University of Twente Student Theses


Estimation of lower limb joint kinematics using electromyography and machine learning

Herijgers, Joost (2020) Estimation of lower limb joint kinematics using electromyography and machine learning.

[img] PDF
Abstract:This study compared different algorithms that estimate lower limb joint kinematics from electromyography (EMG) in different activities. The ultimate goal is to give transfemoral amputees intuitive control over a powered prosthesis in a non-weight-bearing situation. Objectives of this study included an analysis of the machine learning approach in (non-) weight-bearing tasks and determining performance of the approach when applied to transfemoral amputees. Additionally, the influence of inclusion of (historic) information from inertial measurement units (IMUs) was studied. Three datasets were analysed to complete the different research objectives. The analysed activities included non-weight-bearing tasks, sit-to-stand transitions, level ground walking (for able-bodied subjects and transfemoral amputees), stair ascent and stair descent. Eight different algorithms were tested on their ability to estimate lower limb joint angles. An optimal set of hyperparameters for each algorithm was found using a Bayesian optimisation routine, on an activity-generic level. Performance of the different algorithms was analysed using a 5-fold cross-validation routine on a subject-specific level. A convolutional neural network gave the best performance in terms of R2 and RMSE using only EMG data in most tested activities. Including historic information from IMUs significantly increased performance (p < 0.05) for most of the studied activities. Exclusively using the same historic data from IMUs resulted in a significant decrease in performance for several of the studied activities. The developed approach showed to be feasible to apply to transfemoral amputees, as comparable performance is seen for amputees and able-bodied subjects in level ground walking. Results for non-weight-bearing tasks in able-bodied subjects were promising (R2 of 0.956 ± 0.013 for the knee angle). Therefore, further research could focus on studying the applicability of non-weight-bearing tasks in transfemoral amputees. Additional future research could focus on using different methods to extract features from both the EMG and IMU signals to increase performance of the algorithms.
Item Type:Essay (Master)
Faculty:TNW: Science and Technology
Programme:Biomedical Engineering MSc (66226)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page