Paper Title
Attention-Weighted Deep Belief Networks for Interpretable Sleep Disorder Classification From Multimodal Biosignals

Abstract
Sleep-related disorders of consciousness are difficult to diagnose because the underlying neural dysfunctions are often subtle and transient. This paper presents a framework that combines multimodal data fusion with temporal attention mechanisms to predict and analyze these disorders. EEG, ECG, and EMG signals are acquired and processed in MATLAB to capture neural dynamics across sleep stages. A Deep Belief Network (DBN) with three hidden layers (128-64-32 units) extracts hierarchical features that encode increasingly complex temporal patterns of brain activity. Temporal attention zeroes in on EEG segments tied to consciousness disruptions, while multi- modal fusion integrates complementary physiological signals for stronger prediction and interpretability. A total of 48 features are extracted across the three modalities, and four classifiers are com- pared: Random Forest, Support Vector Machine (SVM), DBN- only, and the proposed DBN with temporal attention. Experimen- tal results show that the proposed framework achieves 96.67% classification accuracy, outperforming Random Forest (90.00%), SVM (93.33%), and DBN-only (93.33%). Beyond accuracy gains, the framework provides interpretable attention weights that let clinicians trace predictions back to specific physiological features, supporting early diagnosis and personalized treatment in sleep medicine. Keywords - Sleep, Disorders of Consciousness, EEG, Deep Belief Network, Temporal Attention, Multimodal Data Fusion, MATLAB