Anchor-free temporal localization of apnea events from EEG/EOG with state-space models

Abstract

Objective. Reduced-channel polysomnography (PSG) and electroencephalography/electrooculography (EEG/EOG) signals can support obstructive sleep apnea (OSA) screening, but many learning-based systems use coarse epoch-level classification and generalize poorly across datasets. We introduce ApneaTime, an anchor-free temporal localization framework for apnea/hypopnea boundary detection from EEG/EOG with auxiliary sleep-stage prediction.
Approach. ApneaTime combines multi-resolution time–frequency convolutional encoding, a state-space sequence backbone, and an anchor-free center-offset head for variable-duration respiratory events. Self-supervised pretraining, weak supervision, domain-adversarial learning, and contrastive/prototype regularization are used to improve robustness under limited event labels and cohort shift.
Main results. In in-dataset evaluations, ApneaTime improved event F1 from 70.4% to 78.5%, mean average precision from 75.8% to 84.5%, and mean intersection-over-union from 0.47 to 0.56 compared with the recurrent baseline. Under SHHS-to-MESA transfer,
F1 increased from 60.2% to 70.1% and mean average precision from 65.5% to 78.0%. Expected calibration error decreased from 9.8% to 5.3%, and to 2.1% after temperature scaling. The model has 8.2 million parameters and achieved a GPU real-time factor of approximately 1.5.
Significance. ApneaTime provides calibrated event-level apnea/hypopnea localization from reduced-channel EEG/EOG recordings. The results support EEG/EOG-based sleep apnea monitoring when full respiratory PSG channels are unavailable or impractical.