In many applications, recognizing the depth of sleep (e.g., light, deep, REM sleep) while the subject is sleeping enables innovative features. For instance, in SAE Level 4 autonomous driving, a driver may need to takeover the vehicle control in case the autopilot is exiting its operational design domain. Depending on the depth of the sleep, the subject may need time to takeover effectively; hence, it is particularly relevant to know in which sleep stage the subject is (e.g., light sleep, deep sleep, and REM sleep), and possibly initiate actions to prevent the subject to remain in those sleep stages that lead to longer takeover time. Sleep stage classification can be achieved through an on-the-fly algorithm, which generates output in response to each input portion without knowledge of future inputs, unlike an off-Line algorithm that provides output just after receiving the entire input sequence. Various studies have analyzed algorithms or devices that identify sleep stages during the night; however, these typically require electroencephalography (EEG), which is obtrusive, or specialized devices. This study describes the development of an on-the-fly sleep-scoring algorithm using Heart Rate (HR), RR intervals, which is the distance between two consecutive heartbeats, and accelerometer data from a smartwatch, widespread, non-invasive, and affordable but accurate device. The subjects involved in our study wore a commercial off-the-shelf wearable device during a full night’s sleep, and were also monitored using a reference medical device to establish the ground truth by means of a full polysomnography (PSG) analysis. The on-the-fly sleep scoring algorithm based on smartwatch data was tested against PSG-based scoring, achieving 88.46% accuracy, 91.42% precision, and 93.52% sensitivity in sleep–wake identification. Deep sleep was correctly identified 69.38% of times, light sleep in 50.62%, REM sleep 62.02% and wakefulness 73.48% of times.

Drowsy driving contributes to 10–30% of all vehicle crashes, making it a major road safety concern. Driver Monitoring Systems (DMS) aim to assess driver alertness and are typically categorized into vehicle-, behavior-, and physiology-based approaches. While physiology-based systems offer the highest accuracy, most of them rely on costly and intrusive contact-based cardiac sensors. This study demonstrates, for the first time, the feasibility of predicting driver sleep events using a fully contactless, physiology-based approach that analyzes breathing patterns in real time. Data were collected using a short-range 60 GHz standalone automotive radar in a driving-seat mockup. Crucially, sleep events were objectively validated for the first time in this context using polysomnography (PSG) data reviewed by a medical expert, following American Academy of Sleep Medicine (AASM) guidelines for the Maintenance of Wakefulness Test (MWT) — marking a departure from previous reliance on subjective behavioral observations. The proposed heuristic algorithm achieved 85% overall accuracy, with 100% (95% CI 29%–100%) specificity and 80% (95% CI 44%–97%) sensitivity. This work presents a validated, non-intrusive solution for sleep event prediction in drivers, underscoring its potential for enhancing road safety through practical, clinically supported DMS technologies.

The last decade's market has been characterized by wearable devices, mainly smartwatches, edge, and cloud computing. A possible application of these technologies is to improve the safety of dangerous activities, especially driving motor vehicles. Common enabling technologies, such as system-on-chip, ultra-low-power computational platforms, and wide-band wireless connectivity, push all these trends. On the one hand, wearable devices, thanks to the continuous contact with the user's body, can measure physiological parameters. On the other hand, edge computing and machine learning techniques, alongside cameras, allow the implementation of contactless computer vision systems capable of providing information about the user's current behavior. Another trend is the usage of RADARs in automotive applications, both for collision avoidance and monitoring driver behavior. These technologies can be combined to develop systems designed to aid the driver. For the sake of this paper, we are focusing on warning drivers, allowing them to know whenever they are drowsy and hence risking a sleep onset or are not paying attention to the road. Developing such systems poses many challenges, such as automatic classification of physiological signal patterns, facial expression recognition, head movements and eye gaze detection. These challenges have been individually addressed in the literature. Anyway, we noticed a need for more description on implementing data fusion. Two main reasons for adopting the fusion approach are to improve the quality of the overall representation (increasing accuracy and specificity against drowsy) and make a more reliable system due to redundancy.