SMARTsleep

The challenge

Polysomnography is the clinical reference standard for sleep assessment but is costly, intrusive and unsuitable for long-term or large-scale monitoring. Conducted in specialised sleep laboratories with multiple sensors attached to the patient, polysomnography provides comprehensive data but requires significant resources and disrupts normal sleep patterns. The procedure is impractical for repeated assessments or population-level screening. Whilst wearable devices offer a practical alternative for continuous sleep monitoring in natural environments, existing sleep algorithms often lack transparency, standardisation and clinical-grade validation. Many commercial devices use proprietary algorithms that cannot be independently verified or adapted, limiting their utility in research and clinical applications. The absence of standardised approaches hinders comparison across studies and devices, slowing progress in digital sleep health.

The project: SMARTsleep

TNO developed a standardised, device-agnostic framework for sleep-wake detection using wearable motion sensor data. The project was designed to create transparent, reproducible methods that could be applied across different wearable devices and validated against gold-standard polysomnography.

The framework includes feature extraction from high-frequency accelerometer and gyroscope signals, capturing movement patterns and postural changes that distinguish sleep from wake states. TNO developed machine learning models trained on motion data paired with manually scored polysomnography recordings, ensuring that the algorithms learned clinically meaningful sleep-wake patterns. The development process emphasised transparency and reproducibility, with clearly documented methods that enable independent validation and adaptation by other researchers or developers.

TNO led the scientific coordination and validation activities, conducting rigorous testing against polysomnography data to assess algorithm performance. The validation process examined sensitivity for correctly identifying sleep periods, specificity for detecting wake episodes and overall accuracy across different sleep stages and patient populations. The project delivered a validated sleep detection algorithm with good sensitivity for sleep and moderate specificity for wake, alongside a reusable feature extraction pipeline that other researchers can apply to their own datasets. A deployable tool was created, providing a robust basis for further clinical validation and real-world implementation.

Looking to the future

The standardised framework developed in SMARTsleep enables broader application of validated sleep monitoring across research and clinical contexts. Future work will focus on improving wake detection specificity, which remains challenging due to the heterogeneity of wake behaviours during the night. Integration with additional sensor modalities, such as heart rate or skin temperature, may enhance algorithm performance. The device-agnostic approach facilitates deployment across multiple wearable platforms, increasing accessibility for diverse patient populations and research settings. Collaboration with medical device manufacturers, sleep medicine centres and health technology assessment bodies will support translation into clinical practice, enabling scalable, cost-effective sleep monitoring for conditions where repeated assessment is valuable, such as insomnia, sleep apnoea and circadian rhythm disorders.