Characterizing And Quantifying Diagnostic (Un)Certainty In Medical Reports Through Natural Language Processing

Miscommunication of diagnostic uncertainty can deeply affect the quality of treatment a patient receives. A standardized quantification based on the language used in medical reports is a solution for gaining clarity about the amount of uncertainty an author intended to convey. We use natural language processing techniques to create a dictionary of terms and phrases used in a corpus of radiology reports that are indications of uncertainty or certainty. Using this dictionary, we model reports by analyzing them as both a collection of sentences and a collection of words. We assign reports a rating on a scale of 0-5 to quantify how uncertain a particular report is. Our results suggest that by using a dictionary of both certainty and uncertainty descriptors, we can characterize and quantify diagnostic uncertainty of medical reports.