Rethinking clinical documentation with generative AI
Generative artificial intelligence (AI) is rapidly reshaping the landscape of clinical documentation. Traditionally, writing patient safety narratives (PSNs) for Clinical Study Reports (CSRs) has required hours of manual data extraction and synthesis — a time-consuming process that slows pharmacovigilance workflows.
New advances in large language models (LLMs), such as Google Gemini, are demonstrating how AI can generate coherent, accurate narratives from structured clinical data. By doing so, these models promise to improve both speed and consistency while maintaining compliance with International Council for Harmonization (ICH) standards.
Study overview: Automating PSNs with a RAG framework
In our recent study, we explored how a retrieval-augmented generation (RAG) system could automate PSN drafting for semaglutide-related adverse events. The system merged structured case data with adaptive AI prompting techniques — specifically, Automatic Prompt Engineering (APE) — to optimize the factual accuracy of the generated narratives.
Using an ICH E3–aligned template, the model generated PSNs across four key sections:
- Patient Demographics and Study Information
- Relevant Medical History
- Adverse Event (AE) Details
- Laboratory and Diagnostic Findings
Thirty published case reports were analyzed to assess how well the AI performed in extracting, contextualizing, and summarizing information.
Measuring quality and efficiency
Each AI-generated narrative was evaluated by clinical documentation experts on a 1–10 scale across multiple criteria — including completeness, clarity, and accuracy. Evaluation metrics focused on core demographic details, drug administration data, adverse event description, and diagnostic relevance.
The average processing time per case was approximately 10 seconds, compared to the several hours typically required for manual PSN drafting. This represents a remarkable productivity gain for pharmacovigilance teams.
Key results
The AI-generated narratives achieved an average composite score of 7.5/10 for narrative quality.
- Highest-performing areas included:
- Accuracy of AE/SAE Identification (9.8/10)
- Relevance of Key Findings (9.8/10)
- Disease/Treatment Context Accuracy (9.4/10)
- Extraction of Prior Medications (9.0/10)
These results underscore the model’s strength in synthesizing clinical information into concise, ICH-compliant summaries.
However, the patient demographics section scored lower (6.4–7.0), mainly due to missing temporal details or incomplete demographic data. These gaps reflected the model’s sensitivity to inconsistencies in source reports — a known challenge in real-world data processing.
Discussion: The balance between automation and oversight
Our findings reveal that integrating generative AI within a structured RAG framework can significantly accelerate PSN drafting without compromising clinical accuracy. The approach supports a hybrid workflow in which AI handles repetitive data synthesis, while human reviewers focus on interpretation, validation, and scientific review.
Still, the study also highlights that expert oversight remains essential. Variability across cases — especially when data formats or terminology differ — underscores the importance of human supervision to ensure contextual completeness and regulatory compliance.
The road ahead
Future research will refine prompt design through adaptive APE techniques to improve temporal and contextual accuracy. Expanding the framework across multiple therapeutic areas and languages will be key to scaling adoption in global regulatory environments.
By combining AI-driven generation with expert validation, pharmacovigilance teams can achieve the best of both worlds: faster, more accurate, and more standardized safety documentation.
Key takeaways
AI tools — when integrated with structured RAG systems — hold enormous promise for the future of pharmacovigilance. They can dramatically reduce drafting time, enhance consistency, and allow safety experts to focus where it matters most: interpreting data and protecting patients.