Using Deep Learning to Improve Eating Disorder Treatment | PDF

Overview

Eating disorders (ED) are serious medical conditions that can have disastrous effects on an individual’s mental and physical health. They are pervasive, and do not discriminate based on race, religion, gender, or socioeconomic status. ED are often a lifelong struggle, with approximately ⅔ of patients never achieving a full and sustained remission. They are the product, in part, of increased societal pressures to fit “the thin ideal,” and exposure to this media can be triggering to people with ED as well as those at risk for developing them. Social media platforms are especially rife with these triggers—individuals with ED have created communities where they support one another in the dangerous pursuit of being “thin enough.” These websites teach readers how to act on and hide their ED, putting them at risk for severe physical and mental health complications, including death. Therefore, triggering content online poses a serious risk to social media users with ED and those at risk for developing them. Similarly, it is essential that clinicians and family members be able to identify websites containing images that are associated with the promotion of ED to prevent accidental or intentional exposure to these triggers. However, it is challenging for caretakers to find and stay up-to-date with ED communities and content online. This research aims to automatically detect such triggering material, with the ultimate goal of designing tools to inform clinicians and support patients in their recovery. The main products of this work are a convolutional neural network that identifies images of ED and two novel software tools built from it that assess websites for ED content. These tools would enable clinicians, family members, and those suffering from ED to understand and identify sources of ED content to improve treatment for ED patients.

Intellectual Merit

While other researchers have used machine learning to automatically detect pro-ED content, no one has made a robust deep learning (DL) classifier that accurately detects images of ED in general. Most importantly, to our knowledge, there are no user-centric tools using DL to improve health outcomes for ED patients. With the first tool, clinicians would be able to access a news-feed depicting the latest trends in pro-ED communities online. Another feature of this tool would explain which visual features distinguish ED images from other similar image categories. This feature would use a class activation mapping, a DL interpretability technique that has never been implemented in this context. The second tool would be for patients in recovery, in the form of a browser extension that filters triggering pages and analyzes the content a user consumes. Finally, related work has not incorporated measures of fairness and human diversity into classification; thus, other classifiers likely include biases towards particular demographic groups underrepresented in the classifier’s training dataset. This project would combat this bias by incorporating fairness guarantees into the classifier to improve the fairness of detection across different demographics.

Broader/Commercial Impact

This project has the potential to improve health outcomes for ED patients in practice as well as streamline the process of developing accurate, unbiased image classifiers. The ED classifier would be the first of its kind, demonstrating that it is possible to automatically detect images of ED with high accuracy. With the first proposed tool, ED clinicians could understand the latest trends in pro-ED online media in order to better treat their patients. With the second, ED patients could have a better recovery by seeing less triggering content without impeding their internet use. Finally, this project would contribute to the body of work using deep learning for recognition of objects in high-level categories, as well as incorporating fairness guarantees into classification accurately.

Samsara Counts & Dominique Dalanni

---