Validation of a deep learning-based system for sleep staging, arousal detection, and respiratory event scoring in...

Obstructive sleep apnea is a common yet underdiagnosed sleep disorder in patients with atrial fibrillation (AF). Although polysomnography (PSG) is the gold standard, its costs and complexity limit routine use. Home sleep apnea testing (HSAT) is a more accessible alternative, but it lacks electroencephalography (EEG), limiting direct assessment of sleep stages and arousals. This study evaluates the performance of DeepRESP, an automated scoring solution that infers sleep stages and arousals using respiratory inductance plethysmography (RIP) signals and scoring respiratory events, in a cohort of patients with AF.

In this prospective cohort study, consecutive AF patients underwent ambulatory type II PSG. Automated respiratory event scoring, sleep staging and arousal identification were performed with DeepRESP using only HSAT signals. Performance metrics included epoch-level positive and negative percent agreement, complemented by concordance analyses for the apnea-hypopnea index (AHI), arousal index (ArI), and total sleep time (TST), using Bland-Altman analysis and intraclass correlation coefficients (ICC), with manually scored PSG as reference.

A total of 88 patients (67.0% male) were included. Women were older than men (66.7 ± 7.2 vs. 61.27 ± 10.3 years, p = 0.015). Epoch-level overall percent agreement (OPA) with manual scoring was 0.91 for Wake, 0.95 for rapid eye movement sleep (REM), and 0.87 for non-rapid eye movement sleep (NREM). Arousal detection OPA was 0.80. Strong agreement was observed for the apnea-hypopnea index (ICC = 0.92), TST (ICC = 0.82), and ArI (ICC = 0.83).

In AF patients, DeepRESP reliably estimates sleep architecture, arousals, and sleep apnea severity from HSAT signal, offering a scalable, EEG-independent diagnostic approach.

Obstructive sleep apnea is highly prevalent and often underdiagnosed in atrial fibrillation patients, yet standard home testing lacks the neurophysiological channels necessary to detect arousals and accurate sleep architecture. This study was conducted to validate a deep learning-based system capable of inferring these critical metrics exclusively from home sleep testing signals, improving diagnostics performance within this specific cardiovascular population.

The findings demonstrate that artificial intelligence-driven analysis of breathing patterns achieves high concordance with polysomnography for key diagnostic indices, including arousal index, total sleep time and apnea severity via the apnea-hypopnea index. This provides a scalable, precise diagnostic pathway that bypasses the confounding effects of arrhythmias and cardiac medications, facilitating improved management of sleep-disordered breathing in atrial fibrillation patients.