EDanalysis -- 2019 GW CS Senior Design Project

7 object(s)
 

Writing2

Elevator Pitch | PDF

According to the National Association of Anorexia Nervosa and Associated Disorders, at least thirty million people in the United States suffer from an eating disorder (ED). ED are complex medical conditions with the highest mortality rate of any mental disorder. While ED are generally associated with white women, they affect people of all genders, races, and walks of life. Treatment of ED requires long-term rehabilitation, a strong support system, and sustained behavioral changes, similar to drug dependency rehabilitation. One major setback to recovery from an ED is exposure to triggering images or content that can bring back disordered thoughts and behaviors. This content poses a serious threat to patients, yet can be found anywhere, particularly online and on social media–yet there exists no way to avoid it without avoiding certain platforms altogether, perhaps even the internet, which a substantial handicap to that person’s daily life.

On the other side, it is challenging for clinicians to find such material without a familiarity with the host platform, not to mention understanding trends in posts. This, too, is a serious obstacle, because it is critical for medical professionals who treat patients ED to stay informed on what risks exist outside of treatment. Similarly, family members and caretakers of people with ED would benefit from stay up-to-date with trends and challenges their loved ones might encounter. We propose two software solutions that assess websites for ED content with distinctly different purposes: one for diagnostics, the other for filtering out triggering material. The first is intended for clinicians and caretakers, and the second is for patients. Both tools would enable users to understand and identify sources of ED content with the goal of improving treatment and recovery outcomes for ED patients. Currently, there exists no tool to filter this content available on the market nor software designed for ED clinicians or patients. Therefore, our tools would be the first of their kind, in addition to solving a challenging problem at the forefront of deep learning technology.

Behind the scenes, our tools detect and process triggering images by using a deep learning classifier–a convolutional neural network trained for this specific task. To create an optimal classifier, we combine concepts in fairness in machine learning and computer vision to maximize the performance of our classifier in a challenging problem domain. We optimize for fairness as well as performance in the creation of our classifier. To do this, we programmatically incorporate diverse data in the training set to obtain a dataset that is more representative of the demographic range of triggering content. We also attempt to solve a common problem in the domain of deep learning: unbalanced datasets. Unbalanced can mean different numbers of examples in different categories, but can also mean that within a category there are features that are under-represented. In the context of diagnostic imagery, ensuring that classifiers work on unusual cases is important. With our product, we demonstrate a novel approach to improve classifier performance within the context of accurately classifying images of eating disorders. Hence, our project has both the life-altering potential to improve health outcomes for ED patients in practice and streamline the process of developing accurate, unbiased deep learning classifiers.

Samsara Counts & Dominique Dalanni