Predicting Autism Disorder of an Imbalanced Dataset using Neural Network

Abstract

Autism disorder is a neurodevelopmental disorder that affects the ability to interact, communicate and learn. The nature of autism disorder is complex, requiring a team of several specialists to make an accurate diagnosis. Symptoms of autism can be detected in early childhood. However, autism is often not diagnosed until much later. The development of an accurate machine learning model aids in the early detection of autism, allowing for the reduction of long-term costs associated with delayed diagnosis. Many research works dealt with machine learning algorithms and techniques for predicting different types of diseases such as autism spectrum disorder. Most of these studies have reached good results in diagnosing autism, but there are still some weaknesses and insufficiencies. The datasets used in the previous studies are imbalanced, making the prediction process biased toward the majority class. Also, some works did not consider pre-processing the data and studying the correlation between the data as a preparatory step to the prediction process. In addition, there is still room for improving the accuracy achieved by the current studies using machine learning algorithms and techniques. Thus, the proposed approach in this study includes data cleaning, correlation analysis, oversampling, and parameter selection, dealing with the above shortages. This study applies artificial neural networks to identify patients with autism disorder. The results show that oversampling a dataset with 18 selected features and applying the multi-layer perceptron neural network algorithm with 20 neurons provides the highest Accuracy value of 99.1% compared to the other common and well-known algorithms. Keywords-Autism Spectrum Disorder, Multi-Layer Perceptron, Correlation Analysis, Oversampling