Impact of Wearable Measurement Properties and Data Quality on ADLs Classification Accuracy

Abstract

In the field of automatic recognition and classification of Activities of Daily Living (ADLs), a paramount role to determine the classification accuracy is played by sensor technologies, as the algorithms’ performance is highly affected by the nature and quality of the collected measurement data. This work aims to investigate the influence of the wearable device characteristics and measurement uncertainty on the classification accuracy. For this study, two wearables devices are considered: a top-quality smartwatch (Empatica E4) and a low-cost Arduino-based wristband prototype. These devices have been used to measure the acceleration signal at the dominant wrist of subjects performing some relevant activities in real-life conditions. The experimental evaluation of some ADLs classification algorithms shows that their accuracy fluctuates depending on the choice of the sensor, which in turn affects the amount and type of relevant features to process. As such, the combination of features’ domain, i.e. time or frequency, number and type, which leads to the best classification accuracy has to be tuned on a specific sensor basis, despite the same type of signal, i.e. acceleration, is measured and processed under identical circumstances and processed. Accuracy values of 50-99% and 66-95% in the ADLs classification, are obtained for Empatica E4 and Arduino-based prototype, respectively; the best performance among classifiers is obtained with J48 and Random Forest, confirming that, with an appropriate configuration, satisfactory accuracy may be attained, even by resorting to the use of simple sensors.