Presentation (CTI)

OBJECTIVES: Generative AI has shown promise in automating development of code-based health economic models but struggles to reliably generate Excel models. This is due to additional complexities in managing spreadsheet structures, and Excel’s lack of support for error tracing and validation. A potential solution is to separate model logic generation from spreadsheet implementation, enabling LLMs to work on model logic within a coding environment and without consideration of spreadsheet structures. This study evaluated a novel AI-driven pipeline that generates Excel models through a two-stage process: automated construction of model logic in Python followed by translation into a fully functional Excel workbook.
METHODS: The pipeline was evaluated using two previously published health economic models: a partitioned survival model and a Markov cohort model. For each case study, a health economist provided a free-form model concept and set of datapoints. The pipeline automatically generated a Python implementation of the model, before translating the model into a standalone Excel workbook, applying deterministic layout and formatting processes. The generated Excel models were assessed by two health economists for implementation accuracy and usability.
RESULTS: The pipeline automatically generated Python and Excel implementations for both case-study models. Each set of models were generated in less than 60 minutes at a cost of less than $50. Translation of model logic from Python to Excel was performed successfully, with no change in model results. Manual review demonstrated highly accurate Excel implementation, with no substantive errors identified and close replication of published results. Layout and formatting, generated through deterministic processes rather than LLM outputs, were highly consistent across the models.
CONCLUSIONS: Our two-stage approach may offer a robust method for automating development of Excel-based health economic models using generative AI. This has the potential to significantly reduce the resource required to develop Excel models and to reduce implementation errors.