Automatic sleep staging from CPAP airflow using a dual fusion multi-period convolutional neural network

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Science

Background:

• Continuous Positive Airway Pressure (CPAP) is standard for obstructive sleep apnea-hypopnea syndrome.
• Repurposing CPAP devices for sleep dynamics monitoring is challenging due to signal artifacts and variability.
• Existing CPAP devices primarily record airflow (CPAP-flow) signals, limiting their use beyond therapy.

Purpose of the Study:

• To develop a novel deep learning model for passive sleep dynamics monitoring using CPAP devices.
• To overcome limitations of traditional models in analyzing CPAP-flow signals for sleep staging.
• To create a dual-purpose platform for both CPAP therapy and longitudinal sleep assessment.

Main Methods:

• Introduced a Dual Fusion Multi-Period Convolutional Neural Network (DFMP-CNN) model.
• Leveraged multiple period-specific convolutional kernels and a dual-fusion mechanism.
• Encoded short- and long-range temporal dependencies within CPAP-flow signals.

Main Results:

• DFMP-CNN achieved state-of-the-art performance in CPAP-based sleep staging.
• Achieved 78.5% accuracy (κ=0.605) on the Yale dataset and 73.6% accuracy (κ=0.524) on the Duke dataset.
• Demonstrated cross-dataset transferability across clinical centers and device types, highlighting robustness.

Conclusions:

• DFMP-CNN offers an unobtrusive method for sleep monitoring via CPAP devices.
• The model shows potential for improving clinical assessment and optimizing sleep therapy management.
• This dual-purpose platform integrates therapy delivery with longitudinal sleep assessment capabilities.