EarMeter: Continuous Respiration Volume Monitoring with Earables
Yang Liu, Qiang Yang, Kayla-Jade Butkow, Jake Stuchbury-Wass, Dong Ma, Cecilia Mascolo
Abstract
Respiration volume, i.e., the amount of air inhaled/exhaled during breathing, is a critical measure for health and fitness in daily life, such as helping optimize sports performance, tracking wellness, and early anomaly detection. Current continuous respiration volume monitoring solutions either require specialized and cumbersome instrumentation setup (e.g., RF transceivers), or rely on customized and non-portable wearables (e.g., masks and chest straps), limiting their usage scenarios. In this paper, we introduce EarMeter, the first continuous respiration volume monitoring system that utilizes in-ear microphones on earbuds to seamlessly track respiration volume across varying breathing intensities, making the measurement more accessible in diverse scenarios. The underlying idea is that breathing sounds, which correlate with breathing volume, can propagate through the body to the ear canals, where they are captured by in-ear microphones. To achieve this, we propose a deep-learning approach to address four unique challenges: limited labeled data, faint breathing sounds, interference from footsteps, and generalization to unseen users. Our approach features fine-tuning an audio encoder pretrained on a broad range of audio datasets, knowledge transfer from high-quality nose audio, performance boosting with breathing-heartbeat coupling, and alignment of both earphone channels with normalization. Extensive experiments under the Leave-One-Subject-Out (LOSO) setting across varying breathing intensities demonstrate the effectiveness of EarMeter, with an average Mean Absolute Percentage Error (MAPE) of 18.19%, meeting the clinically required standard of 20%.