LAD-TCNM: A DEEP LEARNING MODEL FOR WEARABLE ECG-BASED SLEEP APNEA DETECTION

Obstructive sleep apnea (SA) remains difficult to monitor in maritime environments, where seafarers are exposed to prolonged voyages, rotating shift schedules, and persistent mechanical noise and vibration. These conditions not only disrupt sleep patterns but also significantly degrade physiological signal quality, rendering conventional polysomnography impractical for continuous screening. To address these challenges, this study collects wearable single-lead ECG data from seafarers and proposes a tailored deep learning framework, termed lightweight attention domain-adaptive temporal convolutional network model (LAD-TCNM), for robust and efficient SA detection in shipboard settings. The proposed framework integrates multi-scale cardiopulmonary feature fusion with a residual temporal convolutional network (TCN) backbone to effectively capture apnea-related temporal dynamics from ECG-derived signals. Specifically, heterogeneous features including heart rate variability (HRV), ECG-derived respiration (EDR), and QRS complex information are jointly modeled across multiple temporal resolutions to enhance physiological representation. To further improve discriminative capability, a lightweight dual attention mechanism combining temporal attention (TA) and channel attention (CA) is incorporated, enabling the model to adaptively emphasize informative time segments and feature channels. Moreover, to address cross-subject and cross-environment variability inherent in wearable ECG acquisition, a domain-adversarial training strategy based on a gradient reversal layer (GRL) is employed to align feature distributions across different users and acquisition conditions. Beyond algorithmic design, the end-to-end system explicitly considers practical shipboard constraints, including low-power operation, intermittent network connectivity with offline buffering and batch uploading, as well as motion-and noise-induced artifacts. Comprehensive experiments conducted on a self-collected seafarer dataset and the public Apnea-ECG benchmark demonstrate that LAD-TCNM consistently outperforms strong baseline methods in terms of accuracy, sensitivity, specificity, and cross-recording robustness. These results highlight the feasibility of continuous, noninvasive, and reliable SA screening using wearable ECG devices in maritime environments, providing a practical solution for large-scale sleep health monitoring among seafarers where conventional clinical diagnostics are difficult to deploy.