On Sparse Functional Linear Regression for EMG Data Analysis
Jun 25, 2021 to Jun 25, 2021
Robotic hand prostheses require a prosthesis controller to translate electromyogram (EMG) signals into the user’s desired movement. State-of-the-art controllers must undergo extensive training on data from a large number of EMG sensors, whereas a biomechanical model for a single movement degree-offreedom shows that relatively few forearm muscles are needed to explain hand movement. A prosthesis controller based on such a biomechanical model should then require fewer EMG sensors to produce accurate predictions under a broad set of conditions. This talk will address three key components of the statistical collaboration on this project. First, I will discuss how understanding the biomechanical model led to developing a functional linear model with position-dependent effects. Second, I will describe the penalized estimation approach taken by Stallrich et al (2020) and identify two pitfalls of their approach: (1) the partial conflation of the sparsity and smoothness tuning parameters and (2) slow computational performance. I will then address these pitfalls by (1) investigating two alternative penalties that promote better exploration of the multidimensional tuning parameter space and (2) utilizing and extending the GLODE algorithm from Yau and Hui (2017) for solving the existing and newly proposed penalized estimation criteria. Variable selection and estimation performance of the competing methods will be compared via simulation and application to the EMG data.