AI may be able to speed up patient care.
Eyal Klang, chief of generative AI, Windreich Department of Artificial Intelligence and Human Health at the Icahn School of Medicine at Mount Sinai, explores one way to do so.
Eyal Klang, MD, is Chief of Generative AI in the Windreich Department of Artificial Intelligence and Human Health at the Icahn School of Medicine at Mount Sinai. His research focuses on using artificial intelligence to improve healthcare delivery and clinical decision-making
Medical coding assigns ICD-10 codes to each diagnosis. These codes go into the medical record, but the process is slow and often wrong. It also takes time away from patient care.
Large language models can read doctors’ notes and suggest codes. But on their own they make mistakes, even inventing codes that don’t exist, a problem called hallucinations.
We tested a method called retrieval-augmented generation, or RAG. The model first suggests a diagnosis from the note. That text is then checked against ICD-10 codes used in our emergency rooms. The system returns the closest matches, and the model picks from those. This keeps the output tied to real codes.
We studied 500 emergency visits. Independent doctors compared provider-assigned codes with RAG-assisted codes. In most cases, they judged the AI codes more accurate and more specific. Even smaller open-source models improved when paired with RAG.
The goal is to cut paperwork and improve records. Better coding means fewer mistakes, cleaner data, and more time for doctors to spend with patients.
Our results show that retrieval-augmented AI can make coding faster, more accurate, and less of a burden for clinicians.
Read More:
[NEJM AI] - Assessing Retrieval-Augmented Large Language Models for Medical Coding
