Optimizing an automated chatbot to achieve efficient, scalable treatment for eating disorders

PROJECT SUMMARY/ABSTRACT Eating disorders (EDs) are common, disabling, and costly, yet less than 20% of those with EDs ever receive treatment. As such, our team established a coached app for EDs. Though coached apps are lower cost relative to therapy, they come with significant implementation barriers given the need for human support. One solution is to program chatbots that mimic aspects of human coaching. Still, it is recognized that engagement with digital mental health interventions can be challenging, making it important that an ED treatment chatbot be efficient, ideally with the most effective components delivered first. To date though, little is known about which components are most effective in reducing ED symptoms. The goal of this study is to fill this gap by optimizing an automated chatbot program to treat EDs. We will build on our programmatic research in this area and prior use of the multiphase optimization strategy (MOST) to test four candidate ED treatment components in an optimization randomized control trial (ORCT). These candidate components, which map onto ED putative mechanisms, target: 1) overevaluation of weight/shape; 2) dietary restraint; 3) emotion dysregulation; and 4) resisting urges to binge. Our multidisciplinary, multisector team has experience developing, evaluating, and deploying digital ED interventions. We include industry and non-profit partners, including the National EDs Association (NEDA; largest EDs non-profit in the U.S.), assuring real, sustained impact. For our ORCT, we will enroll N=800 adults who screen positive for an ED and are not in treatment, recruited through NEDA’s online EDs screen, which reaches 200K+ per year, or social media. We will randomly assign participants in a carefully-balanced 24 factorial design to receive some combination of the candidate ED treatment components (i.e., each combination from zero components to four components). In Aim 1, will estimate the individual and combined contributions of the four candidate ED treatment component on changes in ED psychopathology, as well as on ED behavior frequencies, comorbid symptoms, and clinical impairment at 1-, 2-, and 6-months post- randomization. We will also examine moderation of component effects. In Aim 2, we will examine the effects of each candidate component on its proposed target, and determine if those changes are associated with reductions in overall ED psychopathology. Finally, in Aim 3, we will identify the optimized chatbot package that is effective, producing the best expected improvement in ED outcomes, while remaining efficient in the use of participant time. We will also explore component ordering that may best accommodate participants with different engagement levels. To inform decision-making about immediate dissemination of the optimized chatbot, we will compare the optimized chatbot to the control version. We will also identify any differences in the optimized chatbot as a function of participant priorities and characteristics (personalization). This study has high potential to offer streamlined EDs care via a chatbot that can be scaled infinitely, along with real impact nearly immediately, rather than the long translation gap widely discussed.