Efficient training of neural network models in classification of electromyographic data

Abstract

THE APPLICATION of artificial neural networks (ANt'N) in the diagnosis of neuromuscular disorders based on electromyography (EMG) has recently been proposed (SCHtZAS et aL, 1990; PATTICHIS, 1992). Artificial neural network models have been trained to diagnose normal (NOR), motor neuron disease (MND) and myopathy (MYO) subjects successfully (PATTICHIS et al., 1990; PATTICHIS, 1992). The momentum back-propagation (MBP) training algorithm was used as proposed by Kumelhart et al. (RUMELHART et aL, 1986). The method has a number of limitations; heavy computational and memory requirements, as well as the non-existence of design methodologies for determining the values of the learning coefficient 2, and the momentum coefficient #, number of hidden layers, and architecture size. In addition, the algorithm can exhibit oscillatory behaviour or can even diverge, depending on the values of ~. and #. It is clear that the arbitrary method of choosing the values of and # should be eliminated in order to derive an improved learning algorithm. A method is required that automatically adjusts the values of ). and # so that the resulting algorithm is efficient and reliable. The conjugate gradient method (FLETCHER and REEVES, 1964; POLAK, 1971) is a class of methods for unconstrained optimisation, which basically fulfil this requirement and is based on a sound theoretical basis. With a fairly accurate line search algorithm, these methods are guaranteed to find a local minimum and with a fast rate of convergence. It has recently been demonstrated (BATTn'I, 1990; JOaANSSON et al., 1990; CHARALAMBOUS, 1992; MOLLER, 1993) that the performance of the conjugate gradient backpropagation neural network learning algorithm (CGBP) is superior to that of the standard MBP. The purpose of this study is to apply the CGBP learning algorithm in building neural network models excited with EMG data and compare the results to those obtained by the MBP learning algorithm.