REFINE: personalized medical concepts from budgeted knowledge graphs

A methods brief on patient-specific EHR concept encoding: RL picks how much KG context each code gets, a GNN + frozen LLM refine representations, reported gains on MIMIC-III/IV backbones.

Primary source: arXiv:2609.04415: https://arxiv.org/abs/2609.04415

What’s new: EHR prediction often embeds ICD (and related) codes as if they mean the same thing for every patient. They don’t: context and trajectory change what a code implies. REFINE (arXiv:2609.04415, submitted 3 Sep 2026) is a KG-aware, budgeted LLM graph-refinement framework for patient-personalized medical concept representations.

Starting from a global text-attributed knowledge graph (TKG), REFINE builds patient-specific temporal graphs. A sequential reinforcement-learning policy chooses a personalized KG expansion budget per observed code (how much neighborhood to pull in). A heterogeneous GNN models relation-aware structure; a frozen LLM applies graph-aware soft prompts to refine concept semantics. On MIMIC-III and MIMIC-IV, the authors report consistent improvements across diverse EHR backbones, outperformance versus strong baselines, and robustness under component ablation, KG selection choices, and data-insufficiency settings (details in the paper, the abstract does not publish a single headline percentage).

Why it matters: This is methods work aimed at a real EHR headache: uniform concept encoders ignore patient-specific meaning and blow the context budget if you dump the whole KG into every code. Budgeted expansion + structure (GNN) + semantics (LLM soft prompts) is a concrete recipe other longitudinal predictors can borrow. Keeping the LLM frozen also matters for cost and for not fine-tuning a giant model on every cohort.

Caveats: Preprint. Claims are MIMIC-centric; codes, graphs, and outcomes elsewhere will differ. “Consistently improves” needs the paper’s tables for effect sizes and which prediction heads actually move. RL budget policies can overfit the development graph ontology. Frozen LLM soft prompts still inherit whatever biases and coverage gaps the base model has. Use REFINE as a personalized-encoding pattern with MIMIC evidence, not as a drop-in clinical decision system.