Background: Population-based cancer registries have treatment information for all patients making them an excellent resource for population-level monitoring. However, specific treatment details, such as drug names, are contained in a free-text format that is difficult to process and summarize. We assessed the accuracy and efficiency of a text-mining algorithm to identify systemic treatments for lung cancer from free-text fields in the California Cancer Registry.
Methods: The algorithm used Perl regular expressions in SAS 9.4 to search for treatments in 24,845 free-text records associated with 17,310 patients in California diagnosed with stage IV non-small cell lung cancer between 2012 and 2014. Our algorithm categorized treatments into six groups that align with National Comprehensive Cancer Network guidelines. We compared results to a manual review (gold standard) of the same records.
Results: Percent agreement ranged from 91.1% to 99.4%. Ranges for other measures were 0.71-0.92 (Kappa), 74.3%-97.3% (sensitivity), 92.4%-99.8% (specificity), 60.4%-96.4% (positive predictive value), and 92.9%-99.9% (negative predictive value). The text-mining algorithm used one-sixth of the time required for manual review.
Conclusion: SAS-based text mining of free-text data can accurately detect systemic treatments administered to patients and save considerable time compared to manual review, maximizing the utility of the extant information in population-based cancer registries for comparative effectiveness research.