🟠 Moderate Evidence
Researchers have developed a large language model capable of identifying adverse drug events from unstructured clinical notes in electronic health records (EHRs), according to a study published in Drug Safety (2026). The machine learning approach could improve post-marketing drug surveillance by automating detection of safety signals that might otherwise be missed in vast clinical datasets.
Key takeaways
- A large language model demonstrated high accuracy in detecting adverse drug events from clinical notes in electronic health records
- Automated extraction could enhance pharmacovigilance by identifying safety signals earlier and more efficiently than manual review
- The approach may address critical gaps in post-marketing drug safety surveillance across healthcare systems
Study at a Glance
| Source | Drug Safety |
| Study type | Model development and validation study |
| Application | Post-marketing adverse event detection from clinical documentation |
| Technology | Large language model trained on clinical narratives |
| Focus | Extraction of adverse drug events from unstructured EHR data |
Automated Adverse Event Detection Workflow
Application of large language models to clinical note analysis for pharmacovigilance
Source: Drug Safety Journal, 2026 | Georgian Medical Journal News
Automating Safety Signal Detection
Current pharmacovigilance systems rely heavily on manual review of adverse event reports, a labour-intensive process that may miss safety signals buried within unstructured clinical documentation. The study published in Drug Safety demonstrates that a large language model can systematically extract adverse drug event information from clinical notes with performance metrics comparable to human expert review. This automation could substantially increase the scale and speed of post-marketing safety surveillance.
Manual identification of adverse events from narrative clinical notes remains a significant bottleneck in drug safety systems. Healthcare providers document patient experiences in free-text formats that are difficult to process systematically at scale. The research team developed a model capable of parsing these narratives to identify and categorise adverse events, potentially enabling real-time monitoring across entire hospital networks or health systems. See related Pharmacy & Prescribing updates for additional safety considerations.
Validation and Clinical Integration
The model’s performance was validated against gold-standard adverse event classifications established by clinical pharmacists and patient safety experts. The Drug Safety publication outlines how the system achieved sufficient accuracy to support integration into routine EHR workflows. Healthcare institutions adopting such tools could generate comprehensive adverse event datasets without proportional increases in manual labour or cost.
Integration with existing EHR infrastructure presents both opportunities and implementation challenges. The model requires connection to clinical databases while maintaining patient privacy and data security standards. Early implementation efforts suggest that healthcare systems can deploy the technology with existing technical infrastructure, though standardisation across different EHR platforms remains an ongoing concern. Learn more about Quality & Safety standards in healthcare technology adoption.
Implications for Global Drug Safety Monitoring
Accurate, timely identification of adverse drug events is essential for regulatory agencies including the US Food and Drug Administration (FDA) and the European Medicines Agency (EMA), which monitor drug safety after approval. The automated extraction capability described in the Drug Safety study could accelerate signal detection, potentially enabling faster regulatory action when safety concerns emerge. This is particularly important for rare or delayed adverse events that might not accumulate sufficient reports through traditional spontaneous reporting channels.
The technology also addresses global equity concerns in pharmacovigilance. Many healthcare systems in low- and middle-income countries lack the human resources for comprehensive post-marketing surveillance, placing populations at heightened risk from undetected safety signals. Automated systems could democratise access to advanced safety monitoring capabilities across resource-constrained settings, though implementation would require investment in EHR infrastructure and regulatory capacity building.
A large language model successfully identified adverse drug events from clinical notes in electronic health records, demonstrating that artificial intelligence can enhance post-marketing drug safety surveillance at scale.
— Study published in Drug Safety (2026)
What this means
Frequently asked questions
How does this AI model differ from existing adverse event reporting systems?
Traditional systems rely on voluntary reports submitted by healthcare providers or patients. This model automatically extracts adverse event information directly from clinical documentation already present in EHRs, capturing events that may never be formally reported. The Drug Safety study shows that automated extraction identifies substantially more events than passive reporting systems alone.
Does automated detection replace human pharmacists in drug safety?
No. The model is designed to augment, not replace, human expertise. Clinical pharmacists still interpret findings, assess causality, and determine regulatory significance. The technology reduces time spent on initial data screening, allowing specialists to focus on complex safety assessments and signal evaluation—higher-value activities that require clinical judgment.
What about patient privacy when extracting data from clinical notes?
Implementation requires robust de-identification and data governance frameworks. Healthcare institutions must apply established HIPAA and equivalent regulatory standards before processing notes. The study acknowledges these requirements, though specific technical de-identification methods vary by institution and jurisdiction.
As electronic health records become increasingly central to healthcare delivery globally, the application of natural language processing and machine learning to pharmacovigilance represents a logical evolution in drug safety practice. The successful validation of this language model suggests that healthcare systems and regulatory agencies should prioritise integration of such tools into post-marketing surveillance infrastructure, while ensuring appropriate governance, validation, and clinician oversight.
Was this article helpful?
Disclaimer. This article is health journalism intended for general information and education. It is not medical advice and is not a substitute for professional diagnosis or treatment. Always consult a qualified healthcare provider about your individual circumstances. Full disclaimer →
Related Coverage
Medically reviewed by Prof. Giorgi Pkhakadze, MD, MPH, PhD. Spotted an error? Contact the editorial team.
