Computational analysis and automated classification of eating disorder discourse on Twitter's EDTWT community

Twitter's "EDTWT" community constitutes a prominent online space for eating disorder (ED) discourse, yet large-scale computational characterization remains limited.

To characterize EDTWT behavioral patterns, emotional dynamics, thematic structure, and develop automated content classification methods through computational analysis of a three-year dataset.

We analyzed 48,341 tweets from 18,587 users (January 2022-February 2025). Analyses included engagement patterns, temporal dynamics, clinical keyword prevalence, multi-method sentiment analysis (TextBlob, VADER, clinical affect lexicons), topic modeling (LDA, NMF), and ensemble-based multi-label classification.

User engagement followed a highly skewed distribution in which a small number of highly active users generated the majority of content (62.1% of users posted only once, while the top 10% produced 48.9% of all content), suggesting that a concentrated subset of users may warrant particular clinical attention. Temporal patterns showed a 7.5-fold difference in posting volume between Friday night peaks and Tuesday morning troughs, with consistent nocturnal peaks between 21:00 and 23:00. Clinical keywords appeared in 26.6% of tweets (body image 14.4%, restrictive eating 6.7%, recovery 4.3%). Sentiment was slightly positive (M=0.060, SD=0.282) with moderate subjectivity (M=0.298). Topic modeling revealed ten themes including calorie tracking (13.2%), Spanish (9.7%) and Polish (8.2%) subcommunities, and recovery discourse (9.4%). An ensemble of automated classifiers trained to categorize eating disorder content achieved strong performance (macro F1=0.753, κ=0.703, where higher values indicate better classification accuracy), outperforming a biomedical-domain language model by 21.4% on the least frequently occurring content categories.

EDTWT exhibits complex heterogeneity with coexisting pro-ED content, recovery discourse, and culturally-specific subcommunities. The concentration of activity among a small number of highly active users enables efficient identification of individuals who may be at elevated risk and in need of targeted support. Automated classification enables scalable content monitoring for digital mental health surveillance and evidence-based platform moderation.