Cancer staging data collection using rules-based natural language processing for entity extraction from pathology...

Area of Science:

• Computational pathology
• Medical informatics
• Natural Language Processing (NLP)

Background:

• Population-level cancer staging data is crucial for treatment planning, outcome prediction, and resource allocation.
• Manual cancer staging is resource-intensive, highlighting the need for efficient automated methods.
• Existing manual methods for cancer data collection are resource-intensive.

Purpose of the Study:

• To develop rules-based NLP systems for extracting explicit and implicit tumour, node, and metastasis (TNM) entities.
• To translate extracted TNM values into cancer stages for melanoma, breast, and colorectal cancers.
• To align cancer staging with the American Joint Committee on Cancer (AJCC) 8th edition TNM staging system.

Main Methods:

• Developed rules-based NLP systems in consultation with domain experts.
• Extracted cancer staging information from pathology reports and hospital inpatient morbidity data.
• Evaluated system performance against manual collections using precision, recall, and F1-scores.

Main Results:

• Rules-based NLP systems achieved 87%-90% accuracy in staging cases compared to manual collections.
• Melanoma NLP system achieved weighted average precision of 0.96, recall of 0.94, and F1-score of 0.94.
• Colorectal and breast cancer NLP models demonstrated high performance with weighted average F1-scores of 0.89 and 0.89, respectively.

Conclusions:

• A rules-based NLP architecture can accurately derive TNM components and cancer stage from clinical text.
• The developed methods support partial automation of cancer staging workflows, especially where training datasets are limited.
• The approach relies on domain-expert input rather than data-driven training.