• Braveena K Santhiranayagam, Daniel T H Lai, W A Sparrow, and Rezaul K Begg.
• Institute of Sport, Exercise and Active Living, Victoria University, Melbourne, Victoria 8001, Australia; College of Sport & Exercise Science, Victoria University, Melbourne, Victoria 8001, Australia. Electronic address: braveena.santhiranayagam@live.vu.edu.au.
• J Biomech. 2015 Dec 16; 48 (16): 4309-16.

AbstractFalls are the primary cause of accidental injuries (52%) and one of the leading causes of death in individuals aged 65 and above. More than 50% of falls in healthy older adults are due to tripping while walking. Minimum toe clearance (i.e., minimum height of the toe above the ground during the mid-swing phase - MTC) has been investigated as an indicator of tripping risk. There is increasing demand for practicable gait monitoring using wearable sensors such as Inertial Measurement Units (IMU) comprising accelerometers and gyroscopes due to their wearability, compactness and low cost. A major limitation however, is intrinsic noise making acceleration integration unreliable and inaccurate for estimating MTC height from IMU data. A machine learning approach to MTC height estimation was investigated in this paper incorporating features from both raw and integrated inertial signals to train Generalized Regression Neural Networks (GRNN) models using a hill-climbing feature-selection method. The GRNN based MTC height predictions demonstrated root-mean-square-error (RMSE) of 6.6mm with 9 optimum features for young adults and 7.1mm RMSE with 5 features for the older adults during treadmill walking. The GRNN based MTC height estimation method devised in this project represents approximately 68% less RMSE than other estimation techniques. The research findings show a strong potential for gait monitoring outside the laboratory to provide real-time MTC height information during everyday locomotion. Copyright © 2015 Elsevier Ltd. All rights reserved.

22 keywords...

hide…