Abstract
BACKGROUND: Reddit is a popular social media platform that has faced scrutiny for inflammatory language against those with obesity, yet there has been no comprehensive analysis of its obesity-related content. OBJECTIVE: We aimed to quantify the presence of 4 types of obesity-related content on Reddit (misinformation, facts, stigma, and positivity) and identify psycholinguistic features that may be enriched within each one. METHODS: All sentences (N=764,179) containing "obese" or "obesity" from top-level comments (n=689,447) made on non-age-restricted subreddits (ie, smaller communities within Reddit) between 2011 and 2019 that contained one of a series of keywords were evaluated. Four types of common natural language processing features were extracted: bigram term frequency-inverse document frequency, word embeddings derived from Bidirectional Encoder Representations from Transformers, sentiment from the Valence Aware Dictionary for Sentiment Reasoning, and psycholinguistic features from the Linguistic Inquiry and Word Count Program. These features were used to train an Extreme Gradient Boosting machine learning classifier to label each sentence as 1 of the 4 content categories or other. Two-part hurdle models for semicontinuous data (which use logistic regression to assess the odds of a 0 result and linear regression for continuous data) were used to evaluate whether select psycholinguistic features presented differently in misinformation (compared with facts) or stigma (compared with positivity). RESULTS: After removing ambiguous sentences, 0.47% (3610/764,179) of the sentences were labeled as misinformation, 1.88% (14,366/764,179) were labeled as stigma, 1.94% (14,799/764,179) were labeled as positivity, and 8.93% (68,276/764,179) were labeled as facts. Each category had markers that distinguished it from other categories within the data as well as an external corpus. For example, misinformation had a higher average percent of negations (β=3.71, 95% CI 3.53-3.90; P<.001) but a lower average number of words >6 letters (β=-1.47, 95% CI -1.85 to -1.10; P<.001) relative to facts. Stigma had a higher proportion of swear words (β=1.83, 95% CI 1.62-2.04; P<.001) but a lower proportion of first-person singular pronouns (β=-5.30, 95% CI -5.44 to -5.16; P<.001) relative to positivity. CONCLUSIONS: There are distinct psycholinguistic properties between types of obesity-related content on Reddit that can be leveraged to rapidly identify deleterious content with minimal human intervention and provide insights into how the Reddit population perceives patients with obesity. Future work should assess whether these properties are shared across languages and other social media platforms.