Health misinformation in online patient communities poses significant risks for individuals managing chronic conditions, yet no detection benchmarks exist for rare dermatological dis- eases. We present the first unsupervised multi-model frame- work for detecting health misinformation in Hidradenitis Suppurativa (HS) Reddit communities, requiring zero hu- man annotation. Our framework deploys 28 models across five experimental stages: aspect-based sentiment and emo- tion analysis using RoBERTa models, supervised classifi- cation via auto-labeled RoBERTa-family architectures, un- supervised pattern discovery with UMAP/HDBSCAN clus- tering, zero-shot detection via NLI classifiers, and prompt- based detection via open-source LLMs including Llama-3, Mistral-7B, Flan-T5, BioMistral, and GatorTron. We intro- duce the HS Misinformation Index (HSMI), a composite risk metric fusing multi-model consensus, domain-specific key- word heuristics, and contextual sentiment signals. Applied to 9,838 Reddit texts, our pipeline identifies recurring mis- information patterns including cure claims, pseudoscientific narratives, and commercial promotion, along with their emo- tional correlates. Experiments reveal that NLI classifiers ex- hibit more conservative detection behavior than LLMs, while a closed-source validation using GPT-4o, Claude Sonnet 4, and Gemini models on a seven-category misinformation tax- onomy achieves substantially higher inter-model agreement. Emotion-risk analysis shows that texts expressing disgust and anger carry the highest misinformation risk, while joyful texts carry the lowest. General-purpose models consistently out- perform domain-specific ones across all experimental stages. This work establishes the first annotated HS misinformation benchmark and demonstrates that multi-model consensus can reliably surface health misinformation without human labels, providing a scalable template for underserved chronic disease communities.

PDF