MedFusion-Agent: A Multi-Agent Framework for Conflict-Aware Clinical Pharmacotherapy Planning in Multimorbidity
IEEE Access · 2026 · 14 · 93913-93934
DOI: 10.1109/access.2026.3704838 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Yang-Yen Ou und 2 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref) und nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Multimorbidity pharmacotherapy planning requires reconciling conflicting guideline recommendations, medication safety constraints, and therapeutic completeness. Existing large language model (LLM)-based or collaboration-only systems can generate plausible suggestions, but they provide limited support for stable conflict resolution and regimen-level safety checking. In this paper, MedFusion-Agent is proposed as a neuro-symbolic multi-agent framework that integrates guideline interpretation with optimization-based conflict arbitration, deterministic safety checking, and iterative regimen repair. It transforms unstructured guideline knowledge into structured candidate actions. It then applies mandatory safety screening before finalizing the regimen, while preserving evidence traceability and risk annotations for clinical review. The framework was evaluated using GPT-4o and Gemini 3 Pro backbones on a 500-case synthetic boundary stress cohort designed to probe contraindications, hidden interaction risks, and treatment adjustment errors. Compared with single-LLM, naive retrieval-augmented generation, and natural-language multi-agent discussion baselines, MedFusion-Agent achieved the lowest overall hazard rates under both backbones, reducing them to 1.8% and 1.4%, respectively, while also lowering mean drug-drug interaction rates to 0.080 and 0.063 and average numbers of prescribed drugs to 3.92 and 3.88. These results indicate fewer harmful interacting medication pairs and lower medication burden. On MIMIC-IV, MedFusion-Agent outperformed RETAIN, GAMENet, SafeDrug, COGNet, and AMGNet, improving the F1-score to 0.7121 and reducing the drug-drug interaction rate to 0.0580. These findings suggest that explicit cross-guideline arbitration, mandatory safety verification, and iterative repair improve both safety and recommendation quality in LLM-assisted pharmacotherapy planning for complex multimorbidity.
Zitieren
APA
Ou, Y., Chen, C., & Wang, J. (2026). MedFusion-Agent: A Multi-Agent Framework for Conflict-Aware Clinical Pharmacotherapy Planning in Multimorbidity. IEEE Access, 14, 93913-93934. https://doi.org/10.1109/access.2026.3704838
Vancouver
Ou Y, Chen C, Wang J. MedFusion-Agent: A Multi-Agent Framework for Conflict-Aware Clinical Pharmacotherapy Planning in Multimorbidity. IEEE Access. 2026;14:93913-93934. doi:10.1109/access.2026.3704838
BibTeX
@article{Ou2026,
author = {Ou, Yang-Yen and Chen, Ching-Yuan and Wang, Jhing-Fa},
title = {MedFusion-Agent: A Multi-Agent Framework for Conflict-Aware Clinical Pharmacotherapy Planning in Multimorbidity},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {93913--93934},
doi = {10.1109/access.2026.3704838}
}
EUVIMED