SCM-4-OSA: An End-to-End Explainable Deep Learning Model for Interpretable Obstructive Sleep Apnea Detection Based on...

Polysomnography (PSG) is considered as the gold standard for detecting obstructive sleep apnea (OSA). Recently, the automatic identification of cyclic variation of heart rate (CVHR) in single-lead electrocardiogram (ECG) signal has emerged as a promising alternative approach for OSA detection, offering advantages such as simplicity and cost-effectiveness over PSG. However, existing deep learning-based solutions for identifying CVHR lack explainability regarding how decisions are made by the model. This limits the interpretability of the results, and the potential for clinical adoption. In this paper, by adapting self-contrastive masking (SCM), we present SCM-4-OSA, the first explainable end-to-end deep learning architecture for interpretable OSA detection. Trained and validated on a publicly-available dataset, SCM-4-OSA learns complementary masks during supervised training by performing pairwise temporal comparisons between intervals of ECG signal to detect CVHR related to OSA. Results demonstrate that the model achieves a task-level accuracy of 86.9%, comparable to the state-of-the-art models lacking interpretability, while simultaneously generating visualizations that clearly capture CVHR patterns in the ECG signal for interpretable OSA detection. Overall, this study demonstrates the potential of SCM-based deep learning models for OSA detection with interpretability, paving the way for their adoption in clinical settings.