An open-source LLM accurately assigns CPT codes from interventional radiology postprocedure reports in a proof-of-concept study.
Key Details
- 1Researchers fine-tuned the XLNet model using over 7,000 IR postprocedure reports from the MIMIC-IV dataset.
- 2On an embolization dataset (n=1,590; 17 CPT codes), XLNet achieved AUROC 0.93, AUPRC 0.86, F1 score 0.84.
- 3On an embolization+catheter dataset (n=5,590; 42 CPT codes), AUROC reached 0.99, AUPRC 0.86, F1 score 0.85.
- 4Best performance seen for codes like 37243 and 36597, with some codes (e.g., 36246, 36010) more challenging.
- 5The study suggests LLMs could reduce billing inefficiencies, noting current manual coding takes 12 minutes per procedure and up to 80% of bills contain errors.
- 6Authors recommend stepwise deployment before fully autonomous use, starting with human confirmation.
Why It Matters
Automating CPT code assignment in interventional radiology could significantly improve billing accuracy, reduce administrative burden, and enhance revenue integrity. Early success with open-source LLMs demonstrates potential for broader application in radiology coding workflows.
Source
AuntMinnie
Related News
•AuntMinnie
Study: Computer Vision Models Best LLMs in Chest CT Breast Abnormality Detection
Computer vision models (CVMs) surpass large language models (LLMs) in accurately labeling incidental breast abnormalities on chest CT scans.
•AuntMinnie
Deep Learning Models Rival Radiologists for Pancreatic Cancer Detection on CT
Deep-learning models achieved comparable or superior accuracy to experienced radiologists in detecting pancreatic cancer on CT scans, especially for small tumors.
•Radiology Business
Radiology AI Devices at Elevated Risk for FDA Recalls, Study Finds
Radiology AI devices are more likely to face FDA recalls, largely due to deviations from intended use and incomplete clinical data.