Abstract No. 478 Procedure dictations to t-test: leveraging natural language parsing in interventional radiology for extracting clinically significant medical data from unstructured dictations

Abstract

Interventional radiology generates a large volume of procedure documentation that contain a wealth of medical data but in highly variable free-form text, which can hinder our ability to analyze and utilize the data. We designed a pipeline based on natural language parsing (NLP) and advanced pattern matching for 700 unstructured chest port insertion dictations as a proof of concept to demonstrate how raw text could be turned into quantitative data for analysis. We pulled 700 chest port insertion dictations from our medical system. Only patient MRN, date/time of procedure, sex, age, and the raw dictations were included in our initial dataset. We preprocessed and validated the data with both R and Python code. Glitches and anomalies from the data procurement process were filtered, the note was formatted, and abbreviations were clarified. We worked on implementing pattern matching and natural language parsing to extract useful quantitative information from the raw text. Apache CTakes, an open-source and free NLP program specifically designed for clinical text, and the National Library of Medicine’s Unified Medical Language System were used to standardize categorical and semantical relationships between words. R was used to perform statistical analysis. Extracted variables: From the 700 cases, we extracted out variables including indication for port insertion, left or right internal jugular approach, quantity of Versed or Fentanyl given, and the total sedation time. Of the 700 cases, the program identified 644 right IJ insertions vs. 40 left IJ insertions. A preliminary t-test identified a statistical significance (P = 0.01) in sedation time between right versus left IJ insertions. Our proof-of-concept pipeline demonstrates the utility of NLP in interventional radiology research and potential quality improvement. Although building the pipeline require a significant technical degree of knowledge and effort, it has potential to be a more effective method to quickly analyze a large dataset of unstructured documents. We are currently investigating the incorporation of more advanced NLP techniques and applying our pipeline on a broader scope of practice.