Medical Knowledge Graph: Structure & Insights

• Medical Knowledge Graph (KG) is a structured, semantically enriched network integrating diverse biomedical data from literature, EHRs, and curated databases.
• They employ multi-step pipelines with entity extraction, normalization, and transformer-based relation extraction to capture dynamic biomedical discoveries.
• High-confidence edges derived from sampling-based estimations enable robust applications in question answering and drug repurposing with measurable improvements.

A medical knowledge graph (KG) is a structured, semantically rich representation of biomedical entities and their relationships, automatically or semi-automatically assembled from diverse sources such as scientific literature, ontologies, drug leaflets, electronic health records (EHRs), and expert-curated databases. As a central resource for the organization, retrieval, and inference of clinical and translational knowledge, medical KGs serve a critical role in areas spanning question answering, clinical decision support, drug repurposing, and biomedical discovery. Sophisticated construction pipelines leverage entity normalization, relation extraction, temporal tracking, confidence estimation, integration of heterogeneous data modalities, and downstream reasoning and quality control, providing dynamic, scalable platforms for knowledge discovery and application.

1. Construction Methodologies and Temporal Evolution

Medical KGs are built through orchestrated multi-step pipelines involving entity and relation extraction, normalization, semantic integration, and continuous enrichment with new knowledge. Leading frameworks such as MedKGent simulate the real-time arrival of knowledge by organizing millions of biomedical abstracts into a fine-grained daily time series (1975–2023), matching the temporal dynamics of scientific discovery (Zhang et al., 17 Aug 2025). Each day's batch of documents undergoes entity extraction (e.g., Gene, Disease, Chemical, Variant, Species, CellLine) and normalization, often via domain tools like PubTator3, which provides unique identifiers, alias mapping, and semantic embeddings (e.g., BiomedBERT 768-dim vectors).

For relation extraction, transformer-based LLMs (e.g., Qwen2.5-32B-Instruct) are prompted to propose candidate triples of the form $(e_i, r, e_j)$, where %%%%1%%%% are entities in the abstract and $r$ is drawn from an ontology-defined set of relations. To address LLM stochasticity and provide relevance scores, sampling-based estimation is employed: triples are collected over $N$ LLM runs with controlled temperature, and a confidence score $c(t)$ is computed as the rounded-down frequency of occurrence across samples. Only triples with sufficient confidence (e.g., $c(t)\ge0.6$) are retained.

The Constructor Agent incrementally integrates new triples into a Neo4j-backed property graph, applying temporal reinforcement (monotonic confidence updates), evidence provenance (PubMed IDs, timestamps), and conflict resolution strategies (LLM-guided selection for relation conflicts). The resulting KGs can contain over 150,000 entities and several million relation triples, with high-confidence edges dominating due to repeated reinforcement as evidence accumulates (Zhang et al., 17 Aug 2025).

2. Ontology Design and Graph Schema

Schema design is central to robust KG construction. Entity types commonly include Gene, Disease, Chemical, Drug, Symptom, Phenotype, Procedure, CellLine, and regulatory entities. Relation types are defined both from biomedical ontologies and via open information extraction, covering interactions (e.g., “Associate,” “Negative_Correlate,” “Positive_Correlate,” “Treat,” “Inhibits”), mechanistic relations (e.g., “causes,” “expressed_in,” “participates_in”), and document-level provenance. Ontologies such as UMLS, SNOMED CT, OBO Foundry, Disease Ontology (DOID), and custom domain schemas (e.g., OncoNet Ontology for cancer biomarker discovery) establish relation domains/ranges and semantic constraints (Zhang et al., 17 Aug 2025, Karim et al., 2023).

KGs may incorporate hierarchical and synonym edges to support synonym resolution, cross-granularity mappings, and multi-level entity typing. Some frameworks extend conventional triple-based schemas to quadruples by appending a context field, enabling each assertion to carry minimal, standalone justification for downstream applications and improved explainability (Elliott et al., 5 Aug 2025).

3. Confidence Estimation, Temporal Reasoning, and Conflict Resolution

Ensuring factuality and coherence in dynamic KGs requires explicit uncertainty modeling and temporal reinforcement. Sampling-based confidence, such as $c(t) = \lfloor (f(t)/N) \cdot 20 \rfloor / 20$, quantifies extraction stability and filters noisy triples. Upon integration, repeated confirmations of the same relation incrementally increase edge confidence via a monotonic update rule: $s_\mathrm{new} = 1 - (1-s_\mathrm{old})(1-s')$ (Zhang et al., 17 Aug 2025). Temporal evidence is aggregated using the latest timestamp, enabling recency-aware queries and tracking of knowledge emergence.

Relation conflicts (multiple competing relations between two entities) are disambiguated by prompting LLMs with conflict context (previous confidence scores, evidence, timestamps) at low temperature for semantic consistency, ensuring robust resolution even as new findings accumulate over decades.

4. Automated and Manual Evaluation Protocols

Quality assurance of medical KGs relies on both automated and expert-driven evaluations. LLM-based rubric scoring, e.g., by evaluating extracted triples on a 0–3 scale, consistently yields validity rates above 85%, with inter-model agreement rates confirmed via precision, recall, F1, and Cohen’s $\kappa$ (Zhang et al., 17 Aug 2025). Expert manual review of sampled triples typically confirms these statistics, with >86% validity across multiple domain specialists.

Agreement between LLM and expert annotation is computed using standard metrics, revealing F1 scores around 95% and substantial agreement (Cohen’s $\kappa > 0.6$). These high-quality extraction and validation workflows are indispensable for downstream deployment in clinical or critical research settings.

5. Downstream Applications: Question Answering and Drug Repurposing

Medical KGs are foundational to diverse downstream tasks, enabling both retrieval-augmented generative reasoning and analytical inference over biomedical corpora. In retrieval-augmented generation (RAG), topologically relevant subgraphs (e.g., one-hop neighborhoods of query entities) are supplied to LLMs, with reranking and filtering guided by confidence, semantic similarity, and contextual alignment. RAG-based augmentation across multiple benchmarks (MMLU-Med, MedQA-US, MedDDx, BioASQ-Y/N) demonstrates consistent gains up to +8.6 percentage points over non-augmented baselines (Zhang et al., 17 Aug 2025).

In drug repurposing scenarios, KGs support literature-based causal inference by tracing two-hop or multi-hop patterns (e.g., Chemical—Negative_Correlate→Gene—Positive_Correlate→Disease) and aggregating path-level confidences:

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

Predicted drug–disease pairings, inferred solely from preexisting literature, were validated by independent later publications, illustrating robust prospective discovery with temporally aware, confidence-weighted reasoning (Zhang et al., 17 Aug 2025).

6. Limitations, Extensions, and Future Directions

While current medical KG frameworks provide significant advances in structuring, updating, and leveraging biomedical knowledge, important limitations persist. The restriction to text-based corpora omits complementary data modalities such as clinical trial registries and EHRs. Planned extensions target integration of additional modalities, the adoption of finer-grained uncertainty (e.g., Bayesian estimation), and ongoing replacement of LLMs with newer architectures to further reduce hallucination rates (Zhang et al., 17 Aug 2025).

Generalization to other scientific domains and corpora is straightforward within the agent-based, temporally synchronized extraction paradigm. Closed-loop feedback between KG evolution and downstream utility (QA performance, link prediction accuracy, drug discovery recall) will likely drive automated self-evaluation and continual improvement in future systems.

7. Significance and Impact

Temporally evolving medical knowledge graphs, constructed with robust LLM-powered agent frameworks, constitute the first large-scale resources to capture and track the dynamic emergence, reinforcement, and refinement of biomedical knowledge. The combination of systematic sampling-based confidence modeling, incremental integration with temporal and provenance tracking, and strong empirical results (∼90% extraction accuracy, demonstrable QA and drug-repurposing gains) underpins their practical impact for both research and clinical applications (Zhang et al., 17 Aug 2025). These advances position medical KGs as necessary infrastructure for evidence-based medicine, continuous learning healthcare systems, and automated reasoning across the biomedical discovery pipeline.