FUNNEL: Architectural Citation Enforcement for AI-Assisted Clinical Chart Synthesis
Applied Clinical Informatics · 01.10.2026 · 17 (05) · 1010-1014
DOI: 10.1055/a-2997-2386 ↗Kurz gefasst
Fachartikel in der Zeitschrift Applied Clinical Informatics, erschienen am 01.10.2026. Verfasst von Rugved Parmar und 3 weiteren Autor:innen.
Abstract (Originalsprache)
Background
Clinical decision support systems leveraging large language models (LLMs) risk generating unsourced or fabricated information when synthesizing electronic health record (EHR) data. Existing retrieval-augmented generation (RAG) approaches either provide context without enforcing citation or rely on prompt-based instructions that produce inconsistent source attribution.
Objectives
This study aimed to develop and evaluate a RAG framework that architecturally enforces citation of source documents for every claim in AI-generated clinical summaries, and to quantitatively compare its performance against prompt-based and unconstrained alternatives.
Methods
We developed FUNNEL (Focused Unified Narrative Network via Evidential Layering), a three-stage pipeline comprising hybrid retrieval with temporal weighting, LLM-based reranking, and constrained synthesis with mandatory citation and post-generation validation. We evaluated 50 clinical queries across seven categories under three conditions—vanilla LLM, prompt-based citation, and FUNNEL—on a 128-document synthetic longitudinal corpus with known temporal conflicts and unanswerable queries.
Results
FUNNEL achieved 98.5% citation coverage versus 83.7% for prompt-based citation and 10.3% for vanilla LLM. Citation precision was 98.3 versus 94.9% and 19.8%, respectively. The unsupported claim rate was 0.0% for FUNNEL versus 1.8 and 6.5%. FUNNEL correctly refused all unanswerable queries and adversarial prompts requesting external knowledge. Average cost was $0.012 per query.
Conclusion
Layered architectural enforcement of citation—combining context restriction, constrained synthesis, and post-generation validation—produces measurably superior source attribution compared with prompt-based approaches alone.
Zitieren
APA
Parmar, R., Eldawud, D., Pandal, P., & Budzikowski, A. (2026). FUNNEL: Architectural Citation Enforcement for AI-Assisted Clinical Chart Synthesis. Applied Clinical Informatics, 17(05), 1010-1014. https://doi.org/10.1055/a-2997-2386
Vancouver
Parmar R, Eldawud D, Pandal P, Budzikowski A. FUNNEL: Architectural Citation Enforcement for AI-Assisted Clinical Chart Synthesis. Applied Clinical Informatics. 2026;17(05):1010-1014. doi:10.1055/a-2997-2386
BibTeX
@article{Parmar2026,
author = {Parmar, Rugved and Eldawud, Daoud and Pandal, Perman and Budzikowski, Adam},
title = {FUNNEL: Architectural Citation Enforcement for AI-Assisted Clinical Chart Synthesis},
journal = {Applied Clinical Informatics},
year = {2026},
volume = {17},
number = {05},
pages = {1010--1014},
doi = {10.1055/a-2997-2386}
}
EUVIMED