Upper-Limb Force Modeling using Rotated Ensembles with Fast Orthogonal Search on High-Density Electromyography

Robust estimation of net limb forces has a plethora of important applications. However, the most reliable methods are invasive and are not feasible in most cases. The use of surface electromyography (EMG) for estimating muscle forces is non-invasive, and generally unobtrusive. However, many physiological and non-physiological factors limit the precision of these methods. The rich spatiotemporal information collected from high-density surface EMG (HD-EMG) grids has been used to mitigate the problem; however, the instability of electrode-skin contact for these systems creates outlying channels, which may lead to substantial error. We build non-linear predictors on rotated ensembles, which when aggregated provide robust and highly accurate (RMSE<2.6%) estimation of the force induced at the wrist under isometric contractions.