Estimating Range of Lower Body Joint Angles with a Sensorized Overground Body-Weight Support System Chun Kwang Tan *1 , Bruno Leme 1 , Eleuda Nunez 1 , Hideki Kadone 2 , Kenji Suzuki 1 , Masakazu Hirokawa 1 Abstract— Recent trends in rehabilitation and therapy are turning to data-driven approaches to personalize treatment. Due to such approaches, data collection methods have become more complex and expensive, in terms of financial resources, technological knowledge, and time required to implement the data collection method. Such costs might deter clinical applications of otherwise good data collection methods. Hence, a method to collect data in a non-intrusive manner is proposed. Sensors are embedded into a commonly used rehabilitation tool, the walking trainer, for gait data collection. This study shows that, in principle, lower body joint angles can be collected in a non-intrusive manner, with a slight trade off to precision. In this study, the focus would be on the pelvic and hip movements, since the pelvic segment of the human body is implicated in a variety of gait problems Clinical relevance — The proposed usage model allows clinicians access to additional kinematic data, while minimizing changes to existing clinical evaluation processes and being non- intrusive. Having additional kinematic data would give further insight into a patient’s current state, thereby improving the efficiency of individualized therapy. I. I NTRODUCTION Current trends in medicine and physical therapy indicates that data-driven approaches to personalize interventions are becoming more popular [1], [2], [3]. Although these approaches might have proven to be useful in the laboratory, they require expensive equipment and profound technical knowledge. In the case of the gold standard in gait analysis, the motion capture system (Mocap) represents a huge cost, both financially and technologically, on the clinicians/physical therapists. This might deter applications of objective methods in gait evaluation. Recent works have proposed the use of wearable devices, like IMUs, to alleviate the problem of cost and ease of use [4], [5]. Although IMUs are cost-effective and have reasonable precision, attaching sensors to patients involves an extra step in the clinical evaluation process, which might increase time spent for evaluation. Therefore, it is proposed that data collection should be performed in a non-intrusive manner, by integrating sensors into commonly used rehabilitation devices. In this paper, an overground Body Weight Support (BWS) walking trainer was selected as a candidate to examine the feasibility of such a data collection method. * Corresponding Author: chunkwang@ai.iit.tsukuba.ac.jp 1 University of Tsukuba, Faculty of Engineering, Information and Sys- tems 2 Center for Cybernics Research, Faculty of Medicine, University of Tsukuba This paper examines the feasibility of estimating the pelvic and hip range of joint angles (RoM) with a sensorized BWS trainer [6]. Such a device would be able to collect additional gait information that would help clinicians to better understand the patient’s current condition. The hips and pelvis were the focus for this paper as the pelvis is anatomically important for human gait [7], and constrains to pelvic movement is correlated to the reduction in the RoM of the lower limbs [8]. Furthermore, it has been shown recently that assisting pelvic movement during rehabilitation can improve gait recovery in stroke patients [9]. As pelvic movement play such a central role in gait, being able to collect more data on this phenomena would greatly enhance understanding of how the pelvis affect gait. II. MATERIALS AND METHODS There are two objectives of this study. First, is to validate a method based on a neural network model to estimate the RoM of pelvic tilt, upward obliquity and pelvic rotation. The second objective is to validate another method based on an inverted pendulum model to estimate the RoM of hip flexion. A. Experiment Location Data collection was conducted in a large room with a 10m long walkway. The Vicon Motion capture (VICON MX System with 16 T20S Cameras, Vicon, Oxford, UK, sampled at 100 Hz) was installed in the room for motion capture. B. Sensorized Overgound Walking Trainer A walking trainer (All-in-One Walking Trainer, Ropox A/S, Naestved, Denmark), with attached sensors, was used for participants (Figure 1). Briefly, strain gauges (SG) were added to the lifting arm of the walking trainer to measure the amount of body weight unloaded by the frame of the walking trainer. A Laser Range-Finder (LRF) attached to the under- carriage of the walking trainer (39 cm above the ground), measures the distance of the shin to the undercarriage. This allows us to calculate the step length of the participant. For more information on exact sensor placement and verification tests of on the sensorized walking trainer, please refer to [6]. LRF and SG were sampled at 80 Hz. C. Software The Deep Learning Toolbox (Version 13.0) in Matlab 9.7 (R2019b Update 8) (The MathWorks, Inc., Natick) was used to implement a Bidirectional Long-Short Term Memory (LSTM) network model. This network model was used to 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC) Oct 31 - Nov 4, 2021. Virtual Conference 978-1-7281-1178-0/21/$31.00 ©2021 IEEE 4932