An Attention-Enhanced Explainable Hybrid Deep Learning Framework for Multilingual Opioid Overdose Detection on Reddit

Abstract

Opioid overdose remains a critical global public health challenge, with increasing morbidity and mortality driven by the misuse of prescription and synthetic opioids. Social media platforms provide real‑time signals of opioid‑related behaviors, yet automated detection of overdose‑related content is hindered by informal language, multilingual discourse, and limited availability of fine‑grained annotated datasets. Existing approaches are predominantly monolingual and offer limited interpretability, restricting their applicability in high‑stakes public health settings. This study proposes a multilingual, explainable deep learning framework for opioid overdose detection using English and Spanish social media posts. We construct a novel bilingual dataset annotated with four clinically relevant route‑of‑administration categories through a hybrid LLM‑assisted and expert‑validated process. Two complementary modeling strategies are explored: a translation‑based approach for cross‑lingual standardization and a joint multilingual approach for shared representation learning. A hybrid CNN–BiLSTM architecture with a custom attention mechanism is introduced to capture both local lexical cues and long‑range contextual dependencies, while explainable AI techniques provide interpretable model outputs. Experimental results demonstrate strong performance across all settings, achieving accuracies of 84% on English data, 79% on Spanish data, and 82% on the joint multilingual dataset. These findings highlight the potential of multilingual and interpretable NLP models for real‑time opioid surveillance.