Exploring advanced machine learning models as solutions to the "Other Peoples Cough Problem" - distinguishing user coughs from others in shared environments.
This paper explores the problem of detecting coughs from other people in shared environments when using wearable devices for health monitoring. We present a novel solution using machine learning models to classify coughs as "near" (from the user) or "far" (from others) based on acoustic properties. We collected a unique dataset of coughs recorded at different distances and trained models to capture spectral differences. Among the models tested, convolutional neural networks (CNNs) demonstrated exceptional performance, achieving a 0.94 ROC-AUC score in distinguishing between the user’s own coughs and those of others nearby.
The paper emphasizes the importance of accurate cough detection in remote health monitoring, particularly for patients with chronic respiratory diseases like COPD or during infectious disease outbreaks such as COVID-19. To enhance detection, the paper suggests integrating additional data streams—such as heart rate and motion sensors from wearable devices—to verify that a detected cough originated from the user. This multi-modal approach has the potential to improve the precision of cough monitoring, making the technology more reliable for real-world healthcare applications, from personalized treatment to population-level disease tracking.
To get early access to this paper fill in the form below