Automated Detection of Autism Spectrum Disorder Symptoms using Text Mining and Machine Learning for Early Diagnosis

Abstract

Autism spectrum disorder (ASD) is a neurological condition whose etiology is still insufficiently understood. The heterogeneity of manifestations makes the diagnosis process difficult. Thus, many children are diagnosed too late, which leads to the loss of precious time that can be used for therapy. A viable solution could be to equip medical staff with modern technologies to detect autism in its early stages. The objective of this research was to investigate, through empirical means, how text mining and machine learning (ML) algorithms can aid in the early ASD diagnosis by identifying patterns and ASD symptoms in text data regarding children’s behavior that concerned parents provided. The research involved the design of an innovative technical solution based on text mining for the identification of ASD symptoms in unstructured text data describing children’s behavior and the practical implementation of the solution using Rapid Miner. The dataset was created through a controlled experiment with 44 participants, parents of children diagnosed with ASD, who answered questions about their children’s (35 boys and 9 girls) behavior. Analysis of the performance of models trained with ML algorithms: Naïve Bayes, K-Nearest Neighbors, Deep Learning and Random Forest revealed that the K-Nearest Neighbors classifier outperformed the other methods, achieving the highest accuracy of 78.69%. Results obtained using text mining and ML demonstrated the feasibility of using parents’ narratives to develop predictive models for autism symptoms detection. The achieved accuracy highlights the potential of text mining as an autonomous and time- and cost-effective method for early identification of ASD in children.