A domain adversarial network for small-sample obstructive sleep apnea detection using a single-lead piezoelectric sensor.

Objective.In recent years, increasing attention to sleep health has accelerated development of wearable devices for home monitoring. Two critical challenges remain: providing an unobtrusive sensing solution suitable for long-term deployment, and designing algorithms that generalize across diverse users and acoustic environments when annotated data are limited.Approach.We develop a single-lead bone-conduction monitoring system and a distribution-aware detection model. The wearable system uses a piezoelectric sensor at the lateral throat to capture tracheal vibrations for continuous overnight monitoring, which improves user comfort relative to multi-lead polysomnography (PSG). For robust event recognition, we propose the convolutional memory adaptive prototypic network (CMAP), which combines convolutional-recurrent feature extraction with domain-adversarial training and a prototypical constraint to learn domain-invariant representations under limited labels. We evaluated our solution with a cohort of 95 participants undergoing simultaneous PSG and wearable monitoring.Main results.For snore and breath sound event detection during sleep, the model attained precision of 0.90 and recall of 0.93. The estimated Apnea-Hypopnea Index (AHI) from our system correlated strongly with PSG-derived AHI (Pearsonr= 0.94,p< 0.001). The model achieved 83.2% accuracy in four-class severity classification across healthy, mild, moderate and severe obstructive sleep apnea (OSA).Significance.These results indicate that a single-sensor, distribution-aware approach can provide a reliable acoustic surrogate for AHI and clinically relevant severity classification. The proposed wearable and CMAP model offer a practical, comfortable, and cost-effective solution for early clinical triaging, scalable home-based OSA screening and longitudinal monitoring with low user burden.