Abstract No. 576 Comparison of a model for extracting high-quality and clinically significant medical data from unstructured interventional radiology documentation to manual review

Abstract

Interventional radiology generates a large volume of information that contain a wealth of medical data but in highly variable free-form text, which can hinder the ability to harvest data. We designed a pipeline based on natural language parsing (NLP) and advanced pattern matching for 700 unstructured chest port insertion dictations and compared it to a subset of 100 manually reviewed dictations. We obtained 700 chest port insertion dictations from our medical system. Only patient MRN, date/time of procedure, sex, age, and the raw dictation text were included in our initial dataset. We started with a basic approach to parsing via simple text matching. A list of key words and phrases were defined. The program would report if found. This would only identify if the key words were mentioned, but not complex parameters like why and how much certain medications were used (i.e., semantics). Thus, we have been working on implementing pattern matching and natural language parsing into the program. For example, “insertion” would be detected as a “procedure,” and “left internal jugular” as a nearby “anatomical site” that has semantic relationship. A subset of 100 cases was manual reviewed to compare the specificity and sensitivity of our NLP pipeline for the different variables that we extracted (Table). More importantly, the NLP program performed the analysis of 700 cases on a standard desktop computer in less than 5 seconds compared to manual review of the 100 cases that took 6 hours. The comparison of our pipeline to the gold standard of manual review demonstrates the potential accuracy of NLP in IR research and quality improvement. Although building the pipeline requires a significant technical degree of knowledge and effort, it has strong implications in data harvesting efficiency. Future opportunities to utilize this automated pipeline are being explored including template standardization, and quality metric tracking.Tabled 1VariableIndicated for ChemotherapyRight or Left IJ insertionVersed DoseFentanyl DoseSedation TimeSensitivity0.980.990.990.990.97Specificity0.950.990.990.990.99 Open table in a new tab