Researchers have developed an artificial intelligence model that can track a person’s sleep stages using just three simple body measurements. The system analyzes your heart rate, blood oxygen levels, and abdominal breathing, charting your sleep down to the second without the need for invasive brainwave sensors. The findings, published in the Journal of Sleep Research, provide a path to making clinical sleep tracking more comfortable and accessible to patients at home.
Rest is an active biological process characterized by distinct neurological and physical stages. When a person loses consciousness, they cycle through light sleep, deep sleep, and rapid eye movement sleep. Each of these steps serves a distinct biological purpose, from physical repair to consolidation of new memories.
Accurately charting these changes is essential for diagnosing serious medical conditions such as insomnia and sleep apnea. Traditional methods of evaluating these stages require a procedure known as polysomnography. The test typically requires patients to spend one night in a specialized clinic, during which technicians attach multiple wires and sensors to their body.
The main component of this traditional assessment is the electroencephalogram, or EEG. The device tracks electrical activity in the brain through adhesive electrodes applied to the scalp. Although this brain tracking provides incredibly accurate information, the extensive setup disrupts normal rest patterns and results in long waiting lists at specialized clinics.
Lead author Angel Serrano Alarcón from the Universities of Reutlingen and Seville set out to solve this logistical bottleneck. Alarcón worked with colleagues at HTWG Constance to design an automated sleep tracking method. They aimed to build a system that relied solely on sensors that patients could wear from the comfort of their own beds.
Previous attempts to automate this process using artificial intelligence have faced notable obstacles. Many previous algorithms still relied on raw brainwave data and required similar unwieldy scalp sensors. Other models utilized simpler sensors but lacked transparency as engineers designed them blindly by trial and error without establishing reproducible guidelines.
Another major limitation of previous algorithms involved how time was distributed. Most automated clinical systems divide a patient’s night into arbitrary 30-second windows and assign one sleep phase to the entire block. This averaging method can easily miss small interruptions or brief awakenings that occur within a given 30-minute interval.
To overcome these hurdles, the research team focused on assessing only three specific physiological markers. They decided to track oxygen saturation in the blood, variations in resting heart rate, and the physical expansion of the abdomen during breathing. These specific data points can be easily collected using consumer health rings or basic chest straps.
Researchers obtained primary training data from the Sleep Heart Health Study. They extracted sensor readings related to 855 subjects from this existing medical database. The scientists also collected official medical annotations for these patients to teach the algorithm how the three physical signals correspond to clinical sleep stages.
The engineering team utilized an artificial intelligence framework known as U-Net as the primary algorithm. Computer scientists originally designed this particular type of deep learning model to separate and classify different objects in photos. For this new application, engineers adapted U-Net’s structural logic to interpret a linear 8-hour timeline of body signals.
The researchers avoided manual guesswork by employing an automatic tuning program to optimize the software’s inner workings. This optimization tool has tested thousands of mathematical configurations to find the most accurate settings. Establishing this rigorous mathematical approach will allow other software engineers to independently reproduce the model for future medical research.
Clinicians typically classify human rest cycles into five different categories. These include three progressively deeper stages of wakefulness, rapid eye movement sleep, and dreamless rest, designated N1, N2, and N3. The researchers instructed the tuned algorithm to classify the training data using either the full five-stage clinical model or the compressed four-stage model.
In the four-stage framework, the algorithm grouped the N1 and N2 phases under the generalized umbrella of light sleep. When this four-stage configuration was evaluated against a primary dataset of 855 patients, the deep learning model correctly identified the correct sleep stage with 71% accuracy. The system performed particularly well when distinguishing between periods of wakefulness and periods of rapid eye movement sleep.
Matching a neural network against its own training material may yield overly optimistic results. To prove that the model was indeed learning the underlying biological patterns, the researchers tested it against a completely separate medical database. They input the algorithm sensor data from 931 subjects enrolled in the Multi-Ethnic Study of Atherosclerosis.
This model has proven to be highly resilient when faced with this entirely new group of patients. The accuracy of four-step classification on unseen data remained 66%. The algorithm achieved 65% accuracy on the 5-step classification on the initial dataset and actually improved to 68% accuracy on the external test dataset.
A notable operational success of the resulting software was its exceptionally high resolution. Rather than aggregating data into 30-second blocks, the algorithm generates a clear sleep stage prediction for every second of the night. This granular timeline reflects the natural fluidity of the human body.
Researchers predict that this high-resolution screening will eventually allow doctors to visualize subtle sleep disorders. Short-term physiological arousal, routinely swallowed up by traditional averaging techniques, becomes clearly visible on a second-by-second timeline. This detailed view helps doctors cross-reference minor arousals with other physical disturbances, such as isolated drops in blood oxygen.
Despite these advances, current models still struggle with some specific classification tasks. The algorithm had the most difficulty identifying the N1 stage of sleep. This particular category represents a very brief transitional moment when a person first begins to drift.
The transition to N1 is typically temporary and therefore accounts for a small portion of the available medical data. The algorithm studies relatively few examples, making it difficult for the neural network to distinguish this subtle transition from a general state of wakefulness. The software also sometimes confused the biological characteristics of light and deep sleep.
Currently, systems are constrained by enormous computing power. To streamline the training process, engineers limited the algorithm to analyzing exactly eight hours of sleep at a time. Scientists need to update their models to account for natural variations in rest time, as some patients may sleep for 6 hours while others sleep for 10 hours.
The researchers also applied minimal filtering to the raw sensor data before feeding it into the model. Real-world physical environments contain many electrical and mechanical perturbations. Going forward, engineers will need to test how the algorithm handles sudden sensor disconnections and violent body movements that temporarily disrupt the signal.
Moving from the lab to the home remains the ultimate goal for automated sleep monitoring. Validating algorithms against vast arrays of physical data narrows the gap between unwieldy clinical arrays and accessible wearable technology. Future improvements to this AI framework could eventually allow patients to receive a full clinical-grade sleep assessment from the comfort of their own bed.
The study, “Optimizing sleep stage detection using a minimal non-EEG physiological signal set and deep learning,” was authored by Angel Serrano Alarcón, Maxime Gaiduk, Nativida Martínez Madrid, Juan Antonio Ortega, and Ralph Seepold.
