Cascaded redundant convolutional encoder-decoder network improved apnea detection performance using tracheal sounds...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Respiratory Medicine

Background:

• Acoustic apnea detection is hindered by noise, leading to misdiagnoses.
• Effective denoising of tracheal sounds is crucial for reliable apnea detection in clinical settings like the Post Anesthesia Care Unit (PACU).

Purpose of the Study:

• To enhance apnea detection algorithms in the PACU by developing a tracheal sound denoising method.
• To process tracheal sounds without requiring separate background noise recordings.

Main Methods:

• Tracheal sound data was collected and synthetically corrupted with background noise.
• Frequency-domain features were extracted using Short Time Fourier Transform (STFT).
• A Cascaded Redundant Convolutional Encoder-Decoder (CR-CED) network was trained for denoising, followed by apnea detection.

Main Results:

• The CR-CED network achieved 88% sensitivity and 98.6% specificity in apnea detection.
• Accurate detection of 207 apnea events and 11,305 normal respiratory events was demonstrated.
• The denoising method proved effective on real-world PACU patient data.

Conclusions:

• The CR-CED network provides accurate and reliable tracheal sound denoising for apnea detection in the PACU.
• This approach effectively improves the applicability of tracheal sound analysis in medical environments.
• The method successfully denoises tracheal sounds without needing separate background noise samples.