Robust Cough Detection With Out-of-Distribution Detection

Area of Science:

• Biomedical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Cough detection is vital for monitoring respiratory conditions like asthma.
• Current acoustic cough detection models struggle with real-world noisy data (Out-of-Distribution or OOD).

Purpose of the Study:

• To develop robust cough detection methods that handle OOD data.
• To improve the accuracy and reliability of acoustic cough detection for asthma patients.

Main Methods:

• Proposed two novel cough detection methods incorporating an OOD detection module.
• Implemented techniques include adding a learning confidence parameter and maximizing entropy loss.
• Evaluated performance across various sampling rates and audio window sizes.

Main Results:

• The OOD system reliably distinguishes In-Distribution (ID) and OOD data above 750 Hz sampling rate.
• OOD detection performance improves with larger audio window sizes.
• Increased OOD data proportion enhanced model accuracy and precision, especially at lower sampling rates.

Conclusions:

• Integrating OOD detection significantly improves acoustic cough detection performance.
• These methods offer a practical solution for real-world acoustic cough monitoring in asthma.
• The approach enhances the utility of portable recording devices for patient self-monitoring.