ACIE Agentic RAG: Clinical Extraction at University Medicine Essen

ACIE, an on-premise agentic RAG pipeline, has been deployed at University Medicine Essen, and the associated paper was submitted on 17 June 2026. The system reasons over complete patient contexts and "grounds every answer in source passages for clinician verification." Across 7,326 clinician judgments, reviewers accepted 96.5% of extractions, with per-type acceptance from 80% to 99%.

What did the team build and deploy?

ACIE is an on-premise agentic retrieval-augmented generation pipeline that operates over entire patient contexts and produces source-grounded extractions. The paper describes ACIE as an agentic RAG pipeline deployed at University Medicine Essen that reasons over complete patient contexts and returns answers tied to source passages so clinicians can verify them.

The authors frame ACIE as a response to missing or incomplete document-level metadata. They say standard retrieval-augmented generation fails on clinical datasets because it mishandles temporal reasoning, cross-document dependencies, and missing metadata. ACIE’s design choices aim to address those gaps by explicitly working with full patient contexts and providing provenance for each extracted value.

How was ACIE evaluated and what were the results?

ACIE was evaluated alongside an independent retrospective lymphoma registry study, with nuclear-medicine physicians verifying every extracted value against its cited source passages. The evaluation produced 7,326 clinician judgments and an overall acceptance rate of 96.5% for the system’s extractions, while acceptance by extraction type ranged from 80% to 99%.

The paper quantifies a metadata gap in clinical records and traces how that gap shaped architectural decisions for ACIE. The authors emphasize clinician verification: every extracted value in the lymphoma registry study was checked by nuclear-medicine physicians against the passages ACIE cited.

Why does this matter?

Clinical records span hundreds of heterogeneous documents and thousands of structured datapoints, yet missing document-level metadata undermines retrieval and triage. ACIE’s approach — reasoning over complete patient contexts and grounding outputs in source passages — directly targets those failure modes (temporal reasoning, cross-document dependencies, missing metadata) that the authors identify in standard RAG pipelines. High clinician acceptance in a retrospective lymphoma registry study suggests the approach can produce verifiable extractions clinicians trust.

That trust matters because clinicians must be able to confirm which source supports a given extracted value. By returning citations for each answer, ACIE makes human verification feasible and auditable in settings where incorrect extractions carry clinical risk.

What to watch

Look for external replications that apply ACIE to other disease registries or clinical domains and for published breakdowns of the per-type acceptance rates. The paper’s next confirmatory signals would be deployments or studies that reproduce the 7,326-judgment scale evaluation or show similar acceptance ranges in different specialties.