Related Experiment Video
Updated: Jun 13, 2026
05:47
Evidence-based Knowledge Synthesis and Hypothesis Validation: Navigating Biomedical Knowledge Bases via Explainable AI and Agentic Systems
Published on: June 13, 2025
Rare-Disease Diagnosis on the ZebraMap Multimodal Case Report Dataset: A Hybrid Pipeline with Grounded Explainability
Md Sanzidul Islam1,2, Amani Jamal1,3, Ali Alkhathlan1
1Department of Computer Science, Faculty of Computing and Information Technology, King Abdulaziz University, Jeddah 21589, Saudi Arabia.
Sensors (Basel, Switzerland)
|June 12, 2026
Summary
This study introduces a hybrid AI pipeline for rare-disease diagnosis, combining retrieval and classification to improve differential diagnosis and provide explanations. The system supports clinicians by ranking potential rare diseases and offering evidence-based justifications.
Area of Science:
- Medical Informatics
- Artificial Intelligence in Medicine
- Computational Biology
Background:
- Rare-disease diagnosis presents significant challenges due to vast, imbalanced disease spaces and dispersed evidence across clinical data.
- Existing diagnostic tools often struggle with multimodal data integration and providing interpretable explanations.
Purpose of the Study:
- To develop and evaluate a hybrid AI pipeline for ranked rare-disease diagnosis and grounded explanation generation.
- To assess the performance of multimodal fusion techniques in rare-disease differential diagnosis.
- To quantify the quality and utility of AI-generated diagnostic explanations.
Main Methods:
- A sequential hybrid pipeline combining BM25 lexical retrieval, TF-IDF classification, MedCPT dense retrieval, cross-encoder reranking, and caption-mediated multimodal fusion was developed.
- The pipeline was evaluated on the ZebraMap multimodal case-report dataset using grouped train-validation-test splitting.
- An explanation module was integrated to provide grounded justifications for diagnostic predictions.
Main Results:
- The hybrid pipeline achieved a Mean Reciprocal Rank (MRR) of 0.3905 and Recall@10 of 0.5507.
- A class-balanced TF-IDF classifier, an individual component, showed higher diagnostic accuracy (MRR 0.4200, Recall@10 0.6279).
- The explanation module demonstrated citation coverage of 0.7334 and usefulness of 3.8734, indicating a trade-off between diagnostic accuracy and explanation richness.
Conclusions:
- A hybrid retrieval-and-classification approach effectively supports ranked rare-disease differential diagnosis.
- Grounded explanation quality in AI diagnostic systems can be quantitatively evaluated.
- This computational approach extends support for the complex and prolonged rare-disease diagnostic journey.