Résumé
One of the challenging issues in computational rehabilitation is that there is a large variety of patient situations depending on the type of neurological disorder. Human characteristics are basically subject specific and time variant; for instance, neuromuscular dynamics may vary due to muscle fatigue. To tackle such patient specificity and time-varying characteristics, a robust bio-signal processing and a precise model-based control which can manage the nonlinearity and time variance of the system, would bring break-through and new modality toward computational intelligence (CI) based rehabilitation technology and personalized neuroprosthetics. Functional electrical stimulation (FES) is a useful technique to assist restoring motor capability of spinal cord injured (SCI) patients by delivering electrical pulses to paralyzed muscles. However, muscle fatigue constraints the application of FES as it results in the time-variant muscle response. To perform adaptive closedloop FES control with actual muscle response feedback taken into account, muscular torque is essential to be estimated accurately. However, inadequacy of the implantable torque sensor limits the direct measurement of the time-variant torque at the joint. This motivates the development of methods to estimate muscle torque from bio-signals that can be measured. Evoked electromyogram (eEMG) has been found to be highly correlated with FES-induced torque under various muscle conditions, indicating that it can be used for torque/force prediction. A nonlinear ARX (NARX) type model is preferred to track the relationship between eEMG and stimulated muscular torque. This paper presents a NARX recurrent neural network (NARX-RNN) model for identification/prediction of FES-induced muscular dynamics with eEMG. The NARX-RNN model may possess novelty of robust prediction performance. Due to the difficulty of choosing a proper forgetting factor of Kalman filter for predicting time-variant torque with eEMG, the- presented NARX-RNN could be considered as an alternative muscular torque predictor. Data collected from five SCI patients is used to evaluate the proposed NARX-RNN model, and the results show promising estimation performances. In addition, the general importance regarding CI-based motor function modeling is introduced along with its potential impact in the rehabilitation domain. The issue toward personalized neuroprosthetics is discussed in detail with the potential role of CI-based identification and the benefit for motor-impaired patient community.