Predicting eating disorders from Internet activity

• Published in: The International journal of eating disorders 2020-09, Vol.53 (9), p.1526-1533
• Main Authors: Sadeh‐Sharvit, Shiri, Fitzsimmons‐Craft, Ellen E., Taylor, C. Barr, Yom‐Tov, Elad
• Format: Article
• Language: English
• Subjects: Accuracy; Adolescent; Adult; browsing history; Eating disorders; Feeding and Eating Disorders - diagnosis; Female; Humans; Internet; Internet activity; Machine learning; Middle Aged; online screening; Risk Factors; Surveys and Questionnaires; Well being; Young Adult
• ISSN: 0276-3478; 1098-108X
• DOI: 10.1002/eat.23338

Objective Eating disorders (EDs) compromise the health and functioning of affected individuals, but it can often take them several years to acknowledge their illness and seek treatment. Early identification of individuals with EDs is a public health priority, and innovative approaches are needed for such identification and ultimate linkage with evidence‐based interventions. This study examined whether Internet activity data can predict ED risk/diagnostic status, potentially informing timely interventions. Method Participants were 936 women who completed a clinically validated online survey for EDs, and 231 of them (24.7%) contributed their Internet browsing history. A machine learning algorithm used key attributes from participants' Internet activity histories to predict their ED status: clinical/subclinical ED, high risk for an ED, or no ED. Results The algorithm reached an accuracy of 52.6% in predicting ED risk/diagnostic status, compared to random decision accuracy of 38.1%, a relative improvement of 38%. The most predictive Internet search history variables were the following: use of keywords related to ED symptoms and websites promoting ED content, participant age, median browsing events per day, and fraction of daily activity at noon. Discussion ED risk or clinical status can be predicted via machine learning with moderate accuracy using Internet activity variables. This model, if replicated in larger samples where it demonstrates stronger predictive value, could identify populations where further assessment is merited. Future iterations could also inform tailored digital interventions, timed to be provided when target online behaviors occur, thereby potentially improving the well‐being of many individuals who may otherwise remain undetected.