Improving Meditron for Medical Coding Through Fine-Tuning: A Comparative Evaluation Against GPT-4

Medical coding is essential but burdensome. In a prior study, Meditron3-70B showed poor coding accuracy on clinical vignettes. We compared GPT-4 with Meditron-CHUV, a version fine-tuned on diverse clinical tasks at Lausanne University Hospital, for generating ICD-10 (billing) and SNOMED-CT (clinical) admission diagnoses from anamnesis alone. Five blinded clinicians rated 40 synthetic vignettes. GPT-4 was preferred overall, but Meditron-CHUV performed better in adult medicine. Model choice varied by specialty (p < 0.001), not profession. SNOMED-CT codes were systematically invalid despite plausible labels. Results confirm LLM limitations and support hybrid methods that combine retrieval and ontologies. Real-data validation is planned.