A new parallel man-machine training approach to brain-computer interface (BCI) succeeded through a unique application of machine learning methods. The BCI system could train users to control an animated cursor on the computer screen by voluntary electroencephalogram (EEG) modulation. Our BCI system requires only two to four electrodes, and has a relatively short training time for both the user and the machine. Moving the cursor in one dimension, our subjects were able to hit 100% of randomly selected targets, while in two dimensions, accuracies of approximately 63% and 76% was achieved with our two subjects.

A method is developed for using neural recordings to control functional electrical stimulation (FES) to nerves and muscles. Experiments were done in chronic cats with a goal of designing a rule-based controller to generate rhythmic movements of the ankle joint during treadmill locomotion. Neural signals from the tibial and superficial peroneal nerves were recorded with cuff electrodes and processed simultaneously with muscular signals from ankle flexors and extensors in the cat's hind limb. Cuff electrodes are an effective method for long-term chronic recording in peripheral nerves without causing discomfort or damage to the nerve. For real-time operation we designed a low-noise amplifier with a blanking circuit to minimize stimulation artifacts. We used threshold detection to design a simple rule-based control and compared its output to the pattern determined using adaptive neural networks. Both the threshold detection and adaptive networks are robust enough to accommodate the variability in neural recordings. The adaptive logic network used for this study is effective in mapping transfer functions and therefore applicable for determination of gait invariants to be used for closed-loop control in an FES system. Simple rule-bases will probably be chosen for initial applications to human patients. However, more complex FES applications require more complex rule-bases and better mapping of continuous neural recordings and muscular activity. Adaptive neural networks have promise for these more complex applications.

Two machine learning techniques were evaluated for automatic design of a rule-based control of functional electrical stimulation (FES) for locomotion of spinal cord injured humans. The task was to learn the invariant characteristics of the relationship between sensory information and the FES-control signal by using off-line supervised training. Sensory signals were recorded using pressure sensors installed in the insoles of a subject's shoes and goniometers attached across the joints of the affected leg. The FES-control consisted of pulses corresponding to time intervals when the subject pressed on the manual push-button to deliver the stimulation during FES-assisted ambulation. The machine learning techniques used were the adaptive logic network (ALN) [1] and the inductive learning algorithm (IL) [2]. Results to date suggest that, given the same training data, the IL learned faster than the ALN, while both performed the test rapidly. The generalization was estimated by measuring the test errors and it was better with an ALN, especially if past points were used to reflect the time dimension. Both techniques were able to predict future stimulation events. An advantage of the ALN over the IL was that ALN's can be retrained with new data without losing previously collected knowledge. The advantages of the IL over the ALN were that the IL produces small, explicit, comprehensible trees and that the relative importance of each sensory contribution can be quantified.

We report on our advances in sensory feedback data processing and control system design for functional electrical stimulation (FES) assisted correction of foot drop. We have applied 2 methods of signal purification on the bin integrated electroneurogram (i.e., optimized low pass filtering and wavelet denoising) before training adaptive logic networks (ALN). ALN generated stimulation control pulses, which correspond to the swing phase of the impaired leg when dorsal flexion of the foot is necessary to provide safe ground clearance. However, the obtained control signal contained sporadic stimulation spikes in the stance phase, which can collapse the subject, and infrequent broken stimulation pulses in the swing phase, which can result in unpredictable consequences. In this study, we have introduced adaptive restriction rules (ARR), which are initially used as previously reported and then dynamically adapted during the use of the system. Our results suggest that ARR provide a safer and more reliable stimulation pattern than fixed restriction rules.