Automatic Parsing of Colonoscopy Medical Reports With Llms

Automatic parsing of colonoscopy reports is essential to transform unstructured narratives into structured data that supports systematic monitoring of quality indicators in gastrointestinal practice. This work presents a modular pipeline based on Large Language Models (LLMs) to automatically extract key colonoscopy quality metrics, namely progression, preparation, limitations, alterations, and number of polyps, from routine medical reports. A dataset consisting of 22,577 colonoscopy reports from a Portuguese hospital was used, with a subset of 51 reports annotated by a gastroenterology specialist serving as the gold standard for model evaluation. Six LLMs were first benchmarked under identical conditions, and the best-performing models were integrated into a modular architecture that assigns specialized models to each category. The final pipeline achieves accuracies above 84% for all categories, reaching 98.0% for alterations, 92.2% for several other fields, and correctly detecting cecal intubation in 96.7% and complete colonoscopies in 93.9% of cases, including reliable identification of both presence and absence of polyps. From our tests the pipeline was able to extract this information from 20 reports in a total time of 307.7 seconds. By exporting results in JSON and Excel formats, the system eases integration into existing clinical workflows and quality dashboards, enabling scalable, data-driven assessment of colonoscopy performance and offering a flexible architecture that can be adapted to other endoscopic or narrative-based medical reporting scenarios.