A Cascading Hybrid Strategy for Address Normalization using Fine-Tuned LLMs to Strengthen Cancer Registry Geospatial Analysis
REVISTA COLOMBIANA DE TECNOLOGIAS DE AVANZADA · 28.07.2026 · 2 (48) · 166-180
DOI: 10.24054/rcta.v2i48.4512 ↗Kurz gefasst
Fachartikel in der Zeitschrift REVISTA COLOMBIANA DE TECNOLOGIAS DE AVANZADA, online erschienen am 28.07.2026. Verfasst von Silvio Ricardo Timarán Pereira und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC 4.0).
Abstract (Originalsprache)
The accuracy of geospatial analysis within the Population-based Cancer Registry of the Municipality of Pasto (RPCMP) is severely compromised by the heterogeneous and ambiguous nature of physical addresses collected manually. These source data inconsistencies limit the analytical reach of the YACHAY-SIG infrastructure for generating multiscale territorial intelligence. This paper describes the development and implementation of a hybrid methodological framework designed to optimize spatial data integrity through a two-phase processing workflow.The initial phase deploys a deterministic geocoding engine specifically tailored to the local urban morphology. For records exhibiting high semantic complexity such as non-conventional "block and lot" (manzana y lote) nomenclatures a second stage based on Large Language Models (LLMs) is incorporated. This component integrates Natural Language Processing (NLP) techniques, utilizing TF-IDF and Word2Vec vectorization for the discernment and correction of structural anomalies in the addresses.To maximize efficiency in resource-constrained computational environments, a Supervised Fine-Tuning (SFT) protocol was applied using the Low-Rank Adaptation (LoRA) technique on several open-source models. Experimental results reveal that the integration of the Llama-3.2-3B-Instruct model achieved superior performance, reaching a normalization accuracy of 99.40%. System evaluation over an effective test dataset of 2,324 addresses (extracted from a total corpus of 15,492 unique instances allocated into 70% for training, 15% for validation, and 15% for testing) using standard metrics including precision, recall, and F1-score demonstrates a substantial improvement in spatial recovery, increasing from an initial baseline of 33.02% to 55.13%, effectively mitigating the semantic gap regardless of local cartographic constraints. This architecture stands as a scalable data engineering model for epidemiological registries, with future work focusing on integration with spatial providers such as OpenStreetMap and the Agustín Codazzi Geographic Institute (IGAC) to strengthen national geographic coverage.
Zitieren
APA
Pereira, S. R. T., & Córdoba, J. D. V. (2026). A Cascading Hybrid Strategy for Address Normalization using Fine-Tuned LLMs to Strengthen Cancer Registry Geospatial Analysis. REVISTA COLOMBIANA DE TECNOLOGIAS DE AVANZADA, 2(48), 166-180. https://doi.org/10.24054/rcta.v2i48.4512
Vancouver
Pereira SRT, Córdoba JDV. A Cascading Hybrid Strategy for Address Normalization using Fine-Tuned LLMs to Strengthen Cancer Registry Geospatial Analysis. REVISTA COLOMBIANA DE TECNOLOGIAS DE AVANZADA. 2026;2(48):166-180. doi:10.24054/rcta.v2i48.4512
BibTeX
@article{Pereira2026,
author = {Pereira, Silvio Ricardo Timarán and Córdoba, Jonathan David Viveros},
title = {A Cascading Hybrid Strategy for Address Normalization using Fine-Tuned LLMs to Strengthen Cancer Registry Geospatial Analysis},
journal = {REVISTA COLOMBIANA DE TECNOLOGIAS DE AVANZADA},
year = {2026},
volume = {2},
number = {48},
pages = {166--180},
doi = {10.24054/rcta.v2i48.4512}
}
EUVIMED