Social media platforms have become critical sources of patient-generated health data, yet existing computational approaches fail to capture the interconnected nature of online health discourse. We present a novel framework that integrates graph-based community detection with large language model analysis to understand patient narratives in multimodal social media content. Applied to 10,253 TikTok posts about JAK inhibitors (2020-2024), our approach constructs heterogeneous graphs representing user-content-medical entity relationships and applies community detection algorithms enhanced with context-aware LLM interpretation. Our comprehensive analysis of 10,253 posts (January 2020–September 2024) reveals five distinct patient communities characterized by different discourse patterns: treatment success narratives (873 nodes), medication guidance (642 nodes), side effect discussions (589 nodes), comparative treatment analysis (412 nodes), and dosage optimization (347 nodes). The Louvain algorithm significantly outperformed Girvan-Newman in modularity (0.9931 vs. 0.9928), conductance (0.0002 vs. 0.0006), and computational efficiency (0.14s vs. 54.24s). Temporal analysis demonstrates increasing community cohesion and evolving discourse patterns from cautious inquiry (2020-2021) to experience sharing and specialized sub-communities (2023-2024). This work contributes: (1) a scalable computational framework for multimodal health content analysis, (2) methodological innovations in graph-LLM integration, and (3) insights into platform-specific health communication patterns. The framework has applications in pharmacovigilance, computational social science, and AI-assisted health monitoring systems.

This study evaluates the resilience of large language models (LLMs) against adversarial attacks, specifically focusing on Flan-T5, BERT, and RoBERTa-Base. Using systematically designed adversarial tests through TextFooler and BERTAttack, we found significant variations in model robustness. RoBERTa-Base and Flan-T5 demonstrated remarkable resilience, maintaining accuracy even when subjected to sophisticated attacks, with attack success rates of 0%. In contrast, BERT-Base showed considerable vulnerability, with TextFooler achieving a 93.75% success rate in reducing model accuracy from 48% to just 3%. Our research reveals that while certain LLMs have developed effective defensive mechanisms, these safeguards often require substantial computational resources. This study contributes to the understanding of LLM security by identifying existing strengths and weaknesses in current safeguarding approaches and proposes practical recommendations for developing more efficient and effective defensive strategies