Automation of the Timed Up and Go test using an instrumented walking cane
University of New Brunswick
The Timed Up and Go (TUG) test is used to test a person’s mobility and static and dynamic balance. It measures the time a person takes to stand up from a chair, walk three meters, turn around, walk back to the chair, and sit down. Typically, the TUG test is assessed by a physiotherapist with a stopwatch, limiting its effectiveness and making it prone to user error. This has motivated research into automated approaches capable of assessing the various segments of the TUG test using a range of sensing modalities. This study extends upon this body of work by evaluating the feasibility of segmenting the TUG test using an instrumented walking cane. More general contributions are made by introducing the use of error in transition time, as opposed to accuracy, as the cost function during the design of the machine learning framework, and a time-series inspired binary segmentation approach that facilitates the comparison of only two segments at a time. Data was collected using an instrumented cane that measures loading and movement information from 16 participants with musculoskeletal injuries. As a group, the participants yielded TUG times ranging from 11.12s to 28.57s, and a mean of 17.8s. Results of segmenting the TUG test into six segments - Sitting to standing, Walking, Turning, Walking back, Turning back, Standing to sitting - were validated using a leave-one-trial-out and a leave-one-person-out approach, to test both within- and across-participant performance. Various approaches were explored, including conventional classifiers Linear Discriminant Analysis (LDA) and Support Vector Machines (SVM), and extended time series and deep learning methods such as Hidden Markov Models (HMM), CNN LSTM (CLSTM) and Encoder-Decoder Temporal Convolutional Networks (EDTCN). A binary segmentation approach leveraging the temporal nature of the TUG test was adopted with a Dynamic Time Warping (DTW)-based postprocessing alignment. The calculated segmentation error for every case was recorded as both the performance measurement and the optimization parameter as opposed to the traditional use of accuracy of prediction. The results promisingly suggest that the segments or subtasks of a TUG test can be extracted using data collected from a smart cane, laying the groundwork for its automation.