Evaluating Multi-Modality Deep Learning Models for Automated Sleep Apnea Detection

Sleep apnea is a common sleep disorder characterized by recurring interruptions in breathing during sleep. Accurate classification of its subtypes, obstructive, central, and mixed apnea, is essential for targeted treatment. This study evaluates the performance of four deep learning models; 1D Convolutional Neural Network (CNN1D), CNN- LSTM, Bidirectional LSTM (BiLSTM), and Transformer, in classifying sleep apnea types using time-series physiological data. The models were compared based on classification accuracy, confusion matrix analysis, and training time. The Transformer model achieved the highest accuracy of 94.5%, followed by BiLSTM (92.8%), CNN-LSTM (91.2%), and CNN1D (89.7%). While CNN1D had the shortest training time (1.0 minutes), the Transformer required the longest (1.7 minutes) but yielded superior classification performance with minimal inter-class confusion. These findings demonstrate the effectiveness of deep learning, particularly Transformer-based architecture, in automating and enhancing the diagnosis of sleep apnea in clinical and telemedicine applications.