A Discriminative Model to Predict Comorbidities in ASD

Abstract

Children with autism spectrum disorders face a higher comorbidity burden than the general pediatric population. Machine learning applications have been applied to structured medical data, specifically electronic health records, to cluster ASD patients in terms of the trajectory of their comorbidities over time up to the age of 15 years old. However, social online forums related to ASD provide valuable information in terms of volume, verbosity, and context. Thus, this thesis leverages the unstructured text of posts on social ASD forums to binarily predict the presence of a medically-relevant concept in an author’s later posts on an age-identified ASD subject from those on that subject at a younger age. While this approach faces limitations in terms of verifying that the concepts and ages extracted from posts refer to the same ASD subject, it performed well against LDA topic modeling. However, it slightly under performed against dynamic topic modeling, thus providing a discriminative baseline against which to compare the gains made by a generative approach for this application.