Weakly Supervised Deep Learning for Monitoring Sleep Apnea Severity Using Coarse-grained Labels.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Sleep apnea diagnosis relies on manual annotation of physiological signals, which is time-consuming and costly.
• Current machine learning approaches for apnea detection require extensive, finely labeled data, limiting their clinical application.
• Undiagnosed sleep apnea affects a large patient population, highlighting the need for more accessible diagnostic tools.

Purpose of the Study:

• To develop a weakly supervised deep learning framework for estimating fine-grained sleep apnea severity using only coarse-grained labels.
• To incorporate clinical knowledge into a deep learning model to improve the accuracy of apnea severity estimation.
• To reduce the reliance on costly manual labeling for sleep apnea diagnostics.

Main Methods:

• A novel knowledge-enhanced dual-granularity consistency loss was designed to improve fine-grained apnea severity learning.
• Clinical knowledge was mathematically encoded using ordinal alignment functions to calibrate estimation accuracy.
• The framework utilizes coarse-grained labels (apnea presence) to estimate fine-grained severity.

Main Results:

• The proposed method accurately estimates fine-grained sleep apnea severity in real time.
• Significantly reduced labeling costs were achieved compared to traditional supervised methods.
• The model demonstrated superior performance in monitoring apnea severity at high temporal resolution.

Conclusions:

• Weakly supervised deep learning, enhanced with clinical knowledge, offers an effective solution for sleep apnea severity estimation.
• This approach can significantly lower diagnostic costs and expand access to sleep apnea monitoring.
• The developed framework holds promise for both in-lab and at-home sleep apnea diagnostics.