Domain-Specific NLP for Personalized Radiation Treatment Pathways with LLM Fine-Tuning

Abstract

General-purpose large language models (LLMs) such as GPT-4 and LLaMA-2 exhibit strong general reasoning capabilities but struggle to interpret domain-specific biomedical terminology and complex oncological semantics, often producing factual inaccuracies or hallucinations. To address these limitations, this study introduces the Radiation Oncology Model (ROM) a domain-adapted NLP system for personalized radiation treatment planning. ROM is built upon BioGPT and finetuned using the Radiation Oncology NLP Database (ROND) through a Retrieval-Augmented Fine-Tuning (RAG-FT) approach employing Low-Rank Adaptation (LoRA). The model integrates multimodal patient data, including clinical narratives, genomic information, and imaging features, to generate individualized radiotherapy pathways. Experimental results demonstrate significant performance improvements, achieving an average F1-score of 84.09%, surpassing baseline models such as RadOnc-GPT $(\approx 79.06 %)$. ROM also exhibits reduced hallucination rates and improved contextual grounding aligned with clinical guidelines. Qualitative analyses confirm enhanced interpretability, factual consistency, and entity recognition across biomedical categories. These findings highlight that domain-specific fine-tuning and retrieval grounding substantially enhance clinical reliability and reasoning in AI-driven oncology. The proposed ROM framework establishes a pathway toward clinically interpretable and trustworthy LLMs for precision radiotherapy and broader personalized oncology applications.