Enhancing Automated Autism Detection With Improved Word Embeddings

Abstract

Autism Spectrum Disorder (ASD) is a developmental disorder that affects a significant amount of people. Unfortunately, ASD manifests in a large number of ways, meaning that diagnosing ASD is both time-consuming and inaccurate, which results in many children with ASD not being diagnosed until later childhood. One advantage of an early diagnosis is that it allows for early intervention, which typically leads to much better results. We investigated multiple different machine learning models as potential methods of predicting ASD in children from text segments provided by the child's caregiver. Two promising models are multilayer perceptrons and logistic regression. We investigated different hyperparameters for multilayer perceptrons, such as the number of layers and size of hidden layers. We then conclude that text descriptions of a toddler's behavior given by a caregiver are highly accurate for predicting autism when combined with a multilayer perceptron.