Breathing Patterns Alone: How AI Cracked Sleep Apnea Detection in the Toughest Patient Population
FDA-cleared deep learning system achieves 92% correlation with gold-standard sleep studies in atrial fibrillation patients using only respiratory signals—no brain electrodes required.
Diagnosing sleep apnea in patients with irregular heart rhythms has long presented a clinical catch-22: the very cardiac arrhythmias that make sleep testing most urgent also render conventional home monitoring unreliable. Now, a prospective validation study published in the Journal of Clinical Sleep Medicine demonstrates that DeepRESP, a cloud-based artificial intelligence system from Nox Medical, can accurately identify obstructive sleep apnea in atrial fibrillation patients using breathing signals alone—bypassing the brain electrodes traditionally required for sleep studies and the cardiac signals that fail when heart rhythms turn chaotic.
Validation in a Diagnostic Blind Spot
The study enrolled 88 consecutive atrial fibrillation patients at CUF Tejo Hospital in Lisbon, led by Dr. Susana Sousa with collaborators from Nox Medical in Iceland and the University of Porto. The results revealed a striking diagnostic gap: every single patient met criteria for sleep apnea, with 80% showing moderate-to-severe disease, yet only approximately 25% reported clinically significant daytime sleepiness—the symptom that typically triggers referral for sleep testing.
DeepRESP's Nox BodySleep 2.0 algorithm achieved 0.92 intraclass correlation for apnea-hypopnea index compared to gold-standard polysomnography, using only thoracic and abdominal respiratory inductance plethysmography belts, nasal pressure airflow, and oximetry. The system requires no EEG, EOG, or EMG channels, dramatically simplifying the testing setup while maintaining accuracy in a population where cardiac signal-based algorithms typically struggle.
How Breathing Reveals Sleep Architecture
The technology exploits fundamental physiological patterns: breathing becomes irregular during wakefulness, stable and metabolically regulated during non-REM sleep, and erratic during REM sleep due to muscle atonia. Arousals trigger transient ventilatory responses detectable in plethysmography signals sampled at 25 Hz.
Sleep staging accuracy reached epoch-level agreement of 0.91 for wake, 0.95 for REM, and 0.87 for non-REM sleep versus manual polysomnography scoring. Arousal detection achieved overall percent agreement of 0.80, while respiratory event detection showed overall agreement of 0.94 for apneas and 0.78 for hypopneas. Total sleep time correlation reached 0.82 with a mean bias of 18.5 minutes, and arousal index correlation hit 0.83.
The patient cohort—67% male, with mean ages of 61 years for men and 67 years for women—reflected typical atrial fibrillation demographics, with 50% obese, 50% hypertensive, and 20% diabetic.
Implications for Cardiology and Sleep Medicine
The validation addresses a critical gap in managing the estimated 33 million atrial fibrillation patients worldwide. Sleep apnea sabotages atrial fibrillation management by increasing recurrence after cardioversion and ablation, blunting antiarrhythmic drug efficacy, and elevating cardiovascular mortality. Yet conventional home sleep tests often fail in this population due to irregular heart rhythms and beta-blocker medications that confound cardiac signal-dependent algorithms.
Recent peripheral arterial tonometry-based home testing in atrial fibrillation patients showed only slight-to-fair agreement with polysomnography and a tendency to overestimate disease, according to the study context. The breathing-based approach circumvents these limitations entirely.
DeepRESP holds FDA 510(k) clearance based on validation in approximately 3,500 routine clinical recordings from U.S. sleep clinics. For patients already managing complex medication regimens and frequent cardiology visits, a single-night home test using only breathing belts represents a dramatically lower burden than overnight laboratory stays with full electrode arrays—potentially eliminating months-long wait times and geographic barriers while accelerating treatment initiation for a condition that directly impacts heart rhythm control.
Sources
Reporting & accountability
How we maintain accuracy
Health AI Central separates reported facts, analysis, and public-source data. Readers can review our standards, request a correction, or consult the relevant research methodology before relying on a claim.