Coordinate Attention for 1D Audio-Based Sleep Apnea Detection: A Multi-Seed Empirical Study on Smartphone-Deployable Architectures

Sleep apnea affects an estimated 936 million adults globally, yet over 80% of cases remain undiagnosed due to the inaccessibility of polysomnography (PSG), the clinical gold standard. We present a lightweight neural network architecture for detecting sleep-disordered breathing events directly from smartphone audio, eliminating the need for wearable devices, cloud connectivity, or external sensors. The proposed architecture adapts the Coordinate Attention (CA) mechanism — originally proposed for 2D mobile vision networks — to one-dimensional biomedical time-series classification through a global-local feature fusion design that preserves temporal-position information in the attention map.Operating on a 200 × 3 acoustic feature matrix sampled at 1 Hz — comprising sound pressure level, SPL change rate, and a binary snore-presence indicator produced by a short-window snore-detection CNN as the Stage-1 output of a cascaded two-stage pipeline (the cascade architecture itself is the subject of companion work [Yang L., 2026a]) — we evaluate three architectures as Stage-2 classifier candidates: a vanilla 1D CNN baseline, an SE-Attention CNN, and the proposed Coordinate Attention (Coord-Attn) CNN. All three are evaluated under an identical training protocol across five random seeds (42, 123, 456, 789, 2026) on an internal dataset of 2,953 PSG-labeled 200-second windows drawn from 80 person-nights across 40 participants (in-laboratory PSG + portable PSG with nasal-airflow cannula). Results are reported as mean with 95% bootstrap confidence intervals (10,000 iterations) to characterize variability and pairwise differences are evaluated by paired-seed bootstrap. The Coord-Attn CNN achieves an accuracy of 87.14% (85.14, 89.68) and an F1-score of 86.94% (84.28, 89.77) while using only 14,001 parameters — a 93.2% reduction relative to the 204,801-parameter baseline (83.82% accuracy, 83.99% F1). Paired bootstrap analysis shows that Coord-Attn is statistically distinguishable from the Original baseline in six of seven metrics (Accuracy, Precision, Specificity, F1, AUC-ROC, AUC-PR; Recall is not). Against the SE-Attention CNN of comparable size, Coord-Attn provides a statistically significant advantage in Precision and Specificity but not in F1 or AUC, suggesting that the temporal-position-preserving property of Coordinate Attention is most beneficial when minimizing false alarms matters more than maximizing sensitivity. Data and code availability: The three model architectures, training pipeline, and bootstrap analysis code are publicly available at https://github.com/somnisense/ca1d-sleep-apnea under the MIT License. By design, no audio recordings, labels, or derived feature matrices are distributed with the repository: the original recordings were collected under participant consent that does not cover public release of either the waveforms or the derived features. Competing interests: The author is the founder of SomniAI LLC, the entity that may commercialize the technology described in this work. The methods described are also the subject of three co-filed U.S. provisional patent applications by SomniAI LLC. No external funding was received.