Explainable Cross-Modal Attention Fusion of ViT and BiLSTM for Automated Sleep Apnea Severity Assessment from EEG Signals

Abstract

Obstructive Sleep Apnea (OSA) is a common sleep-related breathing disorder that involves recurrent upper airway collapses during sleep, resulting in intermittent hypoxia, sleep fragmentation, cardiovascular symptoms, and cognitive impairment. Polysomnography (PSG) remains the gold standard for diagnosing OSA; however, it is expensive, complex, and not widely available, thus limiting its use for large-scale screening. Electroencephalography (EEG) is useful for providing information on sleep stages and OSA-related cortical arousals, and thus it is a promising modality for automated OSA severity assessment. In this study, a novel explainable hybrid architecture combining a Vision Transformer (ViT) and a Bidirectional Long Short-Term Memory (BiLSTM) network is introduced for the severity classification of sleep apnea into four classes based on a single-channel EEG signal from the Sleep Heart Health Study (SHHS) database. EEG data were preprocessed by bandpass filtering and then converted to Short-Time Fourier Transform (STFT) time-frequency spectrograms. The ViT branch is responsible for capturing global spectral representations from spectrograms, whereas the BiLSTM branch captures temporal dependencies from raw EEG sequences. To leverage complementary information across different modalities, a cross-modal attention fusion module is introduced, and, to support feature representation and model generalization, a self-supervised contrastive pre-training algorithm is used. The experimental results show that the proposed model achieves an accuracy of 99.0%, which is higher than that of a Convolutional Neural Network (CNN) (92.68%) and a ResNet18 transfer-learning model (98.68%). Furthermore, model interpretability is achieved by using the SHapley Additive exPlanations (SHAP), and the most influential physiological features contributing to severity classification are identified. The proposed framework also achieves superior precision, recall, F1-score and Area Under the Receiver Operating Characteristic Curve (AUC) performance. The results show that this approach is clinically applicable and effective for automated OSA severity assessment and evaluation based on single-channel EEG signals.