Semi-Supervised Learning for Low-Cost Personalized Obstructive Sleep Apnea Detection Using Unsupervised Deep Learning...

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Cardiology

Background:

• Obstructive sleep apnea (OSA) is a prevalent disorder with significant health implications.
• Current OSA detection methods often lack personalization and rely on extensive clinical annotations.
• Single-lead electrocardiogram (ECG) offers a low-cost modality for potential OSA monitoring.

Purpose of the Study:

• To develop and validate an automatic semi-supervised algorithm for personalized, low-cost obstructive sleep apnea (OSA) detection using single-lead ECG.
• To address the limitations of existing research by enabling automated personalization fine-tuning.
• To improve the performance of general OSA detection models through data-driven adaptation.

Main Methods:

• Utilized a convolutional neural network (CNN)-based auto-encoder (AE) with a modified training objective for anomaly detection in OSA.
• Implemented a novel semi-supervised approach to assign pseudo-labels to ECG samples based on model confidence.
• Validated the algorithm on both same-database and cross-database scenarios for generalization assessment.

Main Results:

• Within-database validation showed improvements in accuracy (86.3% to 90.3%), AUC (0.915 to 0.948), and MAE (5.178 to 2.593) for 35 subjects.
• Cross-database validation demonstrated enhanced accuracy (75.6% to 84.3%), AUC (0.800 to 0.881), and MAE (9.149 to 3.509) for 25 subjects.
• The auto-encoder (AE) effectively identified abnormal OSA features and adapted to new data.

Conclusions:

• The proposed semi-supervised algorithm demonstrates high adaptability for personalized OSA detection using ECG.
• The data-driven pseudo-labeling strategy overcomes the need for extensive clinical annotations, enabling cost-effective implementation.
• This approach offers a high-performance, annotation-free solution for personalized obstructive sleep apnea detection.