IPF-RAG: temporal-aware retrieval-augmented generation for idiopathic pulmonary fibrosis care

Personalized clinical management of idiopathic pulmonary fibrosis (IPF) is hindered by its complex longitudinal course and fragmented, unstructured data. Although large language models (LLMs) and retrieval-augmented generation (RAG) show promise in clinical decision support, current methods face challenges in addressing dynamic disease progression due to static knowledge corpora, a lack of temporal awareness, and susceptibility to semantic noise and hallucinations. To address this, we propose IPF-RAG, a temporal-aware RAG framework for longitudinal IPF care. Driven by a curated dataset of 58 patient trajectories, the framework comprises two core modules. The Temporal Retrieval Module augments semantic matching with temporal-aware evidence anchoring, establishing temporal correlations between disease progression and medical orders. Subsequently, the Structured Enhancement Module integrates this temporal knowledge into structured prompts, applying medical logic constraints to guide LLMs toward evidence-based reasoning. Experiments on internal and external datasets, evaluated by automated assessments and clinical experts, demonstrate that IPF-RAG improves the quality of LLM-generated treatment plans across multiple dimensions. These findings indicate that the proposed approach effectively assists LLMs in clinical decision-making for complex chronic diseases.