ADLs Monitoring by Accelerometer-Based Wearable Sensors: Effect of Measurement Device and Data Uncertainty on Classification Accuracy

Abstract

Machine Learning algorithms are often used for automatic recognition and classification of Activities of Daily Living, and they rely on the computation of several features capturing the relevant characteristics of the collected signals, either in the time and frequency domains. While the accuracy of the measurement device used may be assessed by the manufacturer’s specifications or by specific tests, the propagated uncertainty of the computed features is typically not considered in the framework of automatic classification approaches. In this paper, the impact of the measurement devices on data quality, and consequently on the performance of automatic classifiers, is evaluated, in the context of accelerometer-based recognition of Activities of Daily Living with a wrist-worn device. Results show that different accuracy performance may be attained in the classification process, depending on the wearable device used, despite the same environmental and operational conditions.