Purpose: This study aimed to predict human energy expenditure and activity type using a miniature lightweight ear-worn inertia sensor and a novel pattern recognition algorithm for activity detection.
Methods: This study used a protocol of 11 activities of daily living: lying down, standing, computer work, vacuuming, stairs, slow walking, brisk walking, slow running, fast running, cycling, and rowing. Subjects included 25 healthy randomized subjects (18 males and 7 females). Each participant wore the ear sensor to record posture and linear acceleration, as well as the Cosmed K4b2 system for indirect calorimetry. The main outcome measure was the continuous energy expenditure per minute prediction for both task-known and task-blind estimation.
Results: The values for METs predicted using the proposed algorithm and the measured METs using the K4b2 showed good agreement with low values for the systematic bias (lying down = 0.01, standing = −0.02, computer work = −0.04, vacuuming = −0.17, stairs = −0.02, slow walking = 0.01, fast walking = 0.04, slow running = 0.14, fast running = −0.35, cycling = 0.32, and rowing = 0.10). For task-blind prediction, the agreement between predicted and measured METs is also good with low values of the systematic bias (lying down = 0.11, standing = 0.14, computer work = −0.06, vacuuming = 0.47, stairs = −0.47, slow walking = 0.53, fast walking = −0.11, slow running = 0.83, fast running = −1.18, cycling = 0.31, and rowing = −0.67). Activity is also well predicted (for task-blind prediction) with an overall success rate of 88.99% and individual correct classification rates of lying down = 89.62%, standing/computer work = 99.10%, vacuuming = 76.60%, stairs = 89.13%, walking = 85.11%, running = 98.96%, and cycling = 79.79%.
Conclusions: The ear-worn sensor presented in this work is a novel lightweight device that can be used to predict energy expenditure for a range of activities without behavior interference or modification.