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

Area of Science:

• Natural Language Processing
• Computational Pathology
• Oncology Informatics

Background:

• TNM staging is crucial for cancer registries but often buried in unstructured pathology reports.
• Manual extraction of TNM staging data is time-consuming and prone to errors.
• Automated TNM classification using transformer models has shown success in English texts.

Purpose of the Study:

• To evaluate the performance of English-trained TNM classifiers on German pathology reports without fine-tuning.
• To assess the cross-lingual transferability and robustness of these AI models.
• To determine the utility of zero-shot TNM classification in a new linguistic context.

Main Methods:

• Three transformer-based TNM classifiers (T, N, M) were utilized.
• A synthetic dataset of 109 German pathology reports (breast, lung, prostate cancer) was created.
• Expert-assigned TNM labels served as the gold standard for comparison.

Main Results:

• Models demonstrated high precision and specificity for tumor (T) and nodal (N) staging.
• Sensitivity for T and N staging was moderate, while metastasis (M) staging sensitivity was low.
• Frequent prediction of "Unknown" indicated a conservative model behavior.

Conclusions:

• English-trained TNM classifiers can reliably extract staging information from German reports in a zero-shot manner.
• The models exhibit high precision but reduced recall, acting as effective screening tools.
• Performance could be enhanced with limited domain adaptation for improved registry workflows.