Mapping of sensory representation of walking and EMG of prime joint movers: Control of functional electrical stimulation

Abstract

This paper presents machine learning (ML) techniques for development of a control scheme to be used in functional electrical stimulation (FES) of hemiplegic walking. The goal is to make an electrical stimulation pattern by mapping the sensors signals acquired during walking (input) to activities of muscles (output) acting around knee and ankle joints. Two machine learning techniques with ability of time series prediction were analyzed: a nonlinear autoregressive neural network (NARX) and an adaptive-network-based fuzzy inference system (ANFIS). Networks were compared in terms of minimum number of sensors needed for accurate prediction, timing errors, false detections and generalization ability. ANFIS network predicted more accurately, while NARX network needed less sensors, had less false detections and better generalization.