Introduction: The misuse of over-the-counter (OTC) drugs poses a significant global public health challenge. This study proposes a system for detecting and visualizing inappropriate OTC drug use in social media data. Methods: We constructed a corpus of 20,036 labeled Japanese tweets, including 7,000 medication-related posts, to address the linguistic and cultural nuances. By fine-tuning the Japanese bidirectional encoder representations from transformers models, the system identified misuse patterns such as overuse. The system also incorporates a visualization tool to illustrate temporal and categorical trends, aiding public health authorities in real-time pharmacovigilance efforts. Results: The system demonstrated a strong performance in detecting specific misuse patterns and has the potential to provide insights through the visualization of temporal and categorical trends. Error analysis revealed challenges such as ambiguous terms and noise inherent in social media data. Discussion: The model performed well with sufficient data, but struggled with underrepresented categories. Challenges with ambiguous terms and indirect references emphasize the need for improved contextual understanding and the potential benefits of larger language models or data augmentation techniques. Conclusion: Although this study focused on the Japanese context, the system identified OTC drug misuse patterns and provided information through visualization. This holds promise for real-time pharmacovigilance and can be applied to other languages, contributing to global efforts to monitor and mitigate drug misuse trends.
Nishiyama et al. (2025) studied this question.