Harnessing Smartwatch Microphone Sensors for Cough Detection and Classification

TL;DR

This study investigates cough detection and classification using a smartwatch microphone, addressing the need for continuous, everyday respiratory health monitoring. The authors collected 9 hours of smartwatch audio from 32 participants and built a 1D CNN classifier on MFCC and Mel-spectrogram features, achieving $0.9849$ (non-walking) and $0.9823$ (walking) accuracy, and identifying four distinct cough types via clustering. Data augmentation increased cough sample diversity, and manual labeling ensured clean cough segments. The findings support the feasibility of wearables for real-world cough monitoring and nuanced cough-type analysis, with implications for early disease detection and public health surveillance.

Abstract

This study investigates the potential of using smartwatches with built-in microphone sensors for monitoring coughs and detecting various cough types. We conducted a study involving 32 participants and collected 9 hours of audio data in a controlled manner. Afterward, we processed this data using a structured approach, resulting in 223 positive cough samples. We further improved the dataset through augmentation techniques and employed a specialized 1D CNN model. This model achieved an impressive accuracy rate of 98.49% while non-walking and 98.2% while walking, showing smartwatches can detect cough. Moreover, our research successfully identified four distinct types of coughs using clustering techniques.

Figures (11)

• Figure 1: Study Design
• Figure 2: Steps in cough detection
• Figure 3: An example of cough file extraction in Audacity Software.
• Figure 4: MFCC of an audio signal. [Best viewed in color]
• Figure 5: Mel Spectogram of an audio signal.[Best viewed in color]