Prediction of the Sleep Apnea Severity Using 2D-Convolutional Neural Networks and Respiratory Effort Signals

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence in Healthcare

Background:

• Sleep apnea is prevalent, but polysomnography has limitations.
• Automated diagnosis using fewer physiological measures is needed.
• Thoracic and abdominal movement signals offer potential for sleep apnea assessment.

Purpose of the Study:

• To demonstrate the utility of thoracic (THO) and abdominal (ABD) movement signals for estimating sleep apnea severity.
• To evaluate the contribution of central respiratory events in automated sleep apnea diagnosis.
• To develop and assess 2D-convolutional neural networks (CNNs) for sleep apnea severity estimation.

Main Methods:

• Development of 2D-CNN models utilizing both THO and ABD signals.
• Joint analysis of THO and ABD data for automated sleep apnea severity estimation.
• Evaluation of model performance in estimating apnea-hypopnea index (AHI) and central apnea index (CAI).

Main Results:

• The CNN achieved an ICC of 0.75 and RMSE of 10.33 events/h for AHI estimation.
• For CAI estimation, the model reached an ICC of 0.83 and RMSE of 0.95 events/h.
• High accuracies were obtained for classifying AHI severity (up to 94.98%) and central events (up to 99.74%).

Conclusions:

• THO and ABD movement signals, analyzed with CNNs, provide a powerful tool for diagnosing sleep apnea.
• The proposed method is especially effective for patients with a high prevalence of central apnea events.
• Automated sleep apnea diagnosis using limited physiological data is feasible and accurate.