A Machine Learning Approach to Predict Autism Spectrum Disorder

Autism spectrum disorder, often shortened to ASD, is a developmental disability caused by differences in the brain. Individuals with ASD display deficits in social communication and interaction as well as repetitive behaviors or interests among a list of other symptoms [1]. If left untreated, individuals with autism may observe no progression or a worsening of their symptoms as untreated ASD can increase the risk of the development of depression and anxiety. The best outcomes are expected when the individual is diagnosed as early as possible allowing for early intervention and the provision of therapy and treatments. ASD can be detected in individuals as young as toddlers. Much research has been conducted into the application of machine learning in order to predict ASD in many different age groups, with the category of toddlers having the least dedicated research. This paper aims to explore the application of different machine learning techniques in order to predict ASD in toddlers. Two different toddler datasets were used to train and test the models. Many different traditional machine learning models, ensemble models, and deep learning models were considered in the scope of this research and compared on the basis of performance. K-fold cross validation was applied, and the models were tested. The most positive results in terms of the highest accuracy recorded were seen in ensemble learning which near consistent positive results being seen for all the different experimented techniques. Further favorable results were seen in particular with random forest (100%), decision trees (100%), C4.5 (100%) and logistic regression (100%). In terms of deep learning the most positive results were seen in multilayer perceptron (100%), recurrent neural networks (98.6%), and multilayer kernel machines (100%).