Selection of wavelet coefficients for neural networks used in medical diagnosis

摘要

We use a neural network to diagnose patients with multiple sclerosis on the basis of their auditory evoked potentials. The wavelet transform coefficients of the evoked potentials are used as inputs to the network. SInce a limited number of clinical cases in available, only a handful of coefficients can be used to train the network. Several critieria are applied to select the most significant wavelet coefficients. Selections based on the Kolmogorov-Smirnov statistic, as well as methods using Shannon information are introduced. We present a comparison of the results obtained using thses criteria for different kinds of wavelets. Diagnosis of up to 96 percent accuracy have been achieved using the best coefficients.