Presentation (CTI)

OBJECTIVES: Health economic models (HEM) development in rare diseases is often challenging due to multiple factors such as limited patient populations, heterogeneous clinical manifestations, and scarce data on disease progression and outcomes. These issues can complicate model conceptualization and lead to high uncertainty in cost-effectiveness analyses. This proof-of-concept study demonstrates how a proprietary generative AI (Gen-AI) tool—augmented with human feedback, guidelines and domain-specific knowledge—can address these challenges by conceptualizing a HEM for a rare disease with significant data gaps.
METHODS: A proof-of-concept exercise was undertaken to develop a HEM for Transthyretin amyloid polyneuropathy (ATTR-PN), a rare disease characterized by the accumulation of amyloid protein in nerve tissues. The foundational large language model (LLM) was augmented with several core HEM guidelines to train it, which were chunked, embedded and retrieved via retrieval augmented generation (RAG). To incorporate further contextual awareness, the relationships between training documents were transformed into a knowledge graph via Graph RAG, where nodes represented critical case information and edges depicted relationships among these data points. Hypothetical Document Embedding (HyDE) was employed to enrich embeddings and enhance the AI-driven generation of model concepts.
RESULTS: The LLM’s initial recommendation for natural history was further refined using the human-in-loop approach. Initial recommendation was promising and closely aligned with what human experts might have generated through comprehensive manual conceptualization. For the model structure, LLM suggested a semi-Markov model suitable for capturing the complex, non-linear progression of ATTR-PN. It suggested the following health states: Asymptomatic Carrier, Mild Polyneuropathy (PND I), Moderate Polyneuropathy (PND II), Severe Polyneuropathy (PND IIIA and IIIB), End-Stage Polyneuropathy (PND IV).
CONCLUSIONS: This proof-of-concept exercise illustrates the potential of LLMs to expedite and enhance HEM conceptualization in rare diseases, despite challenges with the inherent data limitations and diverse clinical pathways. Incorporating human feedback is crucial to ensure clinical relevance and methodological rigor.