Tracheal Sound Analysis Using a Deep Neural Network to Detect Sleep Apnea

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Portable sleep apnea testing devices struggle to differentiate sleep/wake status, potentially leading to underestimations.
• Tracheal sound (TS) spectrograms contain valuable information for identifying breathing events and sleep states.

Purpose of the Study:

• To develop and validate a deep neural network (DNN) system for sleep apnea testing using tracheal sound (TS) analysis.
• To assess the DNN's capability in detecting breathing events and classifying sleep/wake status.

Main Methods:

• A DNN system was developed and trained using 1,548 polysomnography (PSG) records, analyzing 60-second TS spectrograms labeled with PSG scoring.
• Two DNNs were trained: one for breathing status and another for sleep/wake status discrimination.
• The system was validated on a separate set of 304 PSG records.

Main Results:

• The DNN system demonstrated high accuracy in discriminating sleep/wake status.
• Diagnostic performance for sleep-disordered breathing (SDB) showed high sensitivity (0.92-0.98) and specificity (0.76-0.94) across different apnea-hypopnea index thresholds.
• The convolutional layer DNN architecture proved effective for both breathing and sleep status discrimination.

Conclusions:

• The developed tracheal sound (TS) deep neural network (DNN) analysis system shows excellent performance for sleep-disordered breathing (SDB) testing.
• This DNN-based approach offers a promising solution for more accurate home sleep apnea diagnosis.
• The system's ability to classify sleep/wake status enhances the reliability of portable sleep testing.