Robert​ ​Pless,​ ​PhD​1​,​ ​Rebecca​ ​Begtrup,​ ​DO​ ​MPH​1,2​,​ ​Lulwah​ ​Alkulaib​1​,​ ​Samsara​ ​N.​ ​Counts​1​,​ ​ ​James​ ​Harnett​1​, Justine-Louise​ ​Manning​1​,​ ​Hong​ ​Xuan​1​,​ ​David​ ​A.​ ​Broniatowski,​ ​PhD​1 1​George​ ​Washington​ ​University,​ ​Washington,​ ​DC,​ ​USA;​ ​2​Children's​ ​National​ ​Health​ ​System,​ ​Washington,​ ​DC, USA

● “pro-ana”: 16,000 images from a collection of Tumblr blogs including best-thinspo​ ,​ ​thinniest​ ,​ ​and​ ​wanna-be-skinnyminnie​ . ● “selfie”:​ ​4,500​ ​Tumblr​ ​images​ ​tagged​ ​“selfie”. ● “ootd”:​ ​7,000​ ​Tumblr​ ​images​ ​tagged​ ​“ootd”​ ​(outfit​ ​of​ ​the​ ​day). ● “Greek”:​ ​5,000​ ​images​ ​from​ ​Tumblrs​ ​of​ ​Greek-letter​ ​college​ ​organizations. We randomly choose 4740 (15%) of these images to reserve test data, and train the Resnet Deep Learning neural network​2 to classify the remaining training images into these categories. On test data this gives 78% classification accuracy-a significant improvement over chance (25%). To explore a possible application, we identify 10 additional tumblr accounts, five that we judged to have high pro-ana content, 4 blogs without pro-ana content, and 1 fitness inspiration (fitspo) blog that we judged to contain a mix of content. The table below shows the percentage of images​ ​classified​ ​as​ ​pro-ana​ ​in​ ​each​ ​blog: ​ ​ ​ ​ ​Blog​ ​Type​ ​ ​Title,​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​%pro-ana​ ​ ​ ​ ​ ​Blog​ ​type​ ​ ​ ​ ​Title​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​%pro-ana ​ ​ ​ ​ ​pro-ana​ ​ ​ ​ ​think-thygap​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​73​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​not​ ​pro-ana​ ​ ​abelmvada​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​7 ​ ​ ​ ​ ​pro-ana​ ​ ​ ​ ​thinninglittle​ ​ ​ ​ ​ ​ ​ ​ ​ ​88​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​not​ ​pro-ana​ ​ ​roommysocks​ ​ ​ ​ ​ ​ ​ ​ ​ ​21 ​ ​ ​ ​ ​pro-ana​ ​ ​ ​ ​think-skinny-th0ughts​ ​ ​89​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​not​ ​pro-ana​ ​ ​mathematicalmemer​ ​ ​ ​ ​4 ​ ​ ​ ​ ​pro-ana​ ​ ​ ​ ​oh2beskinny​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​83​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​not​ ​pro-ana​ ​ ​satoshikurosaki​ ​ ​ ​ ​ ​10 ​ ​ ​ ​ ​fitspo​ ​ ​ ​ ​ ​veganpilatesangel​ ​ ​ ​ ​ ​ ​53​ ​ ​ ​ ​ ​ ​ ​ ​ ​ ​not​ ​pro-ana​ ​ ​traitspourtraits​ ​ ​ ​ ​ ​9 These​ ​proof​ ​of​ ​concept​ ​results​ ​suggest​ ​that​ ​it​ ​is​ ​feasible​ ​to​ ​automatically​ ​detect​ ​social​ ​media​ ​sources​ ​with​ ​triggering material,​ ​informing​ ​the​ ​creation​ ​of​ ​tools​ ​that​ ​can​ ​assist​ ​clinicians​ ​and​ ​family​ ​members​ ​to​ ​improve​ ​health​ ​outcomes.