Speaker
Description
Audioplethysmography (APG) enables cardiovascular monitoring through ear-worn devices by detecting cyclical changes to the acoustic properties of the ear canal. However, body-motion-induced artifacts significantly degrade signal quality in real-world applications. This paper investigates the impact of body motion on the accuracy of APG-derived heart-rate estimates and compares a digital signal reconstruction method and three machine-learning-based estimators against an unmitigated baseline. The candidate-ranking model achieved the best overall performance, with a mean absolute heart rate error of 5.36 bpm and a mean absolute percentage error (MAPE) of 6.35%. Its MAPE remained below 5% in all conditions except cycling, where it reached 15.72%. These results show that machine-learning-based candidate selection provides more robust motion-artifact rejection than the signal reconstruction method, although severe body motion remains the primary obstacle to reliable cardiovascular monitoring using APG in everyday scenarios.