A language model restaged 51,242 cancer patients from old radiology reports

Cancer registries are a mess of history. A database that spans two decades will contain patients staged under several successive editions of the tumour-node-metastasis (TNM) system, the standard that classifies how far a cancer has spread. Compare survival across those cohorts and you are partly measuring changes in the ruler, not the disease. A study published on 4 September in npj Precision Oncology describes an attempt to fix that retroactively, using a large language model [s1].

What the researchers built

The team assembled a pipeline that combines formal TNM staging criteria and anatomical knowledge with a large language model and a second "reflection" step — a stage where the model rechecks its own output — to read narrative radiology reports and convert them into a single, unified staging standard, the AJCC/UICC 9th edition [s1]. The target disease was nasopharyngeal carcinoma, a cancer for which imaging-based staging is central [s1].

To test it, they used an expert-annotated cohort of 340 patients and compared several language models [s1]. The best configuration, built on a model the authors call GLM4.5, reached accuracies of 0.95 for tumour (T) classification, 0.86 for nodal (N) classification and 0.99 for metastasis (M) classification against the expert reference [s1]. Nodal staging — often the hardest to infer from free text — was the weakest of the three, which is worth noting rather than glossing over.

Turning it loose on the archive

With that performance established, the framework was applied to a real-world cohort of 51,242 nasopharyngeal carcinoma patients treated between 2010 and 2025 [s1]. A stratified random subset of 200 cases, re-checked against the expert standard, showed high concordance [s1]. In downstream analyses, the AI's unified staging was associated with a higher concordance index — a measure of how well a staging system sorts patients by prognosis — and produced more distinct separation between survival curves than the staging already recorded in the electronic health record [s1].

How much to make of it

The claim here is narrow and, for once, appropriately so. This is a tool for harmonising historical data so that longitudinal analyses are not distorted by shifting staging systems — not a system for staging a patient in front of a clinician, and not a treatment tool. Its accuracy was measured against expert annotations in a single cancer type at, it appears, a setting with the radiology reports to feed it; performance could differ for other cancers, other reporting styles, or messier real-world text. The 200-case validation of the full cohort is reassuring but small relative to 51,242 patients, and "high concordance" is the authors' summary rather than a number a reader can weigh here.

The finding that unified staging separated survival curves better than the recorded staging is the most interesting result, because it suggests the cleanup is not merely cosmetic — inconsistent historical staging may have been blunting the prognostic signal in the original records. Whether that holds in other datasets is the obvious next question.

Why it matters

Much of the excitement around language models in medicine has focused on talking to patients or drafting notes. This is a quieter use: cleaning the retrospective data that oncology research runs on. If it generalises, it could make large legacy cohorts more usable without re-reading every scan by hand — but generalisation across cancers and institutions is exactly what a single-disease, single-cohort study cannot demonstrate.

This article is informational and does not constitute medical advice.