P235: GenAI-assisted Clinical Study Report Writing with Structured Content Authoring: Aligning Model Strengths to Section Needs
Poster Presenter
SAMIR LEDADE
ASSOCIATE DIRECTOR
Alexion Pharmaceuticals Canada
Objectives
To outline how we added Generative AI to Structured Content Authoring (SCA) via development of a standardized process and scorecard to improve efficiency, quality, and regulatory robustness of AI generated outputs.
Method
InteliNotion was configured as an enterprise SCA platform for Protocols, SAPs, and CSRs. GenAI was integrated to generate protocol/SAP based and data driven CSR sections. Multiple LLMs were evaluated per section using a standardized scorecard with human-in-the loop review.
Results
Evaluation of GenAI-assisted authoring on a SCA platform demonstrated measurable efficiency, quality, and content governance across regulatory medical writing workflows. Using a standardized five-point, ten-parameter scorecard, multiple LLMs (Claude 4.0/4.5, GPT4.1/5.0, and o3) were assessed across protocol-based and data driven CSR sections. Model behavior was strongly use-case dependent: o3 produced most concise, data focused summaries; GPT4.1 generated fluent, detailed scientific narratives; Claude 4.0 and 4.5 provided more contextual and interpretive language; and GPT5.0 yielded the most granular outputs, often with unnecessary detail requiring refinement.
The scorecard assessed each output with scientific accuracy, completeness, readability, structure, internal-standard alignment, and ethics/compliance; cross-functional human-in-the-loop review ensured only prompts/models aligned with the scorecard criteria and regulatory expectations were adopted.
This multi-model approach enabled alignment of specific CSR sections (protocol-based vs data driven content) with the most appropriate LLM, improving output quality and reducing post-generation editing effort. Integration of GenAI within the SCA platform allowed metadata-driven document assembly and controlled reuse of approved content, strengthening consistency and traceability across Protocols, SAPs, and CSRs. Propagating updated approved template language into an in-progress protocol avoided an amendment (~$1M cost saving).
Other benefits included faster initial document assembly and improved visibility into how content was sourced, reused, linked across documents, and updated over time. Key challenges included change management, SME availability, evolving AI governance requirements, and variability in LLM behavior; however, structured evaluation, cross-functional collaboration, and human-in-the-loop review mitigated risk and supported compliant adoption of GenAI within regulated medical writing workflows.
Conclusion
GenAI assisted CSR generation, when implemented on a governed SCA foundation, demonstrates a scalable approach to improving the efficiency, quality, and consistency of regulatory medical writing documents while preserving human oversight and GxP expectations. Our approach provides an essential governance layer through centrally managed componentized content, while GenAI enables scalable text generation and controlled language transformation within protocol/SAP based and data driven sections.
A standardized scorecard to evaluate multiple LLM models and prompts was critical to success. Assessing outputs across different parameters ensured that only prompts and models meeting predefined thresholds were adopted. Importantly, no single LLM was optimal across all CSR sections. Aligning model strengths to section specific needs (eg, concise, datacentric models for summary sections and more narrative models for interpretive content) resulted in higher quality, more reliable outputs within a controlled workflow.
Our approach also underscores the importance of robust AI governance in a highly regulated environment. Dedicated security, privacy, AI governance maturity, and validation activities enabled responsible adoption of GenAI at-scale. Cross-functional collaboration with different functions was essential to designing workflows that balanced efficiency with compliance.
Looking ahead, our GenAI enabled SCA model is being extended beyond CSRs to Module 2 summaries and the Table of Studies. Planned enhancements such as expanded prompt libraries, learning from human-in-the loop edits to improve first time right generation, and enterprise knowledge workflows using Agentic frameworks and GraphRAG aim to further enhance accuracy, contextualization, and reviewer focus. Overall, GenAI assisted authoring on an SCA foundation represents a sustainable and adaptable operating model with the governance required to transform medical writing in rare disease and beyond.
