JMIR Publications

ABSTRACT

Background:

The increasing usage of social media platforms has given rise to an unprecedented surge in user-generated content with millions of users sharing their thoughts, experiences, and health-related information. Because of this social media has turned out to be a useful means to study and understand public health. Twitter is one such platform that has proven to be a valuable source of such information for both general public and health officials.

Objective:

In this study, we present a novel dataset consisting of 6,515,470 tweets collected from users self identifying with autism using “#ActuallyAutistic” and a control group. The dataset also has supporting information such as posting dates, follower count, geographical location, and interaction metrics. For better accessibility, reusability and new research insights, we have released the dataset publicly.

Methods:

To create the autism dataset, we specifically targeted English tweets using the search query ‘#ActuallyAutistic’ ranging from January 1, 2014 to December 31, 2022. From these tweets, we identified unique users who had keywords like “autism” OR “autistic” OR “neurodiverse” in their profile description and collected all the tweets from their timeline. To build the control group dataset, we formulated a search query excluding the hashtag i.e. ‘-#ActuallyAutistic’ and collected 1000 tweets per day between the same time period. We illustrate the utility of the dataset through common Natural Language Processing (NLP) applications such as sentiment analysis, tweet and user classification, and topic modeling.

Results:

Our text classifier achieves a noteworthy performance, with 73% accuracy, 0.728 AUC score, and an F1-score of 0.71 using word2vec model with logistic regression and user profile classification achieved 87% accuracy, 0.78 AUC score and an F1 score of 0.805 using attention based Bi-LSTM model. This is a promising start demonstrating the dataset's potential for effective digital phenotyping studies and large scale intervention.

Conclusions:

The textual differences in social media communications can help researchers and clinicians to conduct symptomatology studies, in natural settings, by establishing effective biomarkers to distinguish an autistic individual from their typical peers.