Sleep–Wake Model

A model that travels between devices

A robust generalizable device-agnostic deep learning model for sleep-wake determination from triaxial wrist accelerometry asks whether a single model can remain useful when the wrist sensor, patient population, and sleep disorder change. I am the paper’s second author.

Wrist accelerometry is convenient enough for long recordings outside a laboratory, but movement is only an indirect measure of sleep. Quiet wakefulness is particularly difficult: a motion-based model can easily mistake stillness for sleep, producing reassuring aggregate accuracy while missing the state clinicians most need it to distinguish.

Grounded in clinical sleep testing

The study paired wrist accelerometry with simultaneous polysomnography for 453 adults undergoing clinical sleep testing. Participants spanned a wide age range, included people with and without sleep disorders, and wore one of three accelerometer models. That design made cross-device and clinical generalization part of the evaluation rather than an assumption made after training.

Signals were summarized in 30-second epochs. The network first learned three states—wake, sleep, and sleep with arousals—before a decision tree collapsed those outputs into the final wake-versus-sleep determination. Training deliberately emphasized subjects with low sleep efficiency and/or high arousal index to confront difficult wake detection instead of optimizing around easy sleepers.

What the evaluation showed

On held-out recordings, the model reported an F1 score of 0.86, sleep sensitivity of 0.87, and wake specificity of 0.78. Its estimates also correlated with polysomnography-derived total sleep time (R = 0.69) and sleep efficiency (R = 0.63).

Performance remained consistent across the three accelerometer models and robust in the presence of conditions including sleep apnea and periodic limb movements. The important result is not a single headline score; it is evidence that the learned representation was not tied to one piece of hardware or one unusually tidy patient group.

Collaborative research

As second author, I contributed to the machine-learning research and the work of turning a clinical question into a reproducible evaluation. The project sits at the boundary I find most interesting: models are only meaningful when their validation reflects the people, instruments, and failure modes they will actually encounter. Read the paper on arXiv.