A neural net representation of experienced and nonexperienced users during manual wheelchair propulsion

A neural net approach was used to classify and analyze combinations of the physiological and kinematic responses (the factor patterns) of experienced and novice individuals during wheelchair propulsion, and to determine the key characteristics (individual factors) used in making this determination. A sequence of artificial neural networks (ANN) was developed and used to classify differences between eight nonimpaired controls and seven individuals using wheelchairs, who ranged in age from 24 to 36 years. The subjects propelled a wheelchair on a specially constructed dynamometer at three different velocity levels during which stroke pattern, force, energy, and efficiency data were collected. The data from 10 subjects (5 from each group) were used to train a net, with the data from the remaining 5 subjects used to test the resulting net. The nets correctly classified the training subjects in all 10 cases and correctly classified all 5 test subjects, indicating that the developed networks were able to generalize to new data sets. It was concluded that a minimal net consisting of only three variables, peak VO2 at the high velocity, hand force on the rim at the low velocity, and push angle at the high velocity, could accurately represent the differences between these groups.