Neural Network Predictive Analysis of Echoic Behavior for Children with Autism

As an exploration of machine learning strategies applied to behavior analytic investigations, we sought to predict the likelihood that particular children with Autism Spectrum Disorder (ASD) would develop echoic verbal behavior. After retrieving 143 case records of children assessed with the VB-MAPP from a data repository, we reserved 22 case records for holdout analysis. Through iterative development, the model achieved 100% accuracy in predicting the presence or absence of an echoic repertoire. While the functional value of predicting a known outcome might be questioned, it is only possible to gauge the accuracy of any prediction technique when known outcomes are available. Thus, rather than trying to predict a completely unknown outcome, our experimental preparations were directed at gauging the multilayer perceptron’s (MLP) ability to predict at high levels of precision, sensitivity, specificity, and overall prediction accuracy while accessing a limited dataset from one children’s healthcare setting. We discuss procedures for employing this machine learning strategy in conjunction with a receiver operator characteristics analysis of our findings. We developed this study as a general guide for using one of several quickly evolving neural network methodologies for behavior analytic researchers, and we examine specifics regarding MLP functions and operations as they relate to providing more efficacious treatment for children with ASD.