| 제목 |
Analysis of gait velocity and pattern by plantar pressure measurement using machine learning method. |
| 소속 |
Department of Rehabilitation Medicine, College of Medicine, Chungnam National University, Daejeon, Republic of Korea1, Biomedical institute, Chungnam National University, Daejeon, Republic of Korea2, National Institute for Mathematical Sciences, Daejeon, Republic of Korea3, Department of Rehabilitation Medicine, College of Medicine, Chungnam National University Sejong hospital, Sejong, Republic of Korea4 |
| 저자 |
Jun-Hyeong Han1*, Jin-seon Park2, Chang-Won Moon1, Il-Young Jung1,4, Kang Hee Cho1,2†, Woo-Sik Son3, O-Kyu Kwon3, Yeong-Jin Kim3 |
Objective
Activity monitoring systems are widely used in daily life to measure physical activities objectively and assist the promotion of health management.
Subjects & Methods
In this study, plantar pressure and energy expenditure data were gathered from 175 study subjects (age groups of 20-30, 50-64, and over 65 years old) in each of the following ambulation status:standing, walking, running, stair ascending and descending. The speed of walking ranged between 2-6 km/h while the running speed was between 6-9 km/h. For other ambulation status, the subjects were asked to ambulate at the pace with which they were comfortable. For the assessment, the following tools were used: gait phase, force, area, pressure were measured with the in-shoe pressure measuring system, Pedar-X (Novel GmbH, Munich, Germany); and energy expenditure (kcal/min) was measured with the portable gas analysis system, K4b2 (Cosmed S.r.I, Rome, Italy). Linear regression was used for statistical analysis. We tried to use three machine learning method to analyze a lot of plantar pressure variables. Firstly, Random forest(RF) was used to analyze gait pattern, according to DSR(Double support ratio) gait parameter threshold. Secondly, T-distributed Stochastic Neighbor Embedding(t-SNE) was used to predict gait pattern and stance and interval time parameter used to visualize gait pattern. Lastly, We tried to use multidimensional scaling(MDS) to visualize plantar pressure change Using 198 channel pressure data.
Result
The prediction of ambulation speed revealed adjusted R-squared values of 0.88 for walking and 0.99 for running. The traditional algorithm developed in this study calculated stance time accurately according to various ambulation status. Error percentage was lowest value(0.75%) using Random forest using variable gait parameters(DSR threshold = 25)(Table 1). In T-SNE, The clusters have lowest average stance time is definitely exist(blue dot) and gait pattern prediction accuracy was 0.86(Figure 1). We convert 188 channel into 2-D data to visualize and analyze intuitionally using MDS. We found that gait is a cylclic movement and trajectory is different depending on gait pattern(Figure 2).
Conclusion
It is important to categorize the ambulation status of wearers by which energy expenditure is measured accurately. Maching learning is helpful to analyze the accurate energy expenditure measurement according to ambulation status and speed. |
|
Random forest
T-distributed Stochastic Neighbor Embedding
multidimensional scaling
|
