Presentation (CTI)

OBJECTIVES: Healthcare decision-making increasingly relies on large and complex datasets, making traditional analytical approaches time-consuming and resource-intensive. This study evaluated whether a generative AI-based analytical framework could support the development of a statistical analysis plan (SAP) by generating SAP table shells and accurately populating them with study data, while producing analytically rigorous, reproducible, and regulatory-aligned outputs
METHODS: A generative AI-enabled analytical framework incorporating a large language model (LLM) was implemented to support structured analysis and the development of an SAP. Simulated individual patient-level clinical data (~10,000 records) were uploaded to the tool in a CSV format and pre-processed prior to analysis. SAP table shells were generated in accordance with predefined specifications, and plain-language user queries were translated into executable statistical code to produce descriptive analyses and populate SAP-aligned tables and visualisations. All outputs were independently reviewed by a human subject-matter expert (SME) to assess analytical accuracy, completeness, and reproducibility
RESULTS: The AI-based framework successfully generated 20 SAP table shells and produced complete descriptive statistics for approximately 20 predefined baseline variables across all table shells. SME validation confirmed the accuracy of all numerical outputs and statistical computations. Manual human refinement was required in approximately 3-4% of cases, where some tables and figures needed to be refined to align with SAP presentation standards. These refinements were addressed through iterative review, resulting in finalised SAP-ready outputs. Compared with conventional manual workflows, the AI-enabled approach substantially reduced analytical development and reporting timelines by almost 85%
CONCLUSIONS: The generative AI-based analytical framework demonstrated the ability to support SAP development by generating and populating SAP table shells with high analytical accuracy under continuous human oversight. These findings highlight the potential of AI-enabled analytics to enhance efficiency and reduce analytical timelines in healthcare research while maintaining methodological rigor