MedKGent: Evolving Med Knowledge Graph

• MedKGent is a dual-usage framework, initially for medical textual entailment and later for dynamically constructing confidence-aware biomedical knowledge graphs.
• The 2025 system employs two specialized LLM agents—the Extractor and Constructor—to process over 10 million PubMed abstracts and update a temporally stamped Neo4j graph.
• Evaluations indicate that MedKGent improves medical QA, report generation, and drug repurposing by leveraging conflict resolution and cumulative evidence reinforcement.

Searching arXiv for the MedKGent usage and related papers to ground the article. MedKGent denotes, in its most specific and current usage, a LLM agent framework for constructing a temporally evolving medical knowledge graph from biomedical literature. The 2025 system processes more than 10 million PubMed abstracts published between 1975 and 2023, incrementally building a confidence-aware graph whose edges retain timestamps and literature provenance (Zhang et al., 17 Aug 2025). An earlier usage applied the same label to the Medical Knowledge-enriched Textual Entailment framework, also called Sem-KGN, for recognizing medical question entailment through dual encoding of text and knowledge graphs (Yadav et al., 2020). Subsequent papers also use “MedKGent-style” to describe knowledge-graph-centered medical AI systems for report generation, reasoning-probe construction, pharmacy decision support, and automated graph induction, so the term now functions both as a proper system name and as a recognizable architectural motif in medical AI (Liu et al., 2021, Rosenbaum et al., 2024, Farrugia et al., 22 Jun 2025, Khalid et al., 2024).

1. Scope and nomenclature

Within the cited works here, MedKGent has two direct referents. In the earlier one, it names a medical textual entailment framework that enriches BERT with medical knowledge graphs for MEDIQA-RQE. In the later and more formalized one, it is the title of a two-agent framework for constructing a temporally evolving medical knowledge graph from PubMed abstracts (Yadav et al., 2020, Zhang et al., 17 Aug 2025).

Usage	Paper context	Core task
MedKGent / Sem-KGN	Medical Knowledge-enriched Textual Entailment	Medical recognizing question entailment
MedKGent	LLM agent framework for temporally evolving medical KG	Incremental KG construction from PubMed

The later literature broadens the term’s connotation. KGAE is described as a blueprint for a “MedKGent”-style system in which a medical knowledge graph serves as a shared latent space for unsupervised medical report generation. MedG-KRP is described as a concrete instantiation of what a system like “MedKGent” would do by forcing an LLM to externalize medical knowledge as an explicit causal graph. medicX-KG is presented as a reference blueprint for how a knowledge graph can function as the semantic backbone of a MedKGent-like pharmacy agent (Liu et al., 2021, Rosenbaum et al., 2024, Farrugia et al., 22 Jun 2025). This suggests that MedKGent has come to denote not only a single implementation, but also a recurring design pattern in which a medical knowledge graph provides grounding, retrieval structure, or reasoning scaffolding.

2. The 2025 MedKGent framework

The 2025 MedKGent framework is built around a temporally ordered biomedical corpus. After filtering by abstract length and year, it retains 10,014,314 PubMed abstracts and organizes them as a fine-grained daily time series from 1975-01-01 to 2023-12-31. The system is implemented as two specialized agents powered by Qwen2.5-32B-Instruct: an Extractor Agent and a Constructor Agent (Zhang et al., 17 Aug 2025).

The Extractor Agent operates at the level of a single abstract. It first uses PubTator3 to identify and normalize six entity types—Gene, Disease, Chemical, Variant, Species, and CellLine—then prompts the LLM to infer relations under a 12-relation schema. That schema includes the bidirectional relations Associate, Negative_Correlate, Positive_Correlate, Compare, Cotreat, Interact, and Drug_Interact, as well as the unidirectional relations Cause, Inhibit, Treat, Stimulate, and Prevent. Rather than accepting a single decoding, the Extractor Agent performs sampling-based self-consistency: it runs the LLM 50 times at temperature $0.7$, counts how often each triple appears, discretizes the resulting frequency downward to the nearest 0.05, and retains only triples with confidence at least 0.6.

The Constructor Agent then processes the retained triples in publication-date order and updates a Neo4j graph day by day. If a node is new, it is inserted with biomedical identifiers and textual metadata. If an extracted triple already exists, its confidence is reinforced and its provenance list is extended. If the same entity pair is already connected by a different relation, the Constructor Agent invokes a conflict-resolution prompt that considers competing relations together with confidence scores and timestamps. The result is not a static bulk-aggregated graph, but a temporally evolving knowledge structure whose state reflects repeated evidence, contradiction handling, and chronological accumulation.

3. Confidence, temporal updating, and graph representation

MedKGent’s graph representation is explicitly provenance-rich. Each node stores an Identifier, Entity Type, Terminology, Page Link, Exact Keywords, and a 768-dimensional semantic embedding derived from BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext. Exact Keywords collect lowercased aliases observed in the literature, while the embedding layer supports semantic retrieval when exact lexical matching is unavailable. Each edge stores a relation label, a confidence score, a list of supporting PubMed IDs, and a timestamp corresponding to the latest supporting publication (Zhang et al., 17 Aug 2025).

Confidence is treated as an updateable graph attribute rather than a by-product of extraction. When a triple with confidence $s'$ matches an already stored triple with confidence $s$, MedKGent updates the edge weight by

$s_{\text{new}} = 1 - (1 - s)(1 - s').$

This multiplicative reinforcement rule monotonically increases confidence under repeated supporting evidence and asymptotically approaches 1 under consistent recurrence. The framework therefore operationalizes a simple but consequential distinction between first appearance and repeated corroboration.

Conflict resolution is equally central. When two relations compete for the same head-tail pair, MedKGent does not keep both by default. Instead, it passes the competing relations, their confidence scores, and their timestamps to a low-temperature conflict-resolution prompt and retains only the relation judged most appropriate. This design makes the graph a curated temporal state rather than an unconstrained multiset of LLM outputs. After consolidation, the resulting graph contains 156,275 entities and 2,971,384 relational triples.

4. Evaluation and downstream utility

Quality assessment combines automatic and human evaluation. On a 1% subset comprising 34,725 abstracts and 83,438 triples, GPT-4.1 judged 85.44% of triples valid and DeepSeek-v3 judged 88.10% valid when validity was defined as scores of at least 2.0 on a four-point rubric. On a separate human evaluation over 400 abstracts and 1,060 triples, three domain experts each assessed more than 86% of triples as valid. Pairwise Cohen’s $\kappa$ further indicated substantial agreement for 80% of rater pairs, supporting the claim that MedKGent’s extraction quality approaches 90% under both automated and expert review (Zhang et al., 17 Aug 2025).

The framework was also evaluated as a retrieval substrate for question answering. Its RAG pipeline performs entity extraction on a query, aligns the extracted entities to graph nodes through identifier matching, keyword matching, or BiomedBERT semantic similarity, retrieves one-hop neighborhoods, and then reranks candidate triples with an LLM before passing the top $k=5$ triples to a downstream answer generator. Across seven medical QA benchmarks—MMLU-Med, MedQA-US, PubMedQA*, BioASQ-Y/N, and the three MedDDx subsets—and five LLMs, KG augmentation consistently improved accuracy. On MedQA-US, GPT-3.5-turbo improved from 0.5986 without RAG to 0.6834 with RAG, and on MedDDx-Expert the gains reached up to +8.6 points.

A more exploratory downstream use is literature-based drug repurposing. The paper studies tocilizumab and COVID-19 through a path pattern of the form $$Chemical \xrightarrow{\text{Negative\_Correlate}} Gene \xrightarrow{\text{Positive\_Correlate}} Disease$$. For a candidate chemical-disease pair $(c,d)$, MedKGent defines

$$S(c,d) = \frac{1}{|P|} \sum_{p \in P} \prod_{(e_i, r, e_j) \in p} s(e_i, r, e_j),$$

where $P$ is the set of supporting paths and $s'$0 is the edge confidence. In the tocilizumab case, the framework identifies pre-existing paths through mediators such as FGB and TNF before explicit tocilizumab–COVID-19 treatment evidence appears in the literature, illustrating how temporal path structure and confidence-aware aggregation can support hypothesis generation.

5. Earlier MedKGent: knowledge-enriched textual entailment

Before the 2025 graph-construction framework, the name MedKGent was used for a medical textual entailment system introduced as the Medical Knowledge-enriched Textual Entailment framework and referred to in the paper body as Sem-KGN. The task is MEDIQA-RQE: given a consumer health question $s'$1 and an FAQ-style hypothesis question $s'$2, predict whether $s'$3 adequately answers the information need of $s'$4. The framework was designed to address two specific failures of strong LLMs in this setting: difficulty with multiple word forms, synonymy, and abbreviations, and difficulty with focus or intent understanding in lexically similar but semantically distinct question pairs (Yadav et al., 2020).

Sem-KGN has three major components. A Document Encoder based on BERT base uncased produces token-level contextual representations. A Knowledge-enriched Graph Encoder builds a vocabulary graph from the corpus using normalized pointwise mutual information, expands it with two-hop concepts from UMLS, SNOMED-CT, ICD-10, and Clinical Trials, and applies a 2-layer GCN to obtain knowledge-enriched entity embeddings. A Multi-Headed Aggregator then concatenates textual and graph-derived representations and passes them through a stacked Transformer with 12 blocks and 16 attention heads before classification. The model is therefore a dual-encoding architecture: local textual context is supplied by BERT, while global semantic and factual context comes from the graph encoder.

On MEDIQA-RQE, Sem-KGN reports 56.17 accuracy, 63.18 precision, 56.18 recall, and 59.47 F1. This corresponds to an absolute improvement of 8.27% over the vanilla BERT baseline at 47.90 accuracy and a 4.87-point improvement over the prior “BERT + linear projection” result at 51.30 accuracy. Ablation showed that removing the Knowledge-enriched Graph Encoder reduced accuracy to 47.90, while using the vocabulary graph without medical KG expansion reached 50.65. The earlier MedKGent usage is thus best understood as a knowledge-graph-enriched entailment model rather than a graph-construction framework.

6. Broader design pattern, limitations, and outlook

Later papers use MedKGent as a reference point for several distinct medical AI architectures. KGAE presents a knowledge graph as a shared latent space bridging images and reports in unsupervised, semi-supervised, and supervised report generation. MedG-KRP treats an LLM-generated causal graph as an externalized proxy for internal biomedical reasoning and evaluates it with both human review and BIOS alignment. medicX-KG treats a pharmacist-oriented knowledge graph as the semantic layer of a broader platform supporting queries about authorisation, interactions, adverse reactions, and therapeutic classes. M-KGA emphasizes automated KG generation and enrichment through BioPortal ontologies and Clinical BERT similarity (Liu et al., 2021, Rosenbaum et al., 2024, Farrugia et al., 22 Jun 2025, Khalid et al., 2024). This suggests a stable conceptual core: MedKGent-style systems use medical knowledge graphs not merely as storage, but as latent spaces, reasoning scaffolds, retrieval substrates, or semantic backbones.

The limitations also differ by usage. For the 2025 MedKGent, the corpus is restricted to abstracts rather than full texts, domain coverage still excludes sources such as clinical notes and trial registries, confidence is based on sampling frequency rather than calibrated uncertainty, and daily publication date is only a proxy for the actual chronology of discovery (Zhang et al., 17 Aug 2025). For the earlier entailment-oriented MedKGent, performance depends heavily on entity recognition and linking quality, KG coverage may still be incomplete despite using multiple medical ontologies, and the combined vocabulary graph, graph expansion, and 12-layer Transformer increase computational cost (Yadav et al., 2020).

The documented future directions remain consistent with that broader pattern. The temporally evolving KG framework proposes integrating additional data sources and refining confidence modeling, while the entailment framework proposes multi-domain knowledge graphs and more sophisticated graph embedding techniques. A plausible implication is that MedKGent is evolving from a single named system into a family of knowledge-centric medical AI methods unified by explicit graph grounding, structured uncertainty handling, and close coupling between symbolic biomedical relations and neural generation or inference.