A robust real-time gait detection method for spinal cord injury rehabilitation

Abstract

In a recent first-in-human clinical study, a therapy based on targeted epidural electrical stimulation together with body weight support has been validated to restore voluntary locomotion in chronically paralyzed subjects with Spinal Cord Injury (SCI). However, the system is currently open-loop controlled, meaning that the stimulation pattern cannot be influenced by the patient. Introducing closed-loop control with real-timemotion feedback is expected to improve activity-based plasticity and thereby the therapy, to which this thesis will be contributing. To be specific, this thesis investigates the possibility of a robust real-time gait event detection method for SCI patients, to be used to synchronize the stimulation to movement in real-time. The gait of SCI patients deviates remarkably from the gait of a healthy subject. We identified several factors contributing to the deviation and studied them separately. Datasets were created by mimicking various walking scenarios observed from video references of patients from the ongoing clinical study. A robust real-time gait detection method with three variations of it – one using zero crossings of angular velocity in sagittal plane, another using high frequency contents extracted using wavelet transform and a third using additional trunk kinematics - is proposed and tested in this study with IMU signals from foot, shank and trunk. The results were promising when tested on themimicked datasets and was then tested on the limited patients available. For patient with motor incomplete paralysis (AIS-D), the method seemingly identified events correctly 9 out of 10 times both for TO and HS detection. The performance dropped noticeably for TO detection (to around an F1 score of 0.24) for motor complete patient (AIS-B) while still maintaining the HS detection performance (at around an F1 score of 0.83). We conclude that patient-group-specific strategies might be necessary and that motor complete patients may require additional motion intention recognition strategies. In addition to this, a method to emulate IMU signals from motion capture datasets when limited number of markers are used, is also presented in this study, so as to be able to convert the already existing motion capture datasets to IMU datasets, and to time-synchronize these datasets explicitly when they are collected using independent sensors. Correlation values as high as 0.977 were observed between the emulated IMU signals (sagittal plane angular velocity) and that of the actual IMU signals.