Abstract / Summary
Accurate monitoring of energy expenditure (EE) during exercise is crucial for health management, fitness optimization, and clinical assessment. Although many commercial wearables offer EE levels as a standard feature, they typically rely on simple regressions of heart rate, motion trajectory, or activity intensity. Consequently, such models inherently struggle with cross-activity generalization and demand massive labeled data, leading to significant accuracy degradation in practical deployment. To address this, we propose EarEE, a novel earable system that estimates EE by decoding exercise-associated respiratory sounds into oxygen consumption and carbon dioxide production, which are gold-standard metabolic indicators of EE. EarEE advances existing technologies through three key innovations: i) Dynamics-Aware Deep Modeling , which captures the nonlinear temporal and contextual relationship between respiratory acoustics and gas exchange; ii) Generative Data Synthesis , which expands data from a short calibration session into diverse and realistic respiratory profiles, enabling robust model training with limited labeled data; and iii) Heart Sounds Suppression , which mitigates heart sound artifacts for reliable estimation during intensive activities. Extensive evaluations with 42 participants across 7 exercises demonstrate that EarEE achieves an average relative error of 10.95%, significantly outperforming wearable solutions and approaching the 10% clinical accuracy benchmark, highlighting its potential for practical daily EE monitoring.