Automated Obstructive Sleep Apnea Detection Using Reduced Polysomnography Channels

Obstructive Sleep Apnea (OSA) is a highly prevalent sleep disorder linked to severe cardiovascular diseases and cognitive decline. Traditional diagnosis via manual polysomnography (PSG) is expensive, access-restricted, and labor-intensive. To enable scalable and less invasive screening, this study proposes an automated OSA detection pipeline utilizing a reduced subset of standard physiological channels from a real-world clinical dataset: nasal airflow (FLOW), pulse oximetry SpO2, thoracic effort (THX), and abdominal effort (ABD). We design and systematically evaluate three lightweight hybrid Deep Learning topologies that combine One-Dimensional Convolutional Neural Networks (1D-CNNs) and Bidirectional Long Short-Term Memory (BiLSTM) networks, contrasting expanding, contracting, and multi-scale/parallel spatial feature structures. Models are trained using a robust Focal Loss function and optimized via an independent patient-based cross-validation protocol to ensure true clinical generalizability. Our experimental results demonstrate that the multi-scale parallel architecture (Topology 3) achieved the highest overall performance, yielding a Cohen's Kappa of 0.5996, a Macro F1-score of 0.7996, and an AUC-ROC of 0.8843, while maintaining an accuracy of 81.03%. This parallel design offered the most compact alternative evaluated, reducing trainable parameters by up to 86% compared to deeper serial variations with no trade-off in diagnostic robustness. These findings confirm that integrating concurrent, multi-scale convolutional branches within lightweight networks effectively replaces over-parameterized structures, providing an accurate and computationally efficient mechanism for real-time diagnostic screening systems that use raw, multi-channel physiological streams.