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

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Obstructive sleep apnea (OSA) diagnosis typically relies on polysomnography (PSG), a resource-intensive method.
• Detecting cyclic variation of heart rate (CVHR) from single-lead electrocardiogram (ECG) signals offers a simpler, cost-effective alternative for OSA screening.
• Current deep learning approaches for CVHR detection lack transparency, hindering clinical trust and adoption.

Purpose of the Study:

• To develop the first explainable end-to-end deep learning architecture, SCM-4-OSA, for interpretable obstructive sleep apnea (OSA) detection.
• To address the limitations of current black-box deep learning models in identifying CVHR for OSA diagnosis.
• To enable clinical adoption of AI-driven OSA detection through enhanced model interpretability.

Main Methods:

• Adapted self-contrastive masking (SCM) to create SCM-4-OSA, an explainable deep learning model.
• Employed supervised training with pairwise temporal comparisons of ECG signal intervals to learn complementary masks.
• Validated the model on a publicly available dataset for OSA detection using CVHR identification.

Main Results:

• SCM-4-OSA achieved a task-level accuracy of 86.9% for OSA detection, comparable to state-of-the-art non-interpretable models.
• The model successfully generated visualizations clearly highlighting CVHR patterns in ECG signals.
• Demonstrated the capability of SCM-based models to provide interpretable insights into OSA detection.

Conclusions:

• SCM-4-OSA represents a significant advancement in interpretable deep learning for OSA detection.
• The model's explainability facilitates understanding of CVHR patterns related to OSA, promoting clinical trust.
• This approach shows promise for the widespread clinical adoption of AI in sleep apnea diagnostics.