University of Minnesota
Using Persuasive Strategies to Detect Health and Fitness Misinformation
Abstract
dc:description.abstractHealth misinformation has become increasingly prevalent across social media and online platforms, especially within the fitness and wellness communities. Such content often relies on persuasive writing techniques such as emotional appeals, confident tone, and pseudo scientific claims to make false information sound credible. Traditional AI fact checking systems mainly focus on factual accuracy and often overlook these linguistic and psychological aspects of persuasion. This study aims to integrate persuasive language analysis into misinformation detection, with the goal of understanding how emotional and rhetorical cues influence readers and developing AI models that can identify not only false information but also the persuasive techniques that make it convincing. Two major datasets, PubHealthTab and FakeHealth, were used. Texts were preprocessed, tokenized, and transformed into numerical representations using TF-IDF, Word2Vec, and DistilBERT embeddings. Baseline models were compared across three configurations: TF-IDF with Logistic Regression, Word2Vec with a Support Vector Machine, and a fine-tuned DistilBERT transformer, achieving accuracies of 58.5%, 37.6%, and 65.3% respectively, though the highest-accuracy model failed entirely on the minority REFUTES class, underscoring the limits of accuracy alone as a metric on imbalanced data. To more directly test whether persuasive strategy predicts misinformation, a subset of 149 health-relevant claims was manually annotated according to the classical rhetorical framework of ethos, pathos, and logos. A model using only this persuasion-mode label, with no text features at all, achieved 65.1 percent accuracy, nearly matching the 67.1 percent accuracy of a text-only model built from 500 word-level features. Claims combining multiple persuasive appeals were refuted nearly twice as often as purely logic-based claims. Explainable AI tools, SHAP and LIME, were applied to confirm this pattern and to visualize which specific words most influenced individual model predictions, such as comparative and alarmist phrasing rather than topic-specific vocabulary. Overall, the findings highlight that understanding persuasion is crucial to combating health misinformation, and that combining explainable AI with rhetorical analysis supports the development of transparent and trustworthy health information systems.
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Kamal Kumar, Samuel Jacinth
Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Dc Identifier Citation
- Kamal Kumar, Samuel Jacinth. (2026). Using Persuasive Strategies to Detect Health and Fitness Misinformation. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/283013.