Aleksandar Kostov

Sensory nerve recording for closed-loop control to restore motor functions

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 cats 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 the authors designed a low-noise amplifier with a blanking circuit to minimize stimulation artifacts. They 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.<<ETX>>

Machine learning in control of functional electrical stimulation systems for locomotion

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 subjects 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) and the inductive learning algorithm (IL). 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 ALNs 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.<<ETX>>

Improved Methods for Control of FES for Locomotion

Abstract Two methods of generating the rules for control systems for control of functional electrical stimulation (FES) in locomotion of subjects with incomplete spinal cord injury were studied and qualitatively compared. The aim of the rule-based control system was to synthesize and replace the manual FES-switching function operated by the subject or a skilled physiotherapist. The first method, “hand-crafting” of rules, requires very detailed signal analysis by an experienced person, which may result in systems that are limited either by their functionality or safety for the subject. The second method, i.e. automatic generation of rules through learning from examples by a machine learning program, does not have such limitations because it is capable of learning everything that is presented to it during the training phase. It is obvious that for simple control systems, “hand-crafted” rules are a fast and simple solution, but for complex multichannel FES applications automatic generation of the rules can save many hours of trial-and-error experiments with hand-crafted rules.

Learning of EMG-patterns by adaptive logic networks

An adaptive logic network (ALN) has been used for nonparametric identification of the system consisting of two peripheral nerves and two muscles in the freely moving cat. The aim of this identification was to design a rule-based control for a functional electrical stimulation (FES) of the cats hind limb. We recorded from peripheral nerves and muscles while a chronic cat walked an a powered treadmill. We noticed a very reproducible firing in tibial and superficial peroneal nerves related to patterns of EMG activity in medial gastrocnemius (MC) and anterior tibialis (AT) muscles in the cats hind limb with respect to a phase of the gait cycle (e.g. beginning stance, beginning swing). We succeeded to restore and predict EMG activity of MG and AT muscles from neural recordings after applying Adaptive Logic Networks (ALNs). ALN is a type of artificial neural network having a tree configuration and using a Boolean operations in its nodes. We supplied a training set consisting of coded recordings from nerves and muscles (20 cats steps were sampled at 50 Hz) for supervised training. Applied encoding and training parameters resulted in a reasonably short training time and high correlation between an originai and predicted EMG signals. The variance accounted for (VAF) by the ALN prediction of the test data was 80% qualify- ing this technique as a good candidate for implementation in real-time FES control systems.

Evaluation of Adaptive Logic Networks for control of walking in paralyzed patients

An Adaptive Logic Network (ALN), a type of Neural Network (NN), as evaluated for the control of walking in Spinal Cord Injured (SCI patients. The motivation behind this research was to explore more reliable methods for control of simple Functional Neuromuscular Stimulation (FNS) systems in incomplete SCI patients. The ALN was used to recognize a patients intention to make a step by stimulating muscles in a partially paralyzed leg. Signals from four force sensors, installed under the toes and heels, have been used as inputs and a gating pulse associated with the stimulation as an output for learning and testing the function of the control system. Manual control, by either the patient or a physiotherapist, has been used as a template to be matched by the ALN. Generalization of the learned functions by the· ALN to previously unseen data was also tested. Finally, we manipulated the number of input channels, the inclusion of information from past samples and prediction of future events. The ALN is capable of generating the same time series of output pulses as those generated by “human experts.” Furthermore, it can predict the stimulation event early enough so that the requirement for stimulation can be verified and the patient informed to prepare for stimulation.

Gait event discrimination using ALNs for control of FES in foot-drop problem

Discrimination of stance and swing phases of the gait is required for control of functional electrical stimulation (FES) used to assist with ankle dorsiflexion in foot-drop problem. Simple thresholds applied to a human whole nerve signal processed using a sophisticated digital signal processing technique did not result in a safe and reliable control method. In this preliminary study, the authors use the same sensory signals to evaluate a gait event discriminator (GED), based on Adaptive Logic Networks (ALNs). The evaluation was performed off-line using neural signals for sensory feedback and a signal from a heel switch as the output to the stimulator. The neural signal was recorded using a cuff electrode implanted around the calcaneal nerve in the left leg of a male subject and the heel switch was installed inside the shoe of the same leg. Preliminary results suggest that ALNs can discriminate precise timing of heel contact and heel lift during FES-assisted walking. Restriction rules based on a priori knowledge were used to verify decisions made by ALNs and to eliminate infrequent functional errors providing maximum safety for the subject.

Gait event and user intention detection for FES-control: selecting sensors

The authors follow a design method for event detectors using the ID3 rule induction algorithm. Rule induction was chosen mainly for two reasons: it ranks the relative importance of sensor signal attribute in detecting an event and, secondly, the reasoning of the algorithm may be understood by humans since the rules are organized in the familiar form of decision tree consisting of IF(...) THEN(...) ELSE(...) statements. This method allows the control system designer the freedom to position a set of available sensors in unobtrusive locations, such as braces, walking aids or the waistband, and operate them in less demanding environments. Furthermore, the method does not require a high level of intuition as to the contribution that each sensor makes to the detection of an event. Indeed, it has been shown that human experts perform poorly relative to the algorithm in ranking the importance of the sensors (C.A. Kirkwood, and B.J. Andrews, Proc. 11th IEEE EMBS Conf., Seattle, USA, p. 1020-1, 1989). Here, the authors describe a procedure in which a reliable event detector/predictor can be developed with a minimum of sensors. It will mimic a paraplegics skill in using hand switches to control a simple FES walking system, i.e. it will signal the users implicit intention. This example of skill cloning follows that previously described (Kirkwood and Andrews, 1989).

Machine learning in control of functional electrical stimulation for locomotion

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 map the relationship between sensory information and the FES-control signal by using off-line supervised training. Signals were recorded using pressure sensors installed in insoles of a patients shoes and goniometers attached across the joints of the affected leg. The FES-control signal consisted of pulses corresponding to time intervals when the patient pressed on the manual push-button to deliver the stimulation during FES-assisted ambulation. The machine learning techniques evaluated were the adaptive logic network (ALN) and inductive learning algorithm (IL). Results to date suggest that, given the same training data, the IL learned faster than ALN while both performed the test rapidly. The generalization 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 ALN was that it can be retrained with new data without losing previously collected knowledge. The advantages of IL were that IL produces explicit and comprehensible trees and that the relative importance of each sensory contribution can be quantified.<<ETX>>