Feasibility of Gait Event Detection from Thigh Segment Kinematics in Children with Cerebral Palsy

Introduction: Wearable robotic exoskeletons have been shown to improve walking in children with movement disorders such as cerebral palsy (CP). For these users, the exoskeleton must track limb movements in real-time to deliver precisely timed and scaled torque at each joint. In single joint exoskeletons, the predominant approach is to split the walking cycle into discrete phases by detecting foot-strike and foot-lift-off gait events. Many gait event detection algorithms, including the one developed in our laboratory at the NIH, use Force Sensitive Resistors (FSRs) underneath the exoskeleton’s footbed to determine when the foot contacts the ground. FSR-based gait event detection is suboptimal due to fragile hardware and delays between actual gait events and detection (approximately 150 ms). Using kinematics to detect gait events may offer more robust, precise detection than an FSR because foot loading/unloading is not needed to identify the change in contact state with the ground. Previous literature has found sagittal plane thigh angle has a local maximum at initial foot contact, while hip joint angular velocity peaks during toe-off phase of walking in healthy individuals. The goals of this research were to 1) confirm viability of using highly accurate thigh segment kinematics to detect gait events during overground walking with motion capture and 2) determine feasibility of gait event detection using estimated thigh segment kinematics from an Inertial Measurement Unit (IMU) attached to the thigh of children with CP. Successful IMU-based gait event detection would provide impetus for its eventual use in real-time control of robotic exoskeletons.

Materials and

Methods: Previously collected motion capture (n=3, M=3/F=0, average age=11.4) and IMU kinematics data (n=3, M=3/F=0, average age=12.0) collected during overground walking were used in two sequential analyses. All participants completed informed consent and were previously diagnosed with CP and impaired movement (GMFCS II(4)-III(2)). In both analyses, ground truth gait events were determined by visually inspecting digital video of overground walking and recording the timing of each event. For the first analysis, a peak detection algorithm was created to estimate heel-strike timing from motion capture sagittal plane thigh angle (Figure 1A). Toe-offs were not reliably predicted using sagittal plane data alone, so a weighted angular velocity using sagittal and frontal plane data was created to improve toe-off detection. A grid search was conducted across 25 sagittal-frontal plane weighting pairs using timing accuracy metrics to identify optimal weights for toe-off detection.

In the second analysis, gait event detection using data from an IMU placed over the rectus femoris muscle was evaluated. Raw angular velocity was integrated to estimate thigh angle. To combat integration drift, a novel peak detection algorithm was developed. Peak prominence was calculated as the magnitude difference between each thigh angle peak and the thigh angle magnitude from 400 milliseconds (a tunable parameter) before each peak. Peaks with differences greater than a participant-dependent threshold were retained as a detected heel-strike. This algorithm (Figure 1B) was applied to our IMU-estimated thigh angle to determine heel-strike detection accuracy. Another grid search was performed to select the optimal toe-off weighting for IMU-based detection.

Results, Conclusions, and Discussions: This research piloted the use of thigh segment kinematics for accurate, reliable detection of gait events in children with CP. Thigh angle and angular velocity peaks aligned well with ground truth heel-strikes and toe-offs (Fig. 2), suggesting viable use for gait event detection. Heel-strikes were accurately predicted by motion capture sagittal plane data alone, indicated by an average absolute timing error of 68 ms. For toe-off detection, the weighting pair of [1 sagittal plane, -1 frontal plane] minimized mean absolute timing error (104 ms) and error rate (0.58%) while still obtaining a relatively low mean standard deviation (68 ms) across toe-off events. This weighting pair was optimal for both motion capture and IMU-based detection. Overall, motion capture data (“ground truth” kinematics) yielded a mean absolute timing error under 130 ms across all gait events. As expected, detection using estimated kinematics from IMU data was less accurate (mean absolute timing error < 150 ms across all gait events). On average, IMU-based detection of heel-strike and toe-off lagged behind ground truth gait events. False positive and false negative rates were < 3% in IMU-based detection. Therefore, thigh segment kinematics-based gait event detection may have similar, if not improved, timing accuracy compared to the current FSR system used in the NIH pediatric robotic exoskeleton.

The accuracy and precision of our algorithm for detecting gait events indicates viability for use in real-time control of robotic exoskeletons. However, because the algorithm was applied offline, adaptations are needed for real-time implementation. IMU-based gait event detection from different locations, such as the knee joint, should be investigated to determine if an IMU integrated into the exoskeleton architecture has a comparable accuracy as the IMU in this study, which was placed directly on the limb. If precise gait event detection from an IMU within the knee exoskeleton is infeasible, incorporating an IMU into an existing exoskeleton strap across the thigh is warranted. Overall, kinematics could offer more accurate and robust gait event detection with a lower risk of damage compared to currently deployed FSR hardware, thereby enabling long term gait training with the exoskeleton to be delivered outside clinical settings.

Acknowledgements (Optional): This research was supported by the Intramural Research Program of the NIH Clinical Center. We thank Chris Stanley for his contribution to motion capture data processing.