Zero-Shot TNM Staging from German Pathology Reports Using Pre-Trained Transformer Models (BB-TEN)

TNM staging is essential for cancer registries but is often embedded in unstructured pathology reports, requiring manual extraction. Transformer-based models have recently been proposed for automated TNM classification from English pathology texts.

To assess the robustness and cross-lingual transferability of pre-trained English-language TNM classifiers applied without fine-tuning to German pathology reports.

Three publicly available transformer-based TNM classifiers (T, N, M) were applied to a synthetic German pathology dataset of 109 breast, lung, and prostate cancer reports, using expert-assigned TNM labels as the gold standard.

The models achieved high precision and specificity but moderate sensitivity for tumor and nodal staging and low sensitivity for metastasis, frequently abstaining by predicting "Unknown."

English-trained TNM classifiers can extract staging information from German pathology reports in a zero-shot setting with high reliability when predictions are made, but reduced recall. This conservative behavior supports their use as high-precision screening tools in registry workflows and could be improved through limited domain adaptation.