Home // EXPLAINABILITY 2025, The Second International Conference on Systems Explainability // View article

Bridging the Domain Gap: Evaluating Fact-Grounded Knowledge Graph Narratives for Explainable AI in Clinical Decision Support

Authors:
Valentin Göttisheim
Holger Ziekow
Peter Schanbacher
Djaffar Ould-Abdeslam

Keywords: xplainable Artificial Intelligence; XAI; Knowledge Graphs; Shapley Additive Explanations; SHAP; Narrative Generation; Claim Verification.

Abstract:
Clinicians need transparent reasoning to trust Artificial Intelligence recommendations, but standard explanation methods lack clinical semantics. To address this, we transform an Onkopedia colon carcinoma guideline into a semantically enriched Knowledge Graph by segmenting text, extracting and merging semantic concepts, enriching gaps with registry data, and anchoring features to graph nodes. Using a predictive model, we compute Shapley Additive Explanations feature attributions and generate fact-grounded narratives via large language models that directly reference guideline evidence. We compare three contexts across 65 synthetic colorectal cancer cases (195 narratives) and find that KG-based narratives reduce hallucinations, speculation, and contradictions. Embedding KG-grounded narratives in clinical decision-support tools promises to shorten expert review cycles, surface guideline deviations, and bridge the explainability gap between data scientists and clinicians.

Pages: 28 to 33

Copyright: Copyright (c) IARIA, 2025

Publication date: October 26, 2025

Published in: conference

ISBN: 978-1-68558-318-7

Location: Barcelona, Spain

Dates: from October 26, 2025 to October 30, 2025