TCGA_brest_extraction

Abstract

# Structured extraction of breast cancer pathology reports by five large language models (TCGA cohort) ## Description This dataset accompanies the manuscript *"Secure large language models for structured breast cancerpathology extraction: a consensus- and embedding-based evaluation"* (Astruc et al., manuscript inpreparation / [DOI or journal reference to add once available]). It contains the structured extractions produced by five large language models (three commercialAPI-based models — Mistral Large, GPT-4o-mini, Gemini 2.5 Flash — and two open-weight,HDS-deployable models — Qwen-Next 80B, GPT-OSS-120B) for six clinicopathological variables, appliedto 1,034 breast cancer pathology reports from The Cancer Genome Atlas (TCGA). Two files are included: - `5LLM_TCGA_breast_extractions.xlsx` / `.csv` — the structured extraction dataset (same content, two formats), 1,034 rows x 115 columns, including the full source report text.- `Supplementary_Table_S1_EN_v2.docx` — manual review of the 138 reports flagged during automated report-level quality screening (see the manuscript's Methods, "Agent-based deterministic error attribution"). ## Extraction dataset — column dictionary One row = one TCGA pathology report, identified by `report_id`. - `report_text` — the full, unstructured source pathology report text given to each model as input. For each of the six extracted fields — `pT`, `pN`, `Breast_cancer_stage`, `Tumor_size_mm`,`Metastatic_nodes_n`, `Margin_mm` — the dataset gives, per model (`Mistral`, `OpenAI`, `Gemini`,`Qwen`, `GPT-OSS`): - `{field}_{model}_value` — the value extracted by the model (raw model output; numeric fields may include a unit, e.g. "2 mm" or "0.5 cm" — see the manuscript's Methods, "numeric parsing correction", for how these are normalized in the published analysis).- `{field}_{model}_evidence` — the verbatim (or minimally abbreviated) passage from the pathology report cited by the model as supporting evidence for the extracted value.- `{field}_{model}_reasoning` — the model's stated reasoning: the clinical interpretation, staging rule, or logical process used to derive the value from the cited evidence. And, once per field (not per model): - `{field}_majority_vote` — the consensus reference value (majority vote across the 5 models; see Methods, "Consensus framework and agreement analysis").- `{field}_tie_value` — a second value that was also supported by at least two models, if the field is a tie (empty for the large majority of reports).- `{field}_vote_type` — one of `unanimous`, `majority`, or `tie`, describing the type of agreement underlying the majority vote for that report and field. Finally: - `execution_time_{model}_s` — wall-clock execution time (seconds) for that model's extraction request on that report. ## Notes on data quality The per-model `value` fields are raw model outputs and have not been corrected or normalized in thisfile (e.g. some numeric values are recorded with their original unit). The manuscript's publishedconcordance, discordance, and significance statistics apply a corrected numeric parser (strippingcomparison operators, normalizing whitespace, converting cm to mm) on top of these raw values —see Methods for full detail and the associated analysis code.