Artificial intelligence accurately charts sleep stages without intrusive brain sensors

Researchers have developed an artificial intelligence model capable of tracking a person’s sleep stages using only three simple physical measurements. The system analyzes heart rate, blood oxygen levels, and abdominal breathing to chart sleep second by second without requiring invasive brain wave sensors. The findings, published in the Journal of Sleep Research, offer a pathway for making clinical sleep tracking more comfortable and accessible for patients at home.

Resting is an active biological process characterized by distinct neurological and physical phases. When people drift off, they cycle through periods of light sleep, deep sleep, and rapid eye movement sleep. Each of these phases serves a distinct biological purpose, ranging from physical restoration to the consolidation of new memories.

Accurately charting these transitions is essential for diagnosing serious medical conditions like insomnia and sleep apnea. The traditional method for evaluating these phases requires a procedure known as polysomnography. This test usually requires a patient to spend the night in a specialized clinic while technicians attach multiple wires and sensors to their body.

A primary component of this traditional evaluation is the electroencephalogram, or EEG. This device tracks electrical activity in the brain through sticky electrodes applied to the scalp. While this brain tracking provides incredibly accurate information, the bulky setup interrupts normal resting patterns and generates long waiting lists at specialized clinics.

Lead author Ángel Serrano Alarcón of Reutlingen University and the Universidad de Sevilla set out to resolve this logistical bottleneck. Alarcón worked alongside colleagues from HTWG Konstanz to design an automated sleep tracking method. They aimed to build a system that relies exclusively on sensors that a patient could comfortably wear in their own bed.

Previous attempts to automate this process using artificial intelligence have faced notable roadblocks. Many prior algorithms still relied on raw brain wave data, requiring the same cumbersome scalp sensors. Other models utilized simpler sensors but lacked transparency, as engineers designed them through blind trial and error without establishing reproducible guidelines.

Another major limitation of earlier algorithms involved how they portioned out time. Most automated clinical systems divide a patient’s night into arbitrary 30-second windows and assign a single sleep phase to that entire block. This averaging method can easily miss tiny disruptions or brief awakenings that occur within a given half-minute interval.

To overcome these hurdles, the research team focused on evaluating just three specific physiological markers. They chose to track oxygen saturation in the blood, fluctuations in the resting heart rate, and the physical expansion of the abdomen during breathing. These specific data points can be gathered effortlessly using consumer-grade health rings or basic chest straps.

The researchers sourced their primary training data from the Sleep Heart Health Study. They extracted the relevant sensor readings for 855 subjects from this existing medical database. The scientists also gathered the official medical annotations for those patients to teach their algorithm how the three physical signals corresponded to clinical sleep stages.

The engineering team utilized an artificial intelligence framework known as a U-Net for their primary algorithm. Computer scientists originally designed this specific type of deep learning model to separate and classify distinct objects within photographs. In this novel application, the engineers adapted the structural logic of the U-Net to interpret a linear, eight-hour timeline of bodily signals.

The researchers avoided manual guesswork by employing an automated tuning program to optimize the internal mechanics of the software. This optimization tool tested thousands of mathematical configurations to find the most accurate setup. Establishing this rigid, mathematical approach ensures that other software engineers can independently reproduce the model for future medical studies.

Clinicians typically divide the human resting cycle into five distinct categories. These include wakefulness, rapid eye movement sleep, and three progressively deeper stages of dreamless rest denoted as N1, N2, and N3. The research team instructed their tuned algorithm to categorize the training data using either the full five-stage clinical model or a condensed four-stage model.

In the four-stage framework, the algorithm grouped the N1 and N2 phases together under the generalized umbrella of light sleep. When evaluating this four-stage setup against the primary dataset of 855 patients, the deep learning model correctly identified the exact sleep phase with 71 percent accuracy. The system performed particularly well when distinguishing periods of wakefulness and rapid eye movement sleep.

Checking a neural network against its own training material can sometimes produce overly optimistic results. To prove the model had actually learned the underlying biological patterns, the researchers tested it against a completely separate medical database. They fed the algorithm sensor data from 931 subjects enrolled in the Multi-Ethnic Study of Atherosclerosis.

When confronted with this entirely new group of patients, the model proved highly resilient. It maintained a 66 percent accuracy rate for the four-stage classification on the unseen data. The algorithm achieved a 65 percent accuracy rate for the five-stage classification on the initial dataset, and it actually improved to a 68 percent accuracy rate on the external testing dataset.

A notable operational success of the resulting software was its exceptionally high resolution. Instead of lumping data into 30-second blocks, the algorithm produces a distinct sleep stage prediction for every single second of the night. This granular timeline mirrors the natural fluidity of the human body.

The researchers anticipate that this high-resolution screening will eventually allow doctors to visualize microscopic sleep disruptions. Brief physiological arousals that are routinely swallowed by traditional averaging methods become clearly visible on a second-by-second timeline. This detailed perspective helps doctors cross-reference tiny awakenings with other bodily disturbances like isolated drops in blood oxygen.

Despite these advances, the current model still struggles with a few specific classification tasks. The algorithm had the most difficulty identifying the N1 phase of sleep. This specific category describes the very brief, transitional moment when a person first begins drifting off.

Because the N1 transition is usually fleeting, it makes up a very minor fraction of the available medical data. The algorithm has comparatively fewer examples to study, making it harder for the neural network to differentiate this delicate transition from general wakefulness. The software also occasionally confused the biological signatures of light sleep and deep sleep.

The system is currently constrained by sheer computing power. To streamline the training process, the engineers restricted the algorithm to analyzing exactly eight hours of sleep at a time. Scientists will need to update the model to handle natural variations in resting durations, as some patients might sleep for six hours while others sleep for ten.

The researchers also applied minimal filtering to the raw sensor data before feeding it into the model. Real-world physical environments contain a multitude of electrical or mechanical disruptions. Moving forward, engineers will need to test how the algorithm handles sudden sensor disconnections or intense body movements that temporarily corrupt the signal.

Transitioning from the laboratory to the home remains the ultimate goal for automated sleep monitoring. Validating algorithms on massive arrays of physical data narrows the gap between cumbersome clinical arrays and accessible wearable technology. Future refinements to this AI framework may eventually allow patients to receive clinical-grade sleep evaluations entirely from the comfort of their own beds.

The study, “Optimising Sleep Stage Detection Using a Minimal Non-EEG Physiological Signal Set and Deep Learning,” was authored by Ángel Serrano Alarcón, Maksym Gaiduk, Natividad Martínez Madrid, Juan Antonio Ortega, and Ralf Seepold.