This full text paper was peer-reviewed at the direction of IEEE Instrumentation and Measurement Society prior to the acceptance and publication.
Gait Phase Detection from Thigh Kinematics using Machine Learning Techniques Johnny D. Farah, Natalie Baddour
Edward D. Lemaire
Department of Mechanical Engineering University of Ottawa Ottawa, Ontario, Canada
[email protected],
[email protected]
The Ottawa Hospital Research Institute Faculty of Medicine University of Ottawa, Ottawa, Canada
[email protected]
Abstract-
Intelligent
orthotic
devices
require
accurate
detection of gait events for real-time control. For orthoses that control the knee, an ideal system would only locate sensors at the thigh
and
knee,
thereby
facilitating
sensor
and
electronics
integration with the assistive device. To determine potential gait phase identification approaches, classification was implemented using
J-48
Decision
Tree,
Random
Forest,
Multi-layer
Perceptrons, and Support Vector Machine classifiers, along with 5-fold (5-FCV) and 10-fold cross validation (10-FCV). Knee angle, thigh angular velocity, and thigh acceleration were obtained from 31 able-bodied participants during walking (10 strides each). Strides were segmented into Loading Response, Push-Off, Swing, and Terminal Swing and features were extracted using a 0.1 second sliding window. Gait phase classification was performed with and without the knee angle parameter. J-48 Decision Tree with the knee angle parameter was ranked the best classifier due to its second highest classification accuracy of 97.5% and lowest mean absolute error of 0.014. Results without the knee angle parameter differed by only 0.5% and 0.003. Therefore, an inertial sensor with accelerometer and gyroscope output, located at the thigh,
is
a
viable
approach
for
classifying
gait
phases
for
intelligent orthosis control.
Keywords-Walking, Stride, Machine Learning, Thigh, Acceleration, Angular Velocity, Knee Angle, Cross Validation
I.
INTRODUCTION
Walking is a complex and dynamic system and an important daily living activity. Quantifying walking parameters by gait phase is necessary for deeper understanding of human locomotion and assistive device control. A stride can be divided into stance and swing phases, separated by initial contact (IC) and foot-off (FO) events (Fig. 1) that represent specific functions throughout the walking cycle [1]. Stance and swing phases are subsequently divided into loading response and terminal swing sub-phases. Gait patterns are characterized by kinetic and kinematic parameters associated with body or body-segment rotations and translational motion in planes of progression. Gait analysis has required complex systems, such as three dimensional motion capture and force-plate laboratory systems. Recently, ubiquitous modalities have been employed using inertial measurement units (IMU). IMU's typically output angular velocity and acceleration signals and provide useful information for gait analysis [2]-[6].
978-1-5090-2984-6/17/$31.00 ©2017 IEEE
Fig. 1. Gait cycle phases and events.
Pathological gait caused by knee extensor muscle weakness can inhibit an individual's ability to support themselves during stance. Knee-Ankle-Foot Orthoses (KAFO) are prescribed to provide support to the knee for persons suffering from knee extensor weakness. Unfortunately, KAFO inhibit knee flexion, causing users to develop irregular gait patterns as compensatory mechanisms such as hip-hike, lateral sway, and vaulting [7], which can be chronically detrimental to the body. Stance-Control-Knee-Ankle-Foot Orthoses (SCKAFO) have mechanical knee joints that are engineered to provide knee support during stance and permit free knee movement in swing. SCKAFO improve mobility and offer more natural gait compared to conventional KAFO [7]. Microprocessor controlled SCKAFO (M-SCKAFO) take advantage of various sensor systems to control the knee joint (ex. pressure sensors, IMU's, EMG [4]) and are more reliable compared to mechanically-controlled SCKAFO that require pre-set movements or knee extension moments to lock and unlock the knee [7]. Microprocessor controlled joints employ different control systems that use machine learning techniques [2], [5], [8] or rule-based control systems [3], [4], [6]. However, many M-SCKAFO require multiple sensors at the thigh, shank, and foot for effective stance-control. As a result, these walking-aid devices become increasingly expensive and difficult to personalize. An ideal control system would localize sensors at the thigh and knee, thereby allowing orthotists freedom to customize the shank and foot segments and possibly reduce overall system cost.
Integrating local sensor signals with machine intelligence algorithms should provide an effective method for gait phase identification that is feasible for application in real-time control. The objective of this paper is to determine whether accurate gait phase detection can be implemented using J-48 Decision Trees (J-48 DT), Random Forest (RF), Multi-Layer Perceptrons Neural Network (MLP), and Support Vector Machine (SVM) classifiers strictly using thigh-segment kinematics and knee angle during walking. This research also investigated thigh-only sensor localization by isolating thigh segment kinematics from the knee angle parameter. II.
METHODOLOGY
A. Equipment All walking signals were acquired with CAREN-Extended system (Computer Assisted Rehabilitation Environment) at The Ottawa Hospital Rehabilitation Centre. CAREN-Extended includes a 6 degree-of-freedom platform with integrated force measuring dual-tread treadmill, 1800 screen to project 3D virtual worlds, safety railings, and Vicon 3D motion capture system. A custom "walk through a park" application was developed and used in this study. A full body marker set was used to track 3D kinematics as participants walked on level sections within the 3D world. Visual3D (C-Motion) was used for all biomechanical analysis. Kinematic data were recorded at 100Hz and ground reaction forces (GRF) were recorded at 1000Hz.
B. Participants A de-identified dataset from 31 able-bodied volunteers was used for this study (16 males, 15 females, mass=75.8±13.2kg, height=1.73±0.12m, age=30±lOyears). Each participant walked at their own pace (self-paced). III.
events, the instant the foot leaves the ground indicating the onset of swing phase, were the times when the GRF passed a