Constrained particle filtering for movement identification in forearm prosthesis

Abstract

We formulate the problem of movement identification for the forearm prosthesis using a nonlinear state-space system and the hypothesis of muscle synergies. The synergy activation coefficients contain task-specific information and can be used to identify limb movements. In the proposed framework, the measurements are EMG data and the system state consists of muscle synergy activation coefficients, which are physiologically constrained to be nonnegative on average. Particle filters are the state-of-the-art techniques for optimal state estimation in nonlinear and non-Gaussian systems. However, the very numerical nature of the particle filters, which constitutes their strength, becomes their major weakness in handling constraints on the state. In this paper, we solve the movement identification problem by introducing a constrained particle filter termed as mean density truncation (MiND). We show that MiND minimally perturbs the unconstrained distribution of the state while simultaneously satisfying the desired constraints on the unknown state. We recorded EMG data from forearm muscles of 12 participants for identification of hand and wrist movements. The proposed particle filtering with MiND provided an accurate stream of synergy activation coefficients (p < 0.001) which were used for movement identification with error rates significantly lower (p < 0.05) than currently used heuristics and Linear Discriminate Analysis.