within its own dataset, not against an independent cohort. A companion preprint from the same research group, describing deployment of the sensor across 356 Insight 46 participants, reported that installation was feasible but that most people needed individualised, hands-on help to set the device up, an operational hurdle for any future population-level use.

The researchers say the approach could offer a low-burden way to flag people for further testing in future studies, but stress it was built and checked in one research cohort using cross-validation rather than an outside dataset, and the next step is testing it in independent groups before its accuracy can be considered established.