Large Language Models for Automatic ICD-10 Coding in Cancer Treatment and Billing: An Academic-Industry Partnership

Abstract

The increasing complexity of medical diagnosis and billing necessitates improving International Classification of Diseases (ICD) coding practices. It also requires partnerships between academia and industry to support rigor and implementation in real-world settings. We conducted an academic-industry collaboration involving four institutions to explore pretrained language models for artificial intelligence (AI)-driven ICD-10 coding in cancer-specific populations. We created a cancer-focused dataset comprising 101,224 clinical notes and 36,040 ICD code assignments across five note types, reflecting real-world scenarios. By fine-tuning and evaluating two existing pretrained language-model approaches designed to process long contexts, PLM-ICD and KEPTLongformer, we developed an institutional benchmark for oncology ICD-10 coding using 20 selected three-character ICD-10 parent codes. The best-performing model achieved an F1-macro of 0.768 and an F1-micro of 0.792 on the test set. Our feasibility analysis showed that integrating these models into clinical workflows could potentially reduce coding time by approximately 10 minutes per case, with a GPU runtime of 7 seconds. Additionally, an exploratory clinical-coder evaluation of attention-based model interpretability showed that 13 of 20 predictions contained high-attention tokens aligned with the corresponding ICD code descriptions; coders rated these tokens as "very helpful" for code assignment. The results demonstrate the potential of AI-driven coding support systems within clinical workflows.