Elastic Net Regression for Prosthesis Control in Short Residual Limb Amputees: Performance and Generalizability

Abstract: This Master's thesis in Biomedical Engineering investigates the performance and generalizability of linear regression models in context of prosthesis control for short residual limb amputees. This thesis uses intramuscular electromyography data, and a regression and emplys a regression technique called Elastic Net Regression - a technique that combines L1 and L2-regularization - to predict 1-DOF isometric forces outputs from fingers and the wrist. The elastic net not only functions as a regression model but also as a feature selector, which is especially useful with higher-order interaction terms. The aim of the thesis was not merely to create a working model with high performance metrics but also to possibly train a multi-channel model that can be readily used on a new amputee without need for recalibration. Another goal was to ensure the model remains transparent and easily interpretable. The results however, indicate that while the elastic net regression offers improved performance over standard single-channel models for the same subject, it struggled to generalize across different subjects, likely due to overfitting to individual subjects distinct characteristics. The elastic net regression model generally performed worse with lower R2-scores than the bare bones single-channel model when applying the model to new subjects.