Improving signal reliability for on-line joint angle estimation from nerve cuff recordings of muscle afferents.

Closed-loop functional electrical stimulation (FES) applications depend on sensory feedback, thus, it is important to continuously investigate new methods to obtain reliable feedback signals. The objective of the present paper was to examine the feasibility of using an artificial neural network (ANN) to predict joint angle from whole nerve cuff recordings of muscle afferent activity within a physiological range of motion. Furthermore, we estimated how small changes in joint angle that can be detected from the nerve cuff recordings. Neural networks were tested with data obtained from ten acute rabbit experiments in simulated, on-line experiments. The electroneurograms (ENG) of the tibial and peroneal nerves were recorded during passive ankle joint rotation. To decrease the joint angle prediction error with new rabbit data, we attempted to pretune the nerve signals and re-trained the ANNs with the pretuned data. With these procedures we were able to compensate for interrabbit variability. On average the mean prediction errors were less than 2.0 degrees (a total excursion of 20 degrees) and we were able to predict joint angles from muscle afferent activity with accuracy close to the best-estimated angular resolution. The angular resolution was found to depend on the initial joint angle and the actual step size taken and we found that there was a low probability of detecting joint angle changes less than 1.5 degrees. We thus suggest that muscle afferent activity is applicable as feedback in real-time closed-loop control, when the motion speed is restricted and when the movement is limited to a portion of the joint's physiological range.