Cross-hospital portability of information extraction of cancer staging information

• We evaluate the portability across hospitals of machine learning-based text mining systems for colorectal cancer staging (TNM and ACPS).
• We present an architecture based on feature selection that allows to build a portable classifier with minimum cost, and we reach state-of-the-art performance.
• The results show that it is feasible to apply an existing TNM classifier to a new hospital without extra training, given that there is a feature normalisation step.

ObjectiveWe address the task of extracting information from free-text pathology reports, focusing on staging information encoded by the TNM (tumour-node-metastases) and ACPS (Australian clinico-pathological stage) systems. Staging information is critical for diagnosing the extent of cancer in a patient and for planning individualised treatment. Extracting such information into more structured form saves time, improves reporting, and underpins the potential for automated decision support.Methods and materialWe investigate the portability of a text mining model constructed from records from one health centre, by applying it directly to the extraction task over a set of records from a different health centre, with different reporting narrative characteristics. Other than a simple normalisation step on features associated with target labels, we apply the models from one system directly to the other.ResultsThe best F-scores for in-hospital experiments are 81%, 85%, and 94% (for staging T, N, and M respectively), while best cross-hospital F-scores reach 84%, 81%, and 91% for the same respective categories.ConclusionsOur performance results compare favourably to the best levels reported in the literature, and—most relevant to our aim here—the cross-corpus results demonstrate the portability of the models we developed.