Multi-Task Learning for OSA Detection and Sleep Staging via Multi-Scale Modeling.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Obstructive sleep apnea (OSA) and sleep fragmentation are critical in sleep disorder diagnosis.
• Existing deep learning models often focus on either OSA detection or sleep staging, not both.
• A unified approach is needed for efficient and comprehensive sleep analysis.

Purpose of the Study:

• To develop a unified multi-modal multi-task network for simultaneous OSA detection and sleep staging.
• To introduce MT-TASPPNet (Multi-Task Triple Atrous Spatial Pyramid Pooling Network) for joint sleep analysis.
• To enhance the discrimination of sleep stages using an EOG-guided prior mechanism.

Main Methods:

• Developed MT-TASPPNet, a multi-modal multi-task network integrating EEG, ECG, and airflow signals.
• Employed Atrous Spatial Pyramid Pooling modules for multi-scale temporal-frequency pattern capture.
• Incorporated an EOG-guided prior mechanism to improve sleep stage discrimination.
• Evaluated the model on three large-scale datasets (SHHS1, SHHS2, Sydney Sleep Biobank) using a 3-min input window.

Main Results:

• Achieved OSA detection accuracy between 0.798 and 0.884 (MF1: 0.772 to 0.821).
• Attained sleep staging accuracy between 0.776 and 0.834 (MF1: 0.735 to 0.749, Kappa: 0.697 to 0.77).
• Demonstrated consistent performance across heterogeneous data and individual variability.

Conclusions:

• MT-TASPPNet effectively performs joint OSA detection and sleep staging.
• The model's stability and adaptability are validated for clinical settings.
• This work paves the way for efficient and scalable multi-task sleep analysis systems.