Abstract Background Cancer staging data is vital for treatment planning, outcome prediction, clinical research, and healthcare resource allocation. Collection at the population level can improve insights, however existing manual methods are resource intensive. This study aimed to develop rules-based natural language processing (NLP) systems to: (1) extract explicit tumour, node, and metastasis (TNM) entities from data reported to the Western Australian Cancer Registry; (2) extract implicit entities translatable to individual TNM values; and (3) translate these values into cancer stages for melanoma, breast, and colorectal cancers based on the AJCC 8th edition TNM staging system. Methods Rules-based NLP systems were developed with extensive consultation from knowledge experts to extract staging information and stage colorectal, breast, and melanoma cancers using pathology reports and a hospital inpatient morbidity dataset. Their performance was evaluated against manual collections (ground truth) created by cancer staging project officers using recall, precision, and F1-scores. Results After an iterative development process, the rules-based NLP systems correctly staged 87%-90% of cases compared to manual collections created by cancer staging officers. The melanoma NLP system had a weighted average precision of 0.96, recall of 0.94, and F1-score of 0.94. The colorectal and breast models had weighted average precision, recall, and F1-scores of 0.89, 0.89, 0.89; and 0.90, 0.89, and 0.89, respectively. The rules-based NLP systems demonstrated strong performance, with potential for improved accuracy using additional data sources. Conclusion A rules-based NLP architecture can accurately derive TNM components and cancer stage from routinely collected clinical text. Whilst dependent on domain-expert input rather than data-driven training, the methods described demonstrate a way to support partial automation of cancer staging workflows in a setting with limited training datasets.
Newton et al. (Mon,) studied this question.