# RevealED: Uncovering Pro-Eating Disorder Content on Twitter Using Deep Learning

**Authors:** Jonathan Feldman  
**arXiv:** [2212.13949](https://arxiv.org/abs/2212.13949)  
**Date:** 2022-12-28  
**Summary page:** https://www.summarizepaper.com/en/paper/2212.13949v2/

## Key points

- Alarming increase in adolescents diagnosed with eating disorders during the Covid-19 pandemic
- Proliferation of pro-eating disorder content on social media platforms, particularly Twitter
- Social media inundated with harmful content promoting unhealthy behaviors related to food and body image
- Study aimed to develop a deep learning model to identify posts promoting eating disorders based solely on image data
- Vision Transformer model emerged as the most accurate, achieving an F1 score of 0.877 and 86.7% accuracy on the test set
- Seasonal fluctuations in prevalence of pro-eating disorder content observed, with peaks during summertime
- Steady rise in pro-eating disorder content over the past five years
- Importance of developing tools like deep learning models to identify and mitigate harmful influences online

## Summary

The study "RevealED: Uncovering Pro-Eating Disorder Content on Twitter Using Deep Learning" by Jonathan Feldman delves into the alarming increase in adolescents diagnosed with eating disorders during the Covid-19 pandemic. This surge can be attributed to both the stress of the pandemic and the proliferation of pro-eating disorder content on social media platforms, particularly Twitter. Over the past decade, social media has become inundated with harmful content that promotes unhealthy behaviors related to food and body image. To address this issue, the study aimed to develop a deep learning model capable of identifying posts on social media that promote eating disorders based solely on image data. The researchers collected tweets from hashtags known to promote eating disorders as well as tweets from unrelated hashtags. These images were then labeled as either pro-eating disorder or not based on their associated Twitter hashtag. Several deep learning models were trained on this dataset and evaluated based on metrics such as accuracy, F1 score, precision, and recall. The Vision Transformer model emerged as the most accurate, achieving an impressive F1 score of 0.877 and an accuracy of 86.7% on the test set. Applying this model to unlabeled Twitter image data scraped from "#selfie," the researchers uncovered seasonal fluctuations in the prevalence of pro-eating disorder content, with peaks observed during summertime. These fluctuations not only corresponded to seasonal changes but also reflected external stressors like the Covid-19 pandemic. Furthermore, analysis of Twitter image data revealed a concerning trend - a steady rise in the amount of pro-eating disorder content over the past five years. This upward trajectory suggests that without intervention, this harmful content is likely to continue increasing in prevalence in the future. Overall, this study sheds light on the detrimental impact of pro-eating disorder content on social media platforms and highlights the importance of developing tools like deep learning models to identify and mitigate such harmful influences online.