Automated Error Labeling in Radiation Oncology via Statistical Natural Language Processing

A report published in 2000 from the Institute of Medicine revealed that medical errors were a leading cause of patient deaths, and urged the development of error detection and reporting systems. The field of radiation oncology is particularly vulnerable to these errors due to its highly complex process workflow, the large number of interactions among various systems, devices, and medical personnel, as well as the extensive preparation and treatment delivery steps. Natural language processing (NLP)-aided statistical algorithms have the potential to significantly improve the discovery and reporting of these medical errors by relieving human reporters of the burden of event type categorization and creating an automated, streamlined system for error incidents. In this paper, we demonstrate text-classification models developed with clinical data from a full service radiation oncology center (test center) that can predict the broad level and first level category of an error given a free-text description of the error. All but one of the resulting models had an excellent performance as quantified by several metrics. The results also suggest that more development and more extensive training data would further improve future results.