AI and the World's Hard Problems · Medicine & drug discovery · Nature (computational biology) · 2026-10-23
Instruction-tuning large language models with medical classification data improves ICD coding of German tumor diagnoses
This study tackles automated ICD coding of German tumor diagnosis texts using locally deployable open-weight language models. Eight Qwen, Llama and Mistral models (7–70 billion parameters) were fine-tuned on over 500,000 question-answer pairs built from public classification resources. ICD-10-GM accuracy rose from 1.4% to 24% and then to 41–58%, malformed outputs fell to nearly 0%, and the training dataset and best checkpoints were released publicly.