Artificial intelligence can help with the arduous task of assigning CPT (Current Procedural Terminology) billing codes for interventional radiology procedures, according to new research.
Translating procedures into standardized CPT codes is a critical component of healthcare billing, relying on manual work from specialized experts. This includes interpreting radiologists’ notes, diagnoses and other documentation, researchers detailed in the Journal of Vascular and Interventional Radiology.
Coding can require an average of 12 minutes per procedure, previous studies have found, with up to 80% of medical bills including errors.
“Medical coding is complex across specialties; however, IR procedures pose unique challenges for code assignment,” Hossam A. Zaki, a medical student with Brown University, Providence, Rhode Island, and co-authors detailed Jan. 24. “Notes are highly detailed, unstructured, and variable in style, often involving multiple codes for a single intervention. This complexity is compounded by the fact that IR operates at the forefront of medicine, frequently introducing new procedures for which standardized CPT mappings lag behind clinical practice.”
Large language models offer a potential solution, they believe, helping improve contextualization of procedure notes and allowing for quicker updates as clinical guidelines change. Zaki and colleagues conducted an experiment using XLNet, an open-source LLM, to help predict CPT codes from post-procedure reports. The research team utilized the Medical Information Mart for Intensive Care dataset for their experiment, focusing on the terms “embolization” and “catheter,” along with their associated CPT codes.
Altogether, the study incorporated two different samples—one focused on embolization, including nearly 1,600 reports and 17 distinct CPT code labels. The other combined embolization and catheter cases, yielding 5,600 reports with 42 distinct CPT code labels. XLNet achieved strong performance on the embolization data, researchers found. The best-performing code was 37243 (vascular embolization and occlusion procedures) while the most challenging was 36246 (selective catheter placement in a second-order artery branch).
AI also worked well on the combined catheter-embolization dataset, performing best on code 36597 (other central venous-access procedures) while struggling with codes such as 36010 (IV vascular introduction and injection procedures). Zaki and co-authors believe their proof-of-concept study demonstrates that AI can accurately predict CPT codes from post-procedure reports, though these models must undergo further testing before attempting real-world use.
“IR can benefit from AI-driven billing models, which may reduce administrative burdens, improve accuracy, increase revenue, and lower litigation risk,” the authors concluded. “A key advantage of our study is using XLNet. Its open-source nature allows healthcare institutions to implement and customize the model locally, enhancing HIPAA compliance by keeping patient data secure in-house. Additionally, it democratizes access to advanced AI tools, enabling institutions of all sizes to use cutting-edge technology.”
Read more in the official journal of the Society of Interventional Radiology, including potential study limitations, here.
link