Screening Sleep Apnea Using Random Convolution Kernels of Acoustic Speech Representations.

Area of Science:

• Biomedical Engineering
• Sleep Medicine
• Artificial Intelligence

Background:

• Obstructive sleep apnea (OSA) affects 38% of adults, diagnosed via cumbersome polysomnography (PSG).
• Current alternative screening methods like questionnaires have low specificity.
• Speech-based screening is accessible but underutilizes deep learning.

Purpose of the Study:

• To explore deep neural network potential for OSA severity prediction using speech.
• To evaluate random convolution kernels for simulating CNN effects on speech data.
• To predict apnea-hypopnea index (AHI) thresholds of 10 and 15 events/hour.

Main Methods:

• Utilized random convolution kernels to mimic convolutional neural network (CNN) behavior.
• Analyzed speech acoustic features to predict OSA severity.
• Tested prediction accuracy against AHI thresholds of 10 and 15 events/hour.

Main Results:

• Achieved an F1-score of 0.83 for AHI ≥ 10 events/hour.
• Achieved an F1-score of 0.71 for AHI ≥ 15 events/hour.
• Performance surpassed existing literature for similar sample sizes.

Conclusions:

• Speech analysis with simulated CNNs offers a viable, accurate method for OSA screening.
• This approach provides a more accessible alternative to traditional diagnostic methods.
• Further research into deep learning for OSA detection is warranted.