Hierarchical Attention-Based Multimodal Deep Learning Model for Sleep Apnea Severity Classification

Sleep apnea is a prevalent yet frequently underdiagnosed respiratory disorder characterized by recurrent cessations in breathing during sleep, resulting in oxygen desaturation, autonomic instability, and elevated cardiovascular and metabolic risk. Despite its clinical significance, large-scale diagnosis remains a major challenge, as the gold-standard polysomnography (PSG) is costly, cumbersome, and unsuitable for continuous or home-based screening. Moreover, conventional single-signal or feature-engineered approaches are limited in capturing the complex cardiorespiratory dynamics underlying apnea events. To address these limitations, this study presents a multimodal deep learning framework that integrates convolutional, recurrent, and hierarchical attention mechanisms for automated sleep apnea severity classification, leveraging heart rate (HR), oxygen saturation (SpO2), and oximetry status (OX) signals from the Sleep Heart Health Study (SHHS) dataset. The proposed framework combines modality-specific convolutional–recurrent encoders with hierarchical attention, where Dual Attention refines signal-specific representations and Multi-Head Attention models interactions among complementary physiological signals. Trained on 5,227 SHHS1 recordings, the model attains 0.924 accuracy, 0.896 Cohen’s $\kappa $ , and 0.922 macro-F1, outperforming state-of-the-art benchmarks. The results highlight that hierarchical attention and multimodal integration substantially enhance robustness and generalization. Overall, the proposed framework provides a scalable, non-invasive, and clinically meaningful alternative to PSG, enabling reliable home-based screening and severity assessment of sleep apnea.