This paper proposes analyzing unstructured data from social media platforms like Reddit as a solution to the opioid overdose crisis, a serious public health problem in the United States. Drawing on Reddit user data sharing their experiences with opioid use, we extract information using a natural language processing (NLP) technique leveraging Opioid Named Entity Recognition (ONER-2025). We build a unique, manually annotated dataset of 331,285 tokens and detail the annotation process and challenges associated with it, encompassing eight key opioid entity categories. Furthermore, we analyze linguistic challenges in opioid-related discussions, such as slang, ambiguity, fragmented sentences, and emotionally charged language. We propose a real-time monitoring system that integrates machine learning, deep learning, Transformer-based language models, and advanced contextual embeddings. In 11 experiments conducted with 5-fold cross-validation, Transformer-based models such as bert-base-NER and roberta-base achieved 97% accuracy and F1-score, which is 10.23% better performance than the baseline model (RF=0.88).
