Investigating the Relationship between Cough Detection and Sampling Frequency for Wearable Devices

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence in Healthcare

Background:

• Manual cough counting is time-consuming and expensive for monitoring chronic respiratory conditions.
• Existing Automatic Cough Detection Algorithms (ACDAs) often require high sampling frequencies, leading to high power consumption and difficulty in implementation on wearable devices.
• Reproducibility of ACDA performance is hindered by the use of private clinical data.

Purpose of the Study:

• To present a novel Automatic Cough Detection Algorithm (ACDA) suitable for clinical monitoring.
• To demonstrate reliable ACDA operation at a low sampling frequency.
• To analyze the trade-off between speech obfuscation for privacy and cough detection accuracy.

Main Methods:

• Implementation of an ACDA using a convolutional neural network (CNN).
• Utilizing a publicly available dataset for training and validation.
• Operating the ACDA at a low sampling frequency of 750 Hz.

Main Results:

• The proposed ACDA achieved a sensitivity of 92.7%, specificity of 92.3%, and accuracy of 92.5% at 750 Hz.
• Low sampling frequency effectively preserves patient privacy by obfuscating speech.
• Demonstrated a viable trade-off between speech privacy and cough detection performance.

Conclusions:

• A novel ACDA operating at a low sampling frequency (750 Hz) meets clinical monitoring requirements.
• The developed technique offers a balance between effective cough detection and enhanced patient privacy.
• Publicly available data and CNNs enable reproducible and efficient cough detection for respiratory condition monitoring.