Pipeline for extraction of muscle synergies during isometric rehabilitation exercise, which we then compare across healthy and post-stroke participants.Abstract
Robot-mediated rehabilitation is a promising approach to motor therapy, but lacks sufficiently granular neuromusculoskeletal models to differentiate pathological individuals. Muscle Synergy Analysis (MSA) provides a low-dimensional but comprehensive representation of motor behavior, and is thus a promising framework with which to examine the differences across healthy and pathological individuals.
In this work, we perform preliminary MSA on data of 13 healthy and 2 post-stroke participants performing single- and multi-axis isometric rehabilitation tasks using a custom robot gaming platform (Anand et al. 2025), toward a model that represents pathology and its evolution over time.
Data from 16 tasks were decomposed into synergies via non-negative matrix factorization (NMF). To validate the number of synergies needed to represent human motion, we measured the Variance Accounted For (VAF) using numbers of synergies ranging from 1–8. We then compared the resulting synergies across participants and impairment levels.
All literature-based criteria used to evaluate synergy dimensionality indicated that ~4 synergies were required to represent human motion for both impaired and unimpaired individuals. On comparing each subject’s synergies to all others’ via both cosine similarity and a modified mean-squared error metric of our own devising, we did not find that they could be grouped in any meaningful way.
These negative results, though limited to our small-N, isometric data set, raise questions on the efficacy of NMF as a meaningful decomposition for evaluating post-stroke pathology. Future analyses will include exploration of additional factorization techniques (PCA, ICA, etc.), during expanded and more precisely delimited rehabilitation tasks, to obtain better insight into changes in motor behavior post-stroke.
Text has been modified from initial submission to reflect improved analyses at poster presentation time.
Gabriel Parra
Undergraduate, CS & Applied Mathematics
Chad A. Berghoff
Undergraduate, ME
My research interests include myocontrol, medical and rehabilitation robotics, advanced manufacturing, and rapid prototyping.
Ajay Anand
PhD Student, Robotics
I build robots to improve lives. My research interests include rehabilitation robotics, human sensing and modeling, medical robotics, and computer vision.
Laura A. Hallock
Professor of Mechanical Engineering
My research interests include robotic assistance and rehabilitation, neuromusculoskeletal sensing, and human–robot control interfaces.