The Pulse

Researchers use smartwatch signals to estimate how well people feel they slept

A machine-learning model combined smartwatch measurements with personal characteristics to distinguish participants’ very good and very poor sleep ratings. In data from 4,794 people, it achieved an area under the receiver operating characte

AI.info Team · Written with AI · fact-checked · 2026-11-10

A model for self-reported sleep

Researchers have developed a machine-learning method that uses smartwatch data and personal characteristics to estimate how people rate their sleep. The study reports that its model distinguished between extreme subjective sleep-quality ratings with an area under the receiver operating characteristic curve of 0.88.

The approach targets an aspect of sleep tracking that the authors say still depends on people answering questionnaires: how well they feel they slept. Wearable devices can continuously monitor physiological characteristics, while subjective assessments in the study came from self-reported daily questionnaires.

Three years of records

The researchers trained the model on data from the PerMed prospective observational study. The dataset covered 4,794 participants and included 759,116 self-reported daily questionnaires alongside 1,765,264 days of smartwatch recordings collected over three years.

Inputs included smartwatch sleep data, physical activity, heart rate and stress inferred from heart-rate variability. The researchers also used personal characteristics, including sociodemographic factors, personality traits and health-related attributes.

The study was conducted by Omer Hausner, Eva Mishor, Maya Reiner and Erez Shmueli, whose listed affiliations include Tel Aviv University’s Department of Industrial Engineering. Shmueli is also affiliated with Stanford Institute for Human-Centered Artificial Intelligence.

Performance on extreme ratings

The reported 0.88 score applies to distinguishing extreme subjective sleep-quality ratings. The plain-language summary describes those ends as nights participants rated very good or very poor; the reported result does not describe a score across every possible rating.

The researchers also report that combining smartwatch measurements with personal characteristics performed better than using either type of information alone. Their stated aim is passive, continuous assessment, rather than relying only on questionnaires completed at particular times.

Early version of the paper

Communications Medicine published the article on 10 November 2026 and identifies it as peer-reviewed, accepted research shared early. The journal says this version may receive further edits and will be replaced automatically by the final Version of Record.

The authors say the work was supported by the Israel Science Foundation, through its Israel Precision Medicine Partnership program, and by the Israel National Institute for Health Policy Research. They state that the funding agencies were not involved in the study’s design or in collecting, analyzing or interpreting the data, and declare no competing interests.