A dual-branch metric-learning framework for obstructive sleep apnea detection

Abstract Obstructive sleep apnea (OSA) screening using wearable-compatible signals remains challenging because of signal noise, temporal complexity, and physiological variation among individuals. We propose a multimodal Siamese metric-learning framework that jointly models electrocardiography (ECG) and oxygen saturation (SpO $$_2$$ ) dynamics through a dual-branch encoder and triplet-loss optimization. The architecture integrates a 1D ResNet-18 backbone with bidirectional long short-term memory (BiLSTM)-based temporal fusion to capture complementary physiological patterns. We evaluated the framework using subject-disjoint five-fold cross-validation in a curated cohort from the PSG-Audio dataset. The proposed model achieved an accuracy of $$91.03 \pm 2.75\%$$ , recall of $$90.58 \pm 4.84\%$$ , F1-score of $$91.03 \pm 2.57\%$$ , and AUC of $$0.9506 \pm 0.0342$$ . Patient identities were separated across training, validation, and test partitions, and test data were excluded from triplet construction, model selection, and threshold selection. These results support multimodal metric learning as a competitive within-cohort approach but do not establish generalization to external populations or recording systems.