Wavelet-based CNN for Predicting PAP Adherence Using Overnight Polysomnography Recordings: a Pilot Study.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence in Healthcare

Background:

• Obstructive sleep apnea (OSA) is a prevalent sleep disorder impacting patient health.
• Positive airway pressure (PAP) therapy is the primary treatment, but adherence remains a significant challenge.
• Predicting PAP adherence is crucial for optimizing treatment outcomes.

Purpose of the Study:

• To develop a predictive model for PAP adherence using data from PAP titration polysomnograms (PSG).
• To address challenges of long PSG signal duration and limited dataset size in model development.
• To leverage deep learning and wavelet-based feature extraction for PSG signal analysis.

Main Methods:

• A pipeline was created using a PAP titration PSG database.
• Wavelet-based deep learning model incorporating local feature extraction and P-norm pooling was developed.
• Pre-trained EfficientNet-B7 was utilized for unsupervised feature extraction to overcome data limitations.

Main Results:

• The developed model achieved 78% balanced accuracy on the test set.
• The model demonstrated an 83% Area Under the Curve (AUC) for predicting PAP adherence.
• Performance was evaluated using airflow and frontal electroencephalogram (EEG) signals.

Conclusions:

• The wavelet-based deep learning model shows potential for predicting PAP adherence.
• This pilot study presents a compelling approach for analyzing PSG data to improve OSA treatment.
• Further research is warranted to validate and implement this predictive model in clinical practice.