Continuous Event Modeling for Sleep Apnea Detection

Sleep apnea is a breathing disorder caused by the obstruction of the upper airway during sleep which affects approximately 80 million Americans, but 80% of the cases remain undiagnosed. Overnight polysomnography (PSG) is the gold-standard method of diagnosis and is used to compute the apnea-hypopnea index (AHI), but its high cost and intrusiveness limit its suitability for continuous monitoring. This motivates the development of computational methods that leverage wearable and implantable sensor data for longitudinal assessment of sleep apnea. Although wearable sensors offer a scalable alternative for ambulatory monitoring [1], the signals collected from devices such as accelerometers and electrocardiograms (ECG) are indirect proxies of respiratory physiology and therefore present significant challenges for automated apnea detection. Unlike prior deep learning approaches that frame sleep apnea detection as subject-level screening or fixed-window classification [2,3], we formulate the problem as a continuous multi-label sequence segmentation task that explicitly models event progression, annotation uncertainty, and missing sensor modalities in wearable recordings. The framework performs sample-level detection of apnea, hypopnea, and respiratory arousal events from downsampled chest accelerometer and ECG signals using temporal label smoothing [4] and modality-aware representations. We implement this formulation using a Convolutional Recurrent Neural Network (CRNN) that combines convolutional feature extraction, bidirectional Long Short-Term Memory (BiLSTM) networks for temporal context, and Multi-Head Self-Attention to capture long-range relationships between respiratory and cardiovascular responses. The model is optimized for deployment in wearable and resource-constrained monitoring environment through signal downsampling and lightweight feature processing. The framework was evaluated on data collected from 27 subjects undergoing PSG studies, with performance assessed at sample, event, and subject levels. The proposed approach achieved apnea detection ROC-AUC values above 0.80 and an event-level sensitivity of 60.3%, demonstrating that continuous sequence modeling can recover clinically meaningful respiratory event structure from indirect wearable signals. While exact temporal boundaries remained difficult to localize due to the indirect nature of proxy sensing, the model consistently captured clinically meaningful event morphology and respiratory patterns. At the subject level, predicted apnea severity showed strong agreement with PSG-derived measurements, yielding a Pearson correlation of 0.80 between predicted and true apnea-hypopnea index (AHI) values and a mean AHI difference of only 3.91 events per hour. These results demonstrate that clinically meaningful respiratory events can be localized and quantified from wearable proxy signals despite annotation uncertainty, class imbalance, and missing modalities. By reframing apnea detection as continuous sequence segmentation rather than window-based classification, this work provides a practical and computationally efficient framework for scalable sleep apnea screening and longitudinal monitoring. [1] M. Hafezi et al, “Sleep Apnea Severity Estimation from Respiratory Related Movements Using Deep Learning,” in 41st International Conference of IEEE Engineering in Medicine and Biology Society (EMBC), Jul. 2019, pp. 1601–1604. [2] S. Alarcon, N. M. Madrid, R. Seepold, and J. A. Ortega, “Obstructive sleep apnea event detection using explainable deep learning models for a portable monitor,” Frontiers in Neuroscience, vol. 17, 2023. [3] D. Luszczynski, R. F. Yin, N. Afonin, and A. S. P. Lim, “Sleep Apnea Detection on a Wireless Multimodal Wearable Device Without Oxygen Flow Using a Mamba-based Deep Learning Approach,” Nov. 2025. [4] H. Yeche, A. Pace, G. Ratsch, and R. Kuznetsova, “Temporal label smoothing for early event prediction,” in Proceedings of the 40th International Conference on Machine Learning, ser. ICML’23, vol. 202. Honolulu, Hawaii, USA: JMLR.org, Jul. 2023, pp. 39 913–39 938.