ABSTRACT
Persons with spinal cord injury are generally at least partially paralyzed and are often unable to walk. Some are able to use manually controlled electrical stimulation to act upon nerves or muscles to cause movement of a paralyzed leg so functional walking is achieved. They use crutches or a mobile walker for support, and control stimulation by pressing a switch, usually installed on the walking aid. Machine learning techniques are now making it possible to automate this control. Supervised training can be based on samples of correct stimulation given by the user (e.g. the subject or a resercher), accompanied by data from sensors indicating the state of the person’s body and its relation to the ground during walking. A major issue is generalization: whether the result of training can still be used for automatic control after the passage of time or in somewhat different circumstances. As it becomes possible to increase the number and variety of sensors used and to easily implant more numerous stimulation channels, the need is increasing for fast and powerful learning systems to automatically develop effective and safe control algorithms. In the present study, adaptive logic networks were used to develop an experimental walking prosthesis. Successful generalization has been observed up to several days after training.
