Automated extraction of information from free text of Spanish oncology pathology reports

[摘要] Background: Pathology reports are stored as unstructured, ungrammatical, fragmented, and abbreviated free text with linguistic variability among pathologists. For this reason, tumor information extraction requires a significant human effort. Recording data in an efficient and high-quality format is essential in implementing and establishing a hospital-based cancer registry. Objective: This study aimed to describe implementing a natural language processing algorithm for oncology pathology reports. Methods: An algorithm was developed to process oncology pathology reports in Spanish to extract 20 medical descriptors. The approach is based on the successive coincidence of regular expressions. Results: The validation was performed with 140 pathological reports. The topography identification was performed manually by humans and the algorithm in all reports. The human identified morphology in 138 reports and by the algorithm in 137. The average fuzzy matching score was 68.3 for Topography and 89.5 for Morphology. Conclusion: A preliminary algorithm validation against human extraction was performed over a small set of reports with satisfactory results. This shows that a regular-expression approach can accurately and precisely extract multiple specimen attributes from free-text Spanish pathology reports. Additionally, we developed a website to facilitate collaborative validation at a larger scale which may be helpful for future research on the subject.

[学科分类] 医学（综合）

[关键词] Artificial Intelligence;Linguistics;Biomarkers;Natural language processing;Pathology reports;deep learning;information extraction;multitask learning;Free-form text