Robust Multimodal Cough and Speech Detection using Wearables: A Preliminary Analysis.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Wearable Technology

Background:

• Clinical cough detection methods are accurate but not home-accessible.
• Wearable devices offer accessibility but struggle with audio quality, background noise, and speech privacy.
• Accurate cough detection is vital for monitoring respiratory conditions remotely.

Purpose of the Study:

• To develop a compact, multimodal system for simultaneous cough and speech detection.
• To enhance system robustness against real-world audio challenges using Out-of-Distribution (OOD) detection.
• To improve the accuracy and privacy of cough detection for wearable applications.

Main Methods:

• Developed a small-size, multimodal system integrating cough and speech detection.
• Implemented transfer learning and Out-of-Distribution (OOD) detection algorithms.
• Evaluated system performance in in-subject and cross-subject settings with and without OOD inputs.

Main Results:

• High accuracies achieved: 92.59% (in-subject) and 90.79% (cross-subject) without OOD inputs.
• Maintained strong performance with OOD inputs: 91.97% (in-subject) and 90.31% (cross-subject).
• Demonstrated effective OOD detection, preserving accuracy despite a high volume of non-target audio data.

Conclusions:

• Combining multimodal sensing, transfer learning, and OOD detection significantly boosts cough detection performance.
• The developed system offers a promising, accurate, and privacy-preserving solution for wearable respiratory health monitoring.
• This approach addresses key challenges in real-world wearable-based cough and speech detection.