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) may support obstructive sleep apnea (OSA) screening, but learning-based systems rely on epoch-level classification and generalize poorly across datasets. We introduce ApneaTime, an anchor-free temporal localization framework for apnea/hypopnea boundary estimation 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 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. Evaluation includes event-level localization, subject-level apnea-hypopnea index (AHI) estimation, OSA severity stratification, nightly event-count agreement, calibration, and cross-dataset transfer.Main results.In in-dataset evaluation, ApneaTime improved event1 fromto, mean average precision fromto, and mean intersection-over-union fromtocompared with a CNN+LSTM baseline. Under SHHS-to-MESA transfer, event1 increased fromtoand mean average precision fromto. On the SHHS held-out test set, AHI mean absolute error wasevents/h with a mean bias ofevents/h. Severity stratification achievedaccuracy,macro-1, and weighted, while nightly event-count mean absolute deviation wasevents. Under transfer, AHI mean absolute error wasevents/h and severity accuracy was. Expected calibration error decreased fromfor CNN+LSTM tofor ApneaTime andafter temperature scaling. The model containedmillion parameters and achieved areal-time factor on an NVIDIA Tesla V100 GPU.Significance.ApneaTime supports calibrated, screening-oriented event localization and subject-level burden estimation from reduced-channel EEG/EOG. Limitations include retrospective evaluation on PSG cohorts, dataset-dependent channel availability, no prospective validation against respiratory PSG, and no embedded portable-hardware testing; therefore, the findings do not establish diagnostic or deployment readiness.