Clinical Data is a Graph: Limitations of FHIR References and Context Window Issues
Why FHIR Retrieval Can Overflow Agent Context - bonfireDB
Summary
A benchmark experiment revealed that when answering a single clinical question, an AI agent can fail to find information due to the process of traversing large amounts of data, resulting in incorrect 'cannot find' answers. This failure stems from structural challenges inherent in FHIR reference pointers and the cumulative constraints of the context window.
Details
This article emphasizes that clinical data must be viewed not merely as rows but as a graph structure. Specifically, FHIR references are pointers, not the actual data itself; thus, retrieving necessary information requires traversing node by node. In a benchmark experiment, answering a single question—such as a patient's earliest abnormal HbA1c value—showed that the agent missed the result while sequentially fetching data spread across multiple pages. This failure is structural: the data did not fail to exist; it failed to arrive due to context window overflow (32k tokens). Furthermore, the article highlights that cumulative overflow caused by
Original content copyright by respective publishers