Enhanced Medical Coding Accuracy with AI Using Lookup Steps

A recent study conducted by researchers at Mount Sinai Health System highlights a simple yet effective modification in how artificial intelligence (AI) assigns diagnostic codes, leading to notable improvements in accuracy. The research, published in NEJM AI, suggests that adding a lookup step to the AI process not only reduces errors but can also outperform human physicians in certain cases. This advancement has the potential to streamline medical documentation, reduce billing inaccuracies, and enhance the integrity of patient records.

Traditionally, assigning ICD diagnosis codes—a complex set of alphanumeric identifiers used to catalog medical conditions—has been a time-consuming task for healthcare providers. Large language models like ChatGPT frequently struggle to assign the correct codes, sometimes producing nonsensical or irrelevant results. To address this, the Mount Sinai team implemented a "lookup-before-coding" approach, where the AI first generates a plain-language diagnosis description, then retrieves similar historical cases from a vast database of over a million hospital records to inform its coding decision.

In the study, the team evaluated nine different AI models on 500 emergency department visits, analyzing the effectiveness of the retrieval-enhanced method. The process involved matching the AI-generated diagnosis descriptions with the most relevant similar cases, then selecting the most accurate ICD code based on this information. Emergency physicians and independent AI systems reviewed the coding accuracy without knowledge of whether the codes were generated by humans or AI.

Results showed that models utilizing the lookup step significantly outperformed those without it, sometimes even surpassing physician-coded results. Notably, even small open-source AI models performed well when leveraging this retrieval method, demonstrating the approach's broad applicability. According to Dr. Girish N. Nadkarni, a senior author, this technique embodies smarter support for clinicians rather than automation for automation’s sake, aiming to reduce routine administrative burdens.

The researchers emphasize that this retrieval-augmented method is designed to support, not replace, human oversight. Although it currently focuses on primary diagnosis codes from emergency visits discharged home, the team is optimistic about expanding its use to other clinical settings and incorporating secondary and procedural codes. The pilot integration into Mount Sinai's electronic health records system is underway, with hopes of broader deployment.

This innovation underscores AI's transformative potential in healthcare—liberating clinicians from administrative tasks to spend more quality time with patients, ultimately fostering better care and operational efficiency. As Dr. David L. Reich from Mount Sinai notes, such technology can help build more responsive and compassionate healthcare systems.

Source: https://medicalxpress.com/news/2025-09-adding-lookup-ai-assigning-medical.html