ReXErr: Synthesizing clinically meaningful errors in diagnostic radiology reports
Year: 2024
Type: other
Abstract: Accurately interpreting medical images and writing radiology reports is a difficult yet clinically significant task. In addition to placing an immense burden upon the healthcare system, manually written radiology reports are prone to human oversights including, but not limited to, clinical inaccuracies as well as grammatical and spelling errors. While automated report-generation algorithms show great promise in improving healthcare efficiency, model-generated reports are liable to different forms of errors, with more frequent discrepancies in diagnoses and varying linguistic mistakes. There is a need for more expansive datasets that capture this broad range of errors to better tune future models and potentially train report- correction algorithms. To this end, we introduce a new methodology that leverages LLMs to generate representative clinician and model errors within chest X-ray reports, using error categories developed in collaboration with board-certified radiologists. We propose a novel sampling scheme to ensure that the different classes of errors are injected at comparable frequencies as well as a modular pipeline to improve the diversity and clinical precision of the errors generated through our prompts. Our method, ReXErr, demonstrates impressive diversity, clinical plausibility, and consistency across all categories of the errors injected. We believe our methodology can serve as a baseline for future LLM-based synthetic medical datasets.
Authors: Vishwanatha M Rao, Serena Zhang, Julian N Acosta, Subathra Adithan, Pranav Rajpurkar
Institution: Harvard University
Language: English
Subfield: Health Informatics
Field: Medicine
Domain: Health Sciences
Open Access status: gold
OpenAlex ID: W4416504837