Exploring Siamese-Based Self-Supervised Learning for Sleep Apnea Detection

Obstructive sleep apnea (OSA) is a common and serious sleep disorder characterized by periodic interruptions in breathing lasting more than 10 s (apnea episodes) during sleep. OSA significantly affects quality of life and overall health, highlighting the critical need for an accurate and timely diagnosis. Polysomnography (PSG) is the standard diagnostic technique for OSA, involving the collection of respiratory, oxygen saturation, biometric, and physiological signals. However, manual analysis of these extensive sleep recordings by medical professionals is labor-intensive and time-consuming. To address this challenge, we propose a Siamese Network-based Self-Supervised Learning (SSSL) model for the automatic identification of SA episodes from single-channel electrocardiogram (ECG) signals. Unlike conventional self-supervised methods, our approach does not require a momentum encoder, large batch sizes, or negative-positive pair construction. The model is evaluated using the PhysioNet Apnea-ECG database and employs a two-stage training strategy. In the first stage, the encoder is trained on unlabeled data to learn robust signal representations. In the second stage, the pre-trained encoder and classifier are fine-tuned using labelled data for optimal classification performance. The proposed model achieved high accuracy of $73\%$, $77\%$, and $86\%$ when fine-tuned with $10\%$, $50\%$, and $100\%$ of the labelled training data, respectively, for the classification per segment. These results demonstrate the model's effectiveness in both offline and online diagnostic settings, outperforming state-of-the-art methods.