A neural network representation of electromyography and joint dynamics in human gait

Optimization theory and other mathematical algorithms have traditionally been used to model the relationship between muscle activity and lower-limb dynamics during human gait. We introduce here an alternative approach, based on artificial neural networks with the back-propagation algorithm, to map two different transformations: (1) EMG-->joint angles; and (2) EMG-->joint moments. Normal data for 16 muscles and three joint moments and angles (hip, knee, and ankle) were adapted from the literature [Winter (1987), The Biomechanics and Motor Control of Human Gait]. Both networks were successfully trained to map the input vector onto the output vector. The models were tested by feeding in an input vector where all 16 muscles were slightly different (20%) from the training data, and the predicted output vectors suggested that the models were valid. The trained networks were then used to perform two separate simulations: 30% reduction in soleus activity; and removal of rectus femoris. Net 2, in which electromyography was mapped onto joint moments, provided the most reasonable results, suggesting that neural networks can provide a successful platform for both biomechanical modeling and simulation. We believe that this paper has demonstrated the potential of artificial neural networks, and that further efforts should be directed towards the development of larger training sets based on normal and pathological data.