Artificial Intelligence for Opioid Safety Surveillance from Clinical Text: A Clinically Focused Review

Area of Science:

• Medical Informatics
• Patient Safety
• Artificial Intelligence in Healthcare

Background:

• Opioid-related harms, including opioid use disorder and overdose, pose significant patient safety risks.
• Conventional surveillance using structured data often misses critical safety signals documented in unstructured clinical narratives.

Purpose of the Study:

• To review artificial intelligence (AI) methods for identifying opioid-related harms from clinical text.
• To address the ascertainment gap between structured data and narrative clinical documentation.

Main Methods:

• Synthesized 47 empirical studies (2009-2025) applying AI to clinical text for opioid harm detection.
• Examined the evolution of AI methods from rule-based systems to machine learning, deep learning, and transformer-based approaches, including large language models (LLMs).

Main Results:

• Administrative coding systems (e.g., ICD-10) under-ascertain opioid-related events.
• Text-based AI identifies additional cases and contextual details (e.g., mental status, drug causality, naloxone response) often missed by structured data.
• AI methods have progressed in scalability and extraction richness, with recent LLM-based studies focusing on overdose context and prodromal signals.

Conclusions:

• AI, particularly text-based approaches, can significantly improve the detection of opioid-related harms compared to traditional methods.
• Near-term clinical applications should focus on AI-driven decision support for triage, not autonomous diagnosis.
• Future advancements require standardized definitions, equity auditing, robust validation, and prospective outcome evaluation.