Traditional Chinese Medicine Knowledge Graph

• TCM-KG is a specialized, ontology-rich network that integrates classical texts, clinical records, and biomedical data to represent traditional Chinese medical knowledge.
• The construction involves systematic corpus engineering, advanced NER, dependency parsing, and multi-tiered schema alignment from diverse TCM sources.
• TCM-KGs enable clinical decision support, drug discovery, and research integration by leveraging graph embeddings, LLM-based reasoning, and neuro-symbolic safety mechanisms.

A Traditional Chinese Medicine–Specific Knowledge Graph (TCM-KG) constitutes a rigorously structured, ontology-rich, and multi-relational graph representation of the specialized entities and relationships characteristic of traditional Chinese medical theory and practice. Designed to codify the vast unstructured corpus of classical texts, clinical case records, pharmacologic resources, and emergent biomedical links, modern TCM-KGs provide computational infrastructure for information extraction, structured reasoning, semantic retrieval, clinical decision support, compatibility analysis, drug discovery, and the integration of TCM with contemporary biomedical knowledge. Across research, TCM-KG construction varies in scope—from mono-institutional disorder–symptom bipartite graphs to multi-modular, cross-disciplinary embeddings uniting herbal drugs, molecular targets, diseases, and patient states. Below, the encyclopedic coverage details state-of-the-art TCM-KG methodologies, schemas, embedding strategies, and applications, synthesizing developments from major arXiv contributions.

1. Corpus Engineering, Preprocessing, and Ontology Foundations

TCM-KG construction mandates systematic corpus preparation, entity normalization, and ontological curation. For domain-specific corpora, studies extract diverse sources: scanned classical texts (e.g., 68 gynecological books, 584 TCM PDFs), standardized clinical case records, curated symptom-disease tables, and formula databases (He et al., 28 Apr 2025, Li et al., 26 Feb 2026). Manual expert panel verification and OCR languages correction standardize entity mentions and remove extraneous matter, such as page headers or blank space (Zhao et al., 2024, He et al., 28 Apr 2025).

Entity extraction adheres to fine-tuned schema. For the Yi Shui school, the pipeline segments classical Chinese using BMES/IOB schemas, then applies Conditional Random Fields (CRF) for named entity recognition on herbs, meridians, symptoms, prescriptions, and organs (Zhao et al., 2024). Major ontological axes include:

Axis	Example Entities	References
Herb/HerbReference	“Danggui”, “甘草”	(He et al., 28 Apr 2025)
Disease	“肝虚证”, “hypertensive retinopathy”	(Zhang et al., 1 May 2025)
Symptom	“dizziness”, “wiry pulse”	(Li et al., 26 Feb 2026)
Syndrome/Pattern	“Qi deficiency”, “Blood stasis”	(Li et al., 26 Feb 2026)
Formula/Prescription	“Guizhi Decoction”	(Li et al., 26 Feb 2026)
Protein Target	“NFE2L2”, “PTGS2”	(Zhang et al., 1 May 2025)
Compound	“methyl ursolate”, “triptonide”	(Zhang et al., 1 May 2025)

Ontologies are harmonized with authoritative standards (WHO, ICD-11, MeSH, DOID, GB/T 12346-2006) and custom dictionaries for string normalization and synonym unification (Zeng et al., 2024, Xu et al., 9 Mar 2026). Coreference is resolved via string clustering or embedding comparison (Zhao et al., 2024).

2. Entity and Relation Extraction, Feature Engineering

Relationship extraction leverages dependency parsers, LLM-augmented information retrieval, and manual curation. A typical pipeline (Zhao et al., 2024):

1. CRF NER: Recognize entity spans using contextual, orthographic, and POS-based features; L-BFGS optimization stabilizes learning.
2. TF-IDF weighting: Identify “key entities” per document based on term frequency and inverse document statistics.
3. Dependency Parsing: A neural parser (e.g., HanLP/Chen & Manning) assigns labeled dependency arcs (e.g., nsubj, dobj, amod, prep_of).

Entity–relation triples (head, relation, tail)—e.g., (Herb:黄柏, prep_of.downstream, Organ:小肠)—form the factual backbone. Schema definitions range in granularity; policy- and safety-focused KGs incorporate “ProhibitedFor” and “IncompatibleWith” edges for acupuncture safety (Xu et al., 9 Mar 2026).

Multi-tiered KG schemas extend from macro book–chapter–section hierarchies to micro-level (entity, relation, entity) triples extracted per document snippet, aligning to high-level ontologies for synonym consolidation (Li et al., 26 Feb 2026, He et al., 28 Apr 2025).

3. Graph Construction, Embedding, and Mathematical Architecture

TCM-KGs are formalized as heterogeneous, typed, and often weighted property graphs:

• $G = (V, E, A)$, with $V$ partitioned among entity types and $E = \{(u, v, \tau, w)\}$ representing edges (relation type $\tau$, weight $w$) (Li et al., 15 Jul 2025).
• Node attributes and embeddings $a_n: V \to \mathbb{R}^{d_n}$, edge attributes $a_e: E \to \mathbb{R}^{d_e}$.
• Edge weights assigned via normalized Pointwise Mutual Information (nPMI), Tanimoto similarity, or learned attention.

Graph representation learning integrates node2vec, Word2Vec, PCA, and heterogeneous GNNs—most notably:

• Node2vec and DGI extract and refine node embeddings via random walks and mutual information maximization (Zhang et al., 1 May 2025, Zeng et al., 2024).
• Message passing: GAT, GTN, and Hypergraph Neural Network propagate representations over complex, multi-relational structures with attention and global context (Zeng et al., 2024).
• Transformer-based autoencoders map symptoms and herbs into a unified TCM embedding space (TCM-ES), training via autoencoding and an InfoNCE contrastive loss. Clinical, biomedical, and pharmaceutical entities are projected into this space for KNN-based link prediction (Li et al., 15 Jul 2025).

Graph construction and similarity scoring support both knowledge completion and retrieval. For example, in the DoPI system, the graph is a weighted, in-memory symptom–disease bipartite with disease scoring derived via disease–symptom binary vectors and cosine similarity (Sun et al., 7 Jul 2025).

4. Integration with Clinical Reasoning, Retrieval, and Downstream Systems

TCM-KGs operationalize clinical reasoning via integration with LLMs, RAG, or neuro-symbolic pipelines. Notable strategies:

• Retrieval-Augmented Generation (GraphRAG): Clinical queries are decomposed into entity mentions by NER, embedded, and retrieved in the KG via cosine similarity and K-hop expansion. Retrieved subgraphs are serialized into LLM prompts for contextualized question answering (He et al., 28 Apr 2025).
• Chain-of-Thought Reasoning: TCM-DiffRAG parses queries into sequences of reasoning triples, retrieves matched KG facts, fuses them into enriched LLM prompts, and orchestrates attention-based knowledge infusion for explicit, auditable clinical decision paths (Li et al., 26 Feb 2026).
• Safety Enforcement: CORE-Acu encodes safety rules (e.g., acupuncture point contraindications) as KG edges and applies a Symbolic Veto Mechanism in a generate–verify–revise loop. The Lexicon-Matched Entity-Reweighted Loss (LMERL) increases learning signal for high-risk concepts (Xu et al., 9 Mar 2026). This system achieved SVR=0/1,000 for explicit KG rules, far surpassing unconstrained LLMs.

In recommendation and diagnosis, KGs enable sequential patient modeling—embedding patient history, symptom trajectories, and herb compatibility via attention-based GNNs and RNNs for personalized prescription generation (Liu et al., 2023).

5. Quantitative Validation, Case Studies, and Error Analysis

TCM-KG systems have undergone multi-faceted, large-scale evaluation:

• NER/Relation Extraction: Pilot study for Yi Shui NER yields $F_1 \approx 92.7\%$, relation extraction $F_1 \approx 88.6\%$ (Zhao et al., 2024). Customized prompts with Chinese-oriented LLMs (e.g., Kimi, DeepSeek) delivered extraction $F_1$ up to $V$0 in OpenTCM (He et al., 28 Apr 2025).
• Graph-Based Inference: Node2Vec-DGI-EL achieved AUC = 0.9987, AUPR = 0.9545 for ingredient–disease association, with ablation studies confirming the necessity of each module (Zhang et al., 1 May 2025).
• Safety and Interpretability: CORE-Acu’s GVR reduced SVR to zero in 1,000 held-out acupuncture planning cases; LLM baselines remained at 2-8.5% (Xu et al., 9 Mar 2026).
• Clinical correlation: TCM-ES distances are negatively correlated with symptom/herb co-occurrence and improvement in telemedicine cohorts. Drug efficacy metrics (bi-directional z-score) correlate with FDA outcome scores for known drug–disease pairs (Li et al., 15 Jul 2025).
• Downstream QA and retrieval: OpenTCM’s GraphRAG pushed mean expert scores to MES=4.5 in herb retrieval, surpassing earlier systems (He et al., 28 Apr 2025).

Major error sources across studies were polysemy (e.g., “经”), ellipsis, limited annotation for rare herbs, and (in LLM contexts) entity hallucination outside KG constraints.

6. Applications: Clinical Decision Support, Drug Discovery, and Beyond

TCM-KGs are operationalized in a diversity of platforms:

• Clinical Diagnosis: DoPI leverages a weighted symptom–disease KG for proactive, multi-turn interrogation and diagnosis, with ongoing expert KNN-driven KG refinement (Sun et al., 7 Jul 2025).
• Syndrome Differentiation: TCM-DiffRAG exploits macro/micro KGs and LLM+CoT fusion, enabling individualized, context-sensitive diagnosis and explicit, interpretable reasoning chains (Li et al., 26 Feb 2026).
• Drug/Ingredient Discovery: Graph-based models (Node2Vec-DGI-EL, GAT) predict novel ingredient–disease links, validated via molecular docking (e.g., triptonide–PGR, methyl ursolate–NFE2L2) and network pharmacology (Zeng et al., 2024, Zhang et al., 1 May 2025).
• Safety Assurance: Neuro-symbolic vetting with explicit veto rules, full audit trails, and entity-weighted loss minimization ensures zero-risk clinical recommendation in high-risk settings, notably in acupuncture (Xu et al., 9 Mar 2026).
• Multi-Modal Integration: KG architectures unify textual, molecular, and biomedical data, supporting drug repurposing (as in TCM-ES proximity-based target prediction for rheumatoid arthritis), and precision subtyping (e.g., RA Heat vs. Cold embeddings) (Li et al., 15 Jul 2025).

Open-source resources for TCM-KGs, including data, code, and graph indexing/embedding tools, are available for independent validation and extensibility (Zeng et al., 2024).

7. Methodological Innovations, Challenges, and Future Directions

TCM-KGs embody multiple technical advancements:

• Integration of deep NER, dependency parsing, and TF-IDF filtration for robust entity–relation extraction from raw, ambiguous classical texts (Zhao et al., 2024).
• Multi-stage, attention-driven heterogeneous GNN architectures and Transformer-based embedding spaces for knowledge representation (Zeng et al., 2024, Li et al., 15 Jul 2025).
• KNN edge formation in continuous embedding space, aligning molecular and clinical modalities.
• Dual-ontology structures: global “macro” (document layout, book, section) and local “micro” (entity–relation triple), unified via mapping dictionaries for traceable inference (Li et al., 26 Feb 2026).
• Safety-centric neuro-symbolic reasoning with deterministic veto subgraphs and lexicon-reweighted optimization (Xu et al., 9 Mar 2026).

Persistent challenges include polysemy, data scarcity for rare herbs or symptoms, shifting professional standards, and bridging differences between traditional and modern biomedical ontologies. Future work is oriented toward domain-adaptive representation learning (ELMo/BERT for TCM NER), large-scale manual annotation, integration of radical/phonetic features in low-resource settings, multimodal fusion (EHR, images), and longitudinal patient modeling.

TCM-KGs thus serve as a computational substrate for unifying millennia-old clinical experience, modern biomedicine, and advanced AI-driven reasoning into an extensible platform for both research and clinical innovation (Li et al., 15 Jul 2025, Zeng et al., 2024, Li et al., 26 Feb 2026, He et al., 28 Apr 2025, Zhao et al., 2024, Xu et al., 9 Mar 2026, Zhang et al., 1 May 2025, Sun et al., 7 Jul 2025, Liu et al., 2023).