SelANet: decision-assisting selective sleep apnea detection based on confidence score

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Sleep apnea syndrome is a common disorder diagnosed via polysomnography, which is labor-intensive and costly.
• Automated detection algorithms using limited biological signals are being developed to overcome polysomnography's limitations.
• Uncertainty in AI judgments due to limited data necessitates improved diagnostic approaches.

Purpose of the Study:

• To develop an automated sleep apnea detection algorithm using selective prediction based on confidence scores.
• To address the uncertainty issue in AI-driven diagnoses from limited biological signals.
• To provide a clinician-support tool that determines the need for polysomnography.

Main Methods:

• Utilized polysomnography data from 994 subjects.
• Employed feature extraction from latent vectors using an autoencoder.
• Designed and trained a 1D Convolutional Neural Network-Long Short-Term Memory (1D CNN-LSTM) model with a selective prediction function based on confidence scores.

Main Results:

• The developed model achieved 90.26% accuracy, 91.29% sensitivity, and 89.21% specificity in classifying sleep apnea.
• Selective prediction improved all performance metrics by approximately 7.03% compared to models without this feature.
• The algorithm demonstrated convergence of empirical coverage to the target coverage with minimal coverage violation (0.067).

Conclusions:

• The selective prediction algorithm effectively minimizes diagnostic uncertainty arising from limited biological data.
• This approach is suitable for wearable devices and can serve as a simple screening tool or a complementary diagnostic method.
• The algorithm enhances the efficiency and accessibility of sleep apnea diagnosis.