UMATBrush: A dataset of inertial signals of toothbrushing activities

Abstract

Smartwatches and other commercially available wrist-worn devices have become a low-cost tool which, in recent years, has gained enormous popularity for monitoring habits associated with a healthy lifestyle. In this regard, the increasing computational power of smartwatches is facilitating the integration of complex machine learning and deep learning algorithms, which implement manual activity recognizers based on the inertial sensor signals that these wearables natively include. One specific application of such human activity recognition (HAR) systems is the monitoring of toothbrushing, aimed at fostering oral health habits among the population. For the evaluation and testing of these types of detectors, having access to databases of inertial signals captured by smartwatches is of paramount importance. This work describes the UMATBrush repository, which results from monitoring four experimental subjects during a large number of toothbrushing sessions using three commercial smartwatches. In contrast to other similar repositories, which are focused on the generic development of detectors for a limited set of manual activities, this repository also includes long periods of monitoring of the subjects during their daily lives. In the dataset, each acceleration sample captured by the watches is binary labelled as either corresponding or not to a toothbrushing session. In this way, potential classifiers using these traces could be trained and validated under realistic conditions, by learning to distinguish the toothbrushing operation from other real-life activities.