FHIR-Hopper: A Neuro-Symbolic Agent for Reasoning over EHRs

Chanwoo Kim, Gunjan Jalori, Justin Anthony Chen, Dimitrios Antos, Xin Liu, A. Ali Heydari

Keywords: Electronic health records, FHIR, Agentic AI, Large Language Models, Clinical question answering

TL;DR: FHIR-Hopper is a neuro-symbolic agentic framework that projects complex, nested EHR data into an episodic chronological graph, allowing LLMs to reason more accurately while reducing input token usage by 10x compared to previous methods.

Abstract: Electronic health records (EHRs), increasingly exchanged via the Fast Healthcare Interoperability Resources (FHIR) standard, encode patient histories as deeply nested, multi-relational bundles that are difficult for large language models (LLMs) to reason over directly. We introduce \emph{FHIR-Hopper}, a neuro-symbolic agentic framework that projects a FHIR bundle into an episodic chronological graph, materializes a budgeted, saliency-ranked linearization as the LLM's initial context, and exposes deterministic graph-traversal tools so the agent can recover details on demand. Across three realistic FHIR clinical benchmarks, FHIR-Hopper attains the highest accuracy across multiple base LLMs \emph{while keeping average input-token usage stable at \(\sim20\)k tokens, a typical \(\sim10\times\) reduction over the strongest retrieval baseline on long records}. These results suggest that decoupling structure-aware retrieval from neural reasoning is an effective design for clinical question answering over structured EHRs.

Submission Number: 151