The paper describes how the use of statistical tools and the implementation of neural networks have been jointly exploited to process PD patterns derived from experimental data relevant to electrical treeing phenomena in EPR compounds. We show that the overall output of the proposed analysis method allows us to discriminate between different materials, in which the treeing structures are present. In particular, we present evidence that a NN approach can be successfully adopted to select an appropriate subset of ten parameters (out of 56 different quantities computed applying to the "raw" data statistical tools). Such a subset can then be used as a new input data set to train/test other neural networks with comparable error rates.