Utilizing Large Language Models to Predict ICD-10 Diagnosis Codes from Patient Medical Records

The application of Large Language Models (LLMs) has been a deeply explored topic, but with little focus on utilizing LLMs for predicting ICD-10 patient diagnoses in the medical field. Using LLMs to predict these patient codes has significant potential to streamline medical processes and reduce human burden drastically. In this work, we use GPT-4o to test four prompting techniques: Base GPT-4o, GPT-4o with Chain of Thought (CoT) prompting, GPT-4o with Retrieval-Augmented Generation (RAG), and GPT-4o with both CoT and RAG. Our results show that combining CoT and RAG significantly improve predictive accuracy of GPT-4o in patient diagnosis.