Presentation (CTI)

OBJECTIVES: To develop and assess a modular inference framework capable of generating preliminary clinical drug reports by integrating outputs from multiple prompt-specific language models.
METHODS: The framework was compatible with a wide range of transformer-based models, including domain-adopted and instruction-tuned variants, and requires no task-specific retraining. Nine independent prompts were used to elicit key clinical report components (e.g., FDA-approved indications, efficacy descriptions, dosing tables, and adverse reaction profiles). Each section was generated separately, automatically merged, and formatted into a cohesive, human-readable report. The pipeline prioritized reproducibility, modularity, and standardized section structures to enable large-scale batch processing. A qualitative assessment examined formatting consistency, clinical usability, and compatibility with downstream workflows, including markdown-to-PDF conversion.
RESULTS: The framework produced consistently structured drug reports across diverse underlying models, demonstrating stable formatting, section-level integrity, and clinically interpretable output. Automated merging of independently generated sections yielded cohesive documents suitable for further review.
CONCLUSIONS: The framework supported batch inference across heterogeneous drug datasets. Its emphasis on modularity, reproducibility, and standardized outputs positions it for integration into drug evaluation and health services delivery workflows, with the potential to enhance efficiency, reduce manual effort, and ensure consistent information synthesis across large drug repositories.