Exploring eating disorder topics on twitter: Machine learning approach

Background: Eating disorders (EDs) are a group of mental illnesses that have an adverse effect on both mental and physical health. As social media platforms (eg, Twitter) have become an important data source for public health research, some studies have qualitatively explored the ways in which EDs are discussed on these platforms. Initial results suggest that such research offers a promising method for further understanding this group of diseases. Nevertheless, an efficient computational method is needed to further identify and analyze tweets relevant to EDs on a larger scale. Objective: This study aims to develop and validate a machine learning–based classifier to identify tweets related to EDs and to explore factors (ie, topics) related to EDs using a topic modeling method. Methods: We collected potential ED-relevant tweets using keywords from previous studies and annotated these tweets into different groups (ie, ED relevant vs irrelevant and then promotional information vs laypeople discussion). Several supervised machine learning methods, such as convolutional neural network (CNN), long short-term memory (LSTM), support vector machine, and naïve Bayes, were developed and evaluated using annotated data. We used the classifier with the best performance to identify ED-relevant tweets and applied a topic modeling method—Correlation Explanation (CorEx)—to analyze the content of the identified tweets. To validate these machine learning results, we also collected a cohort of ED-relevant tweets on the basis of manually curated rules. Results: A total of 123,977 tweets were collected during the set period. We randomly annotated 2219 tweets for developing the machine learning classifiers. We developed a CNN-LSTM classifier to identify ED-relevant tweets published by laypeople in 2 steps: first relevant versus irrelevant (F₁ score=0.89) and then promotional versus published by laypeople (F₁ score=0.90). A total of 40,790 ED-relevant tweets were identified using the CNN-LSTM classifier. We also identified another set of tweets (ie, 17,632 ED-relevant and 83,557 ED-irrelevant tweets) posted by laypeople using manually specified rules. Using CorEx on all ED-relevant tweets, the topic model identified 162 topics. Overall, the coherence rate for topic modeling was 77.07% (1264/1640), indicating a high quality of the produced topics. The topics were further reviewed and analyzed by a domain expert. Conclusions: A developed CNN-LSTM classifier could improve the efficiency of identifying ED-relevant tweets compared with the traditional manual-based method. The CorEx topic model was applied on the tweets identified by the machine learning–based classifier and the traditional manual approach separately. Highly overlapping topics were observed between the 2 cohorts of tweets. The produced topics were further reviewed by a domain expert. Some of the topics identified by the potential ED tweets may provide new avenues for understanding this serious set of disorders.