SiCRNN: A Siamese Approach for Sleep Apnea Identification via Tracheal Microphone Signals.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence in Medicine
• Sleep Medicine

Background:

• Sleep apnea syndrome (SAS) affects millions globally but is frequently undiagnosed.
• Current diagnosis via polysomnography (PSG) is intrusive and relies on subjective clinician interpretation.
• There is a need for objective, accessible SAS detection methods.

Purpose of the Study:

• To develop and evaluate a novel decision support system for detecting sleep apnea events.
• To utilize a tracheal microphone and a deep learning model for objective SAS detection.
• To overcome the limitations of traditional polysomnography.

Main Methods:

• A deep learning model, SiCRNN (Siamese Convolutional Recurrent Neural Network), was developed.
• The system processed Mel spectrograms from tracheal microphone recordings.
• Apnea event detection was finalized using k-means clustering.

Main Results:

• The SiCRNN model achieved a Recall score of up to 95% for apnea events.
• Experimental runs optimized network configuration and data parameters.
• The Siamese training paradigm demonstrated superior performance compared to a fully convolutional baseline.

Conclusions:

• The proposed tracheal microphone and SiCRNN system provide an effective method for sleep apnea detection.
• This approach shows promise for improving the accessibility and objectivity of SAS diagnosis.
• The Siamese deep learning paradigm enhances the identification of sleep apnea events.