INTRODUCTION
Functional electrical stimulation (FES) can been used to provide walking in paraplegic patients. Improvement of these neural prostheses can be approached from several perspectives. Cosmesis and patient acceptance of the neural prosthesis can be addressed through the implantation of electrodes and stimulator. The use of multiple channels of muscle stimulation not only minimizes the requirement for external bracing, enhancing cosmesis, but also increases the patient's functional capabilities. The functional capability of the neural prosthesis can also be addressed through more effective use of stimulation.
The work presented here is part of a larger effort to explore how the effective use of stimulation can enhance neural prostheses for gait. The effective use muscle stimulation means tailoring the stimulation to each particular patient and activity. The muscle stimulation also requires adjustment to accommodate varying physiological and environmental conditions.
In many existing FES gait systems, the stimulus pulse width and frequency is modulated for each muscle channel to produce the desired gait. The stimulus pattern is based upon the muscle activity of normal gait and tailored to an individual patient by trial and error. This fixed stimulation pattern is then repeated each gait cycle, or walking stride, at the command of the patient [l]. The quality of gait is constrained not only by the limited strength of electrically stimulated muscle, but also by the limited precision in specifying muscle activation levels and timing to accommodate physiological and environmental variables. The sensitivity of paralyzed muscle to stimulation varies due to uscle length, short-term muscle fatigue, daily physiological changes and long-term muscle conditioning. Outside a closely monitored laboratory setting, the environment contains varying surface textures, slopes, obstacles and unexpected disturbances. One method to accommodate for these physiological and environmental variables is to adapt or modify the muscle stimulation in response to changes detected in the gait.
Feedback control systems for FES walking must take into consideration muscle latencies and the timescale of the desired gait. The time delay between stimulation and muscle force production (100-300 msec) may be long compared to the time spent in certain limb postures during gait [2]. By thC time the effect of the stimulation can be observed, it is too late to modify the it; the stimulation for future motions is already being sent. For this reason, the control system under development operates at the slower timescale of the gait cycle. This next-step, or cycle-to-cycle controller modifies the stimulation pattern for future gait cycles depending upon observations of the previous gait cycles.
In order to evaluate the quality of gait, the cycle-to-cycle controller requires, as a minimum, the knowledge of when each gait cycle begins and ends. Control parameters, or measures of the quality of gait, can then be compared to desired values, at certain times during the gait cycle. Traditional feedback controllers [3]or derived measures of the quality of gait over the past steps can be used to determine stimulation pattern modifications for the next cycle. For more precise evaluation of the gait, the timing of several gait phases is needed. The transition between each phase is a gait event. Detecting the times of the gait events is equivalent to determining the timing of the gait phases.
In this paper a gait event detector for use with FES walking is presented. The detector is comprised of two levels. The lower level is based upon a fuzzy logic phase of gait classifier [4, 51. The upper level consists of innovative supervisory rules that improve the performance of the gait event detector and determine the times when the gait events occurred. This multi-level gait event detector is implemented in real time. The accuracy of event detection is quantified for two paraplegic FES patients. An innovative aspect of this work is the demonstration that gait event detection can be performed in real time. The performance of the detector over time is further analyzed for one patient.