Classifier Personalization for Activity Recognition Using Wrist Accelerometers

• Published in: IEEE journal of biomedical and health informatics Vol. 23; no. 4; pp. 1585 - 1594
• Main Authors: Mannini, Andrea, Intille, Stephen S.
• Format: Journal Article
• Language: English
• Published: United States IEEE 01.07.2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accelerometers; Accelerometry - methods; Active learning; Activity recognition; Algorithms; Classification; Classification algorithms; Classifiers; Human Activities - classification; Humans; incremental learning; Machine Learning; personalization; Signal Processing, Computer-Assisted; Support Vector Machine; Support vector machines; Testing; Training; Uncertainty; Wearable Electronic Devices; wearable sensors; Wrist; Wrist - physiology
• ISSN: 2168-2194; 2168-2208; 2168-2208
• DOI: 10.1109/JBHI.2018.2869779

Intersubject variability in accelerometer-based activity recognition may significantly affect classification accuracy, limiting a reliable extension of methods to new users. In this paper, we propose an approach for personalizing classification rules to a single person. We demonstrate that the method improves activity detection from wrist-worn accelerometer data on a four-class recognition problem of interest to the exercise science community, where classes are ambulation, cycling, sedentary, and other. We extend a previously published activity classification method based on support vector machines so that it estimates classification uncertainty. Uncertainty is used to drive data label requests from the user, and the resulting label information is used to update the classifier. Two different datasets- one from 33 adults with 26 activity types, and another from 20 youth with 23 activity types-were used to evaluate the method using leave-one-subject-out and leave-one-groupout cross validation. The new method improved overall recognition accuracy up to 11% on average, with some large person-specific improvements (ranging from -2% to +36%). The proposed method is suitable for online implementation supporting real-time recognition systems.