Predicting Sleep Intent Boundaries from Wrist Actigraphy

Purpose : Insomnia is common, affecting roughly one in six people, and car-ries substantial personal and economic cost. A core diagnostic quantity is thedelay between the start of sleep attempts and actual sleep onset, which requiresknowing the moment a person begins trying to sleep. This moment is typicallyrecorded via sleep diaries, laboratory technicians, or a button-press on a wristdevice, all of which are costly or unreliable. We investigated whether wrist-wornaccelerometry alone can locate this event automatically, at low cost and withminimal burden on patients. Methods: We analysed wrist accelerometry from a two-arm, parallel-group ran-domised controlled trial in Norway of 582 adults with multiple sclerosis andclinically significant insomnia, comparing a digital therapy with a sleep-educationcontrol (NorseMS). Both arms were pooled for this secondary analysis. Partici-pants’ own button-presses served as reference labels, and concurrently collectedsleep diaries served as a second independent reference. We present Lights-OutTime from Unobtrusive Sensing (LOTUS), a framework to predict the nightlybutton-press event, using a widened probabilistic target rather than a singletime point and decoding the predicted distribution to a clock time. Performancewas assessed by five-fold cross-validation split by participant, with localisationerror in minutes as the primary measure, against a personal-median baseline, anestablished rule-based sleep-detection algorithm, and a pretrained self-supervisedencoder. We conducted ablations and assessed agreement with diaries usingBland–Altman plots. Results: LOTUS achieved a mean localisation error of 14.0 minutes and amedian of 6.7 minutes. Its mean error was lower than all comparators; in terms1of median performance, a self-supervised baseline was comparable but failed onatypical nights, so LOTUS was more robust overall. Agreement between LOTUSand diaries was comparable to that between button-presses and diaries. Down-sampling and a latency analysis indicated LOTUS captures pre-sleep behaviourrather than the button-press movement or sleep onset itself. Conclusions: Accelerometry alone can locate sleep intent about as closely asexisting self-report measures do, potentially recovering nights without a button-press and reducing reliance on diaries and manual marking in insomnia trials andcare.

Authors

Institutions

Publication Details

Journal
Research Square
Published
2026-11-19
DOI
https://doi.org/10.21203/rs.3.rs-11280339/v1
Primary Topic
Sleep and related disorders
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Predicting Sleep Intent Boundaries from Wrist Actigraphy

Kerstin Bach, Simen Berg Saksvik, Stuart Gallina Ottersen
Research Square
Sleep and related disorders
preprint

Predicting Sleep Intent Boundaries from Wrist Actigraphy

Kerstin Bach, Simen Berg Saksvik, Stuart Gallina Ottersen
preprint en

Abstract

Purpose : Insomnia is common, affecting roughly one in six people, and car-ries substantial personal and economic cost. A core diagnostic quantity is thedelay between the start of sleep attempts and actual sleep onset, which requiresknowing the moment a person begins trying to sleep. This moment is typicallyrecorded via sleep diaries, laboratory technicians, or a button-press on a wristdevice, all of which are costly or unreliable. We investigated whether wrist-wornaccelerometry alone can locate this event automatically, at low cost and withminimal burden on patients. Methods: We analysed wrist accelerometry from a two-arm, parallel-group ran-domised controlled trial in Norway of 582 adults with multiple sclerosis andclinically significant insomnia, comparing a digital therapy with a sleep-educationcontrol (NorseMS). Both arms were pooled for this secondary analysis. Partici-pants’ own button-presses served as reference labels, and concurrently collectedsleep diaries served as a second independent reference. We present Lights-OutTime from Unobtrusive Sensing (LOTUS), a framework to predict the nightlybutton-press event, using a widened probabilistic target rather than a singletime point and decoding the predicted distribution to a clock time. Performancewas assessed by five-fold cross-validation split by participant, with localisationerror in minutes as the primary measure, against a personal-median baseline, anestablished rule-based sleep-detection algorithm, and a pretrained self-supervisedencoder. We conducted ablations and assessed agreement with diaries usingBland–Altman plots. Results: LOTUS achieved a mean localisation error of 14.0 minutes and amedian of 6.7 minutes. Its mean error was lower than all comparators; in terms1of median performance, a self-supervised baseline was comparable but failed onatypical nights, so LOTUS was more robust overall. Agreement between LOTUSand diaries was comparable to that between button-presses and diaries. Down-sampling and a latency analysis indicated LOTUS captures pre-sleep behaviourrather than the button-press movement or sleep onset itself. Conclusions: Accelerometry alone can locate sleep intent about as closely asexisting self-report measures do, potentially recovering nights without a button-press and reducing reliance on diaries and manual marking in insomnia trials andcare.

Research Square
Norwegian University of Science and Technology (NO)
Sleep and related disorders
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon