Presentation: Variable Selection in Functional Linear Model with Varying Smooth Effects
Speaker: Ana-Maria Staicu, Ph.D., Associate Professor of Statistics, North Carolina State University
Abstract: State-of-the-art robotic hand prosthetics generate finger and wrist movement through pattern recognition (PR) algorithms based on features from a large number of forearm electromyography (EMG) signals. This data-driven approach requires extensive training and is prone to poor predictions under conditions that do not exactly match those used for model training as shown in previous literature. We propose a novel functional data-based approach to prosthesis control by viewing EMG signals as functional measurements. The proposed functional linear model quantifies the effect of the EMG signals using a smooth bivariate coefficient that varies with additional covariate information. Unlike the feature-based PR algorithms, the proposed algorithm uses full EMG information and selects important EMG signals. The model is made parsimonious and interpretable through a two-step variable selection procedure which we call Sequential Adaptive Functional Empirical (SAFE) group LASSO. A final model is then fit on the selected subset to reduce shrinkage bias of the regression functions. Numerical study confirms great properties of the proposed methodology in terms of variable selection and prediction when compared to available alternatives. For our motivating dataset, the method helps to uncover correctly the EMG signals that are believed to be important for an able-bodied subject with negligible false positives.