Get Up!: Assessing Postural Activity & Transitions using Bi-Directional Gated Recurrent Units (Bi-GRUs) on Smartphone Motion Data

Abstract

Many health conditions can affect a person’s mobility. Consequently, a person’s ability to perform transitions between activity states (e.g. sit-to-stand) are accurate measures of their mobility and general health. Mobility impairments can manifest either as discomfort while performing certain activity transitions or a complete inability to perform such transitions. The Timed up and Go (TUG) is an important clinical test that assesses patients’ sit-to-stand abilities. Research into passive methods to assess the quality of patients activity transitions and thus conduct the Timed Up and Go autonomously as they live their lives, have recently become popular. Machine and deep learning analysis of smartphone accelerometer and gyroscope data have demonstrated promising activity and transition recognition results. In this paper, we present Get Up!, a novel deep learning-based method to detect whether a person is performing a certain postural activity or transitioning between activities. Get Up! analyzes data from the accelerometer and gyroscope of the patient’s smartphone using Bi-Directional Gated Recurrent Units (Bi-GRU) neural networks with an attention mechanism. Our method outperforms TAHAR, the current state of the art machine learning method, achieving an error rate of 1.47% for activity classification and an accuracy of 97%. We also achieved an error rate of 0.17% with an accuracy of 93.3% when classifying postural transitions. As Get Up! segments activities and transitions, individual TUG sub-components can be timed to identify sub-components that patients find challenging.