AI Reads Breathing Alone to Diagnose Sleep Apnea in Heart Rhythm Patients

For millions of people living with atrial fibrillation, a dangerous and often hidden companion goes undetected night after night. Obstructive sleep apnea is extraordinarily common in patients with this irregular heart rhythm, yet it frequently escapes diagnosis because the symptoms of the two conditions blur into one another. Fatigue, breathless nights, and fragmented sleep are routinely blamed on the heart alone, while the true culprit in the bedroom goes unnoticed. Now, a prospective validation study published in the Journal of Clinical Sleep Medicine offers a striking glimpse of how artificial intelligence could close this diagnostic gap, showing that a deep learning system can reconstruct the essential features of a full sleep study from breathing signals alone, without ever touching the brain electrodes that sleep medicine has relied on for decades.

The study, led by Susana Sousa of CUF Tejo Hospital in Lisbon together with collaborators at Nox Medical in Iceland and the University of Porto, set out to test a cloud-based software called DeepRESP in a population where conventional home testing struggles most. The system had already earned FDA 510(k) clearance after validation on nearly 3,500 routine clinical recordings from sleep clinics across the United States, demonstrating non-inferiority or superiority to established predicate devices. But the atrial fibrillation population poses a unique challenge. Irregular heart rhythms and rate-controlling medications such as beta-blockers scramble the cardiac signals that many simplified sleep-staging algorithms depend on, potentially undermining their accuracy precisely in the patients who need reliable testing the most.

The researchers enrolled 88 consecutive patients with atrial fibrillation referred for sleep assessment at a cardiology outpatient unit, excluding only those with prior sleep apnea diagnoses or recent acute cardiac events. Two-thirds of the participants were men, and the women were significantly older than the men, averaging nearly 67 years compared with about 61. The cohort carried a heavy burden of comorbidity: half were obese, nearly half had hypertension, and roughly one in five had diabetes. Most telling of all, every single participant met criteria for sleep apnea, with an apnea-hypopnea index of at least five events per hour, and 80 percent had moderate to severe disease. Yet only about a quarter reported clinically significant daytime sleepiness, underscoring how easily the condition hides in this cardiac population.

Each patient underwent ambulatory level II polysomnography with a full complement of physiological channels, including six-lead electroencephalography, electrooculography, submental electromyography, nasal pressure airflow, thoracic and abdominal respiratory inductance plethysmography, oximetry, and body position sensing. Certified sleep physicians and technicians manually scored these recordings according to American Academy of Sleep Medicine criteria, establishing the gold-standard reference. DeepRESP then processed the same recordings, but crucially it was allowed to see only the subset of signals available in a home sleep apnea test: oximetry, nasal pressure airflow, and the respiratory inductance plethysmography belts. No electroencephalography, no eye movement channels, no muscle electrodes, and no human editing of the automated output.

The technical heart of the system is the Nox BodySleep 2.0 algorithm, which infers sleep states and arousals exclusively from the thoracic and abdominal breathing belts sampled at 25 hertz. The underlying physiology is elegant. Breathing is irregular and behaviorally influenced during wakefulness, becomes remarkably stable and metabolically regulated in non-REM sleep, and turns erratic again in REM sleep, when skeletal muscle atonia reshapes thoracoabdominal mechanics. These state-dependent signatures are imprinted on the plethysmography signals, and the neural network learns to decode them. Arousals add a further fingerprint: each one triggers a transient ventilatory response, a gasp-like perturbation in the breathing pattern that the algorithm can detect and quantify, enabling it to score hypopneas that end in arousal even when oxygen saturation barely dips.

The results were remarkable for a system flying blind without brain waves. At the epoch level, overall agreement with manual scoring reached 0.91 for wake, 0.95 for REM sleep, and 0.87 for non-REM sleep. Arousal detection achieved an overall percent agreement of 0.80, with a positive percent agreement of 0.72 and a negative percent agreement of 0.84. Respiratory event detection was even stronger, with overall agreement of 0.94 for apneas, 0.78 for hypopneas, and 0.81 for respiratory events combined. Bootstrapping with 10,000 iterations generated confidence intervals for every metric, and the estimates held tightly across resamples.

At the level of the clinical indices that actually drive treatment decisions, the concordance was equally persuasive. The apnea-hypopnea index, the single most important number in sleep medicine, showed an intraclass correlation coefficient of 0.92 against manual polysomnography, with Bland-Altman analysis revealing a mean bias of 6.36 events per hour and limits of agreement stretching from minus 6.83 to 19.56. That bias was driven almost entirely by the hypopnea index, which contributed a bias of 5.45 events per hour, while the apnea index remained highly stable at just 0.91 events per hour. Total sleep time correlated at 0.82 with a mean bias of about 18.5 minutes, and the arousal index reached an intraclass correlation of 0.83 with a low mean bias of 1.44, although individual variance widened at higher arousal frequencies.

Why does this matter beyond the sleep laboratory? Untreated sleep apnea is a recognized saboteur of atrial fibrillation management. It increases recurrence after cardioversion and catheter ablation, blunts the efficacy of antiarrhythmic drugs, and elevates cardiovascular mortality through intermittent hypoxemia, sympathetic activation, and structural cardiac remodeling. Yet conventional home sleep apnea tests, lacking electroencephalography, systematically underestimate disease in patients whose hypopneas terminate in arousals rather than significant desaturation, and in those whose recorded sleep duration is short. A recent evaluation of peripheral arterial tonometry-based home testing in atrial fibrillation patients found only slight to fair agreement with polysomnography for severity classification, with a tendency to overestimate disease. By contrast, the breathing-based approach sidesteps the confounding effects of arrhythmia and cardiac medication entirely, because respiratory effort signals are robust to the electrical chaos of a fibrillating heart.

The authors are candid about the caveats. The cohort was modest in size and drawn from a single tertiary cardiology center, raising the possibility of referral bias. Although recordings were ambulatory, they were acquired as level II polysomnography under technologist supervision, so respiratory belt quality may have exceeded what unattended home testing typically delivers, where sensor displacement and suboptimal positioning are common. The arousal index showed wider limits of agreement than the other parameters, suggesting that breathing-derived arousal estimates, while clinically informative, may not fully replace electroencephalography for fine-grained characterization of sleep fragmentation. Comorbidities such as heart failure, Cheyne-Stokes respiration, and hypoventilation syndromes can produce breathing patterns that mimic obstructive events, and the system has not been assessed in patients already using respiratory support. Nor does DeepRESP score respiratory effort-related arousals.

Even with these limitations, the study marks a genuine inflection point for precision medicine in sleep diagnostics. It challenges the long-standing assumption that neurophysiological channels are indispensable for reliable sleep architecture assessment, at least in the populations where home testing is needed most. A scalable, electroencephalography-independent pathway that reproduces gold-standard indices for apnea severity, total sleep time, and arousal burden could transform screening for the vast, underdiagnosed population of atrial fibrillation patients, many of whom are minimally symptomatic and would otherwise never reach a sleep laboratory. The authors call for validation in fully unattended home recordings and for studies of long-term impact on clinical management and cardiovascular outcomes. If those efforts succeed, the humble breathing belt, read by a neural network, may become the stethoscope of twenty-first-century sleep medicine.

Subject of Research: Validation of a deep learning system for EEG-independent sleep staging, arousal detection, and respiratory event scoring in atrial fibrillation patients

Article Title: Validation of a deep learning-based system for sleep staging, arousal detection, and respiratory event scoring in atrial fibrillation patients

Article References: Sousa, S., Teixeira, C., Sigmarsdóttir, T., Finnsson, E., Ágústsson, J., Sigþórsson, S., Bjarkason, S., Drummond, M., & Bugalho, A. (2026). Validation of a deep learning-based system for sleep staging, arousal detection, and respiratory event scoring in atrial fibrillation patients. Journal of Clinical Sleep Medicine, 22(1), Article 130. https://doi.org/10.1007/s44470-026-00145-0

Image Credits: AI Generated

DOI: 10.1007/s44470-026-00145-0

Keywords: deep learning, atrial fibrillation, obstructive sleep apnea, home sleep apnea testing, polysomnography, respiratory inductance plethysmography, sleep staging, arousal detection, apnea-hypopnea index, machine learning, sleep-disordered breathing, clinical validation