A Hybrid Address Normalization Framework Using Fine-Tuned LLMs to Enhance Geospatial Intelligence in Cancer Registries

The Population-based Cancer Registry of the Municipality of Pasto (RPCMP) faces significant challenges in spatial data integrity due to the high variability, inconsistency, and ambiguity of manually collected addresses. Such limitations restrict the analytical capacity of the YACHAY-GIS platform to generate precise geospatial intelligence across multiscalar levels. To address these issues, this paper proposes a novel hybrid address normalization framework structured in two sequential stages. First, a specialized deterministic geocoding engine, optimized for local urban morphology, is utilized. Second, for highly non-standard records—characterized by “block and lot” (manzana y lote) nomenclatures—a workflow based on Large Language Models (LLMs) is integrated. To ensure computational efficiency in resource-constrained environments, a Supervised Fine-Tuning (SFT) process was implemented using the Low-Rank Adaptation (LoRA) technique across various open-source LLMs. Experimental results demonstrate that the workflow integrated with the Llama-3.2-3B-Instruct model achieved the highest performance, with a normalization accuracy of 99.40%. The end-to-end evaluation shows an increase in spatial recovery from a 33.02% baseline to 55.13%, effectively isolating the semantic gap from the underlying cartographic limitations of the region. This architecture provides a scalable data engineering model for oncological registries, with future work focused on integrating external spatial providers such as OpenStreetMap and Colombia’s Agustín Codazzi Geographic Institute (IGAC) to further enhance geographic coverage.