Subject recognition using wrist-worn triaxial accelerometer data

Abstract

This study demonstrates how a subject can be identified by the means of accelerometer data generated through wrist-worn devices in the context of clinical trials where data integrity is of utmost importance. A custom vector of features extracted from the daily accelerometer time series is defined. Feature selection is adapted to take account of the sequential structure in features. Several classifiers are compared within three different learning frameworks: binary, multi-class and one-class. A simple algorithm like logistic regression shows excellent performance in the binary and multi-class frameworks.