New Research Advances AI-Powered Medical Knowledge Systems for Clinical Decision Support

Researchers have introduced MedKGent, a large language model agent framework that constructs a temporally evolving medical knowledge graph from over 10 million PubMed abstracts. The resulting graph contains 156,275 entities and 2,971,384 triples, achieving a 90% validity rate and improving retrieval-augmented generation across seven benchmarks. Concurrently, platforms like Medvolt's MedGraph EDGE leverage structured biomedical knowledge graphs to support clinical decision systems, target discovery, and translational medicine.

Medical imaging annotation tools

• ▪M. Aljabri et al. published a 2022 survey in Multimedia Tools and Applications analyzing annotation tools for medical imaging.
• ▪M. Neves and U. Leser published a 2014 survey in Briefings in Bioinformatics evaluating annotation tools for biomedical literature.

Large language models in healthcare

• ▪The MedKGent framework uses over 10 million PubMed abstracts from 1975 to 2023 to incrementally construct a medical knowledge graph daily.
• ▪The MedKGent medical knowledge graph contains 156,275 entities and 2,971,384 triples, achieving triple-validity rates approaching 90% in evaluations.
• ▪The MedKGent framework utilizes an Extractor Agent to identify knowledge triples and a Constructor Agent to integrate them into a temporal graph.
• ▪The MedKGent-derived knowledge graph improved retrieval-augmented generation for five large language models across seven medical question-answering benchmarks.

Medical image segmentation techniques

• ▪Q. Nie et al. developed a reparameterized multi-scale transformer for deformable retinal image registration, published in 2025.
• ▪H. Shao et al. developed MCANet, a medical image segmentation method utilizing multi-scale cross-axis attention, published in 2025.

Clinical decision support systems

• ▪Biomedical knowledge graphs face challenges in clinical decision support systems due to data quality, computational scalability, and model interpretability.
• ▪The Medvolt MedGraph EDGE platform integrates multi-source biomedical data into structured graphs to support target discovery, biomarker identification, and translational medicine workflows.

Medical data pre-training methods

• ▪Y. Qiu et al. published a 2023 survey in Machine Intelligence Research focusing on pre-training methods in medical data.
• ▪K. Sun et al. analyzed the applications, challenges, and future directions of medical multi-modal foundation models in clinical diagnosis in 2025.

3 sources