Adverse Drug Event (ADE) detection from user-generated content on social media offers critical real-world pharmacovigilance insights, especially when performed across multiple languages. This paper presents a zero-shot, prompt-based approach using GPT-4o for the SMM4H-HeaRD 2025 Shared Task 1: binary classification of ADE presence in posts written in English, German, French, and Russian. While our English-language approach is detailed in a separate submission titled Uncovering Underreported Adverse Drug Reactions in Epilepsy Communities via Interpretable Social Media Mining, this work focuses on the remaining three languages. Without using any model training, we collaboratively refined a multilingual prompt to classify batches of posts. We observed that GPT-4o achieves high recall across languages but tends to over-predict ADEs, resulting in lower precision. A rerun of the model on only predicted positives in smaller batches significantly improved F1 scores. This study highlights GPT-4o's utility for zero-shot multilingual ADE detection and suggests prompt complexity and batch size as critical parameters in zero-shot inference performance.

PDF