Autism diagnosis requires validated diagnostic tools employed by mental health professionals with expertise in autism spectrum disorders. This conventionally requires lengthy information processing and technical understanding of each of the areas evaluated in the tools. Classifying the impact of these areas and proposing a system that can aid experts in the diagnosis is a complex task. This paper presents the methodology used to find the most significant items from the ADOSG tool to detect Autism Spectrum Disorders through Feed-forward Artificial Neural Networks with back-propagation training. The number of cases for the network training data was determined by using the Taguchi method with Orthogonal Arrays reducing the sample size from 531, 441 to only 27. The trained network provides an accuracy of 100% with 11 different cases used only for validation, which provides a specificity and sensitivity of 1. The network was used to classify the 12 items from the ADOS-G tool algorithm into three levels of impact for Autism diagnosis: High, Medium and Low. It was found that the items "Showing", "Shared enjoyment in Interaction" and "Frequency of vocalization directed to others", are the areas of highest impact for Autism diagnosis. The methodology here presented can be replicated to different Autism diagnosis tests to classify their impact areas as well.