Background Extracting structured data from free-text medical records is laborious and error-prone. Traditional rule-based and early neural network methods often struggle with domain complexity and require extensive tuning. Large language models (LLMs) offer a promising solution but must be tailored to nuanced clinical knowledge and complex, multipart entities.

Methods We developed a flexible, end-to-end LLM pipeline to extract diagnoses, per-specimen anatomical-sites, procedures, histology, and detailed immunohistochemistry results from pathology reports. A human-in-the-loop process to create validated reference annotations for a development set of 152 kidney tumor reports guided iterative pipeline refinement. To drive nuanced assessment of performance we developed a comprehensive error ontology— categorizing by clinical significance (major vs. minor), source (LLM, manual annotation, or insufficient instructions), and contextual origin. The finalized pipeline was applied to 3,520 internal reports (of which 2,297 had pre-existing templated data available for cross referencing) and evaluated for adaptability using 53 publicly available breast cancer pathology reports.

Results After six iterations, major LLM errors on the development set decreased to 0.99% (14/1413 entities). We identified 11 key contexts from which complications arose-including medical history integration, entity linking, and specification granularity-which provided valuable insight in understanding our research goals. Using the available templated data as a cross reference, we achieved a macro-averaged F1 score of 0.99 for identifying six kidney tumor subtypes and 0.97 for detecting metastasis. When adapted to the breast dataset, three iterations were required to align with domain-specific instructions, attaining 89% agreement with curated data.

Conclusion This work illustrates that LLM-based extraction pipelines can achieve near expert-level accuracy with carefully constructed instructions and specific aims. Beyond raw performance metrics, the iterative process itself—balancing specificity and clinical relevance—proved essential. This approach offers a transferable blueprint for applying emerging LLM capabilities to other complex clinical information extraction tasks.

Azure compute credits were provided to Dr. Jamieson by Microsoft as part of the Accelerating Foundation Models Research initiative.

This work was supported by the NIH sponsored Kidney Cancer SPORE grant (P50CA196516) and endowment from Jan and Bob Pickens Distinguished Professorship in Medical Science and Brock Fund for Medical Science Chair in Pathology.

I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.

Yes

The details of the IRB/oversight body that provided approval or exemption for the research described are given below:

Ethics committee/IRB of The University of Texas Southwestern Medical Center gave ethical approval for this work

I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.

Yes

I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).

Yes

I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.

Yes

The data used in this study contain patient identifiers and cannot be shared publicly due to privacy regulations and institutional policies. Code for applying this workflow is available at github.com/DavidHein96/prompts_to_table