JMIR Publications

ABSTRACT

Background:

Mental stress-induced hypertension (MSI-HBP) was recognized as a clinically distinct hypertension subtype in the 2021 expert consensus and revised post-pandemic in 2022, with 64% of adults reporting stress as a primary health concern. Traditional Chinese medicine (TCM) provides a syndrome-differentiation framework mapping psychosocial stressors to characteristic pathogeneses (Liver Qi Stagnation, Liver Yang Hyperactivity, Phlegm-Dampness Obstruction, Heart-Kidney Disharmony) and herbal interventions; however, this knowledge remains fragmented across hundreds of contemporary peer-reviewed publications. Existing TCM knowledge graphs are either large-scale general-purpose systems lacking MSI-HBP focus, restricted to ancient texts, or do not support multi-relation natural language querying.

Objective:

To construct an open-source domain-specific TCM knowledge graph for MSI-HBP from contemporary peer-reviewed literature, supporting natural language question answering for clinical decision-support, with a fully documented and replicable construction pipeline.

Methods:

From 419 full-text peer-reviewed publications (2015–2025) retrieved from PubMed, CNKI and Wanfang, we constructed MSI-HBP-KG using Qwen2.5-7B-Instruct—an open-source 7-billion-parameter Chinese-capable LLM—accessed via the SiliconFlow inference platform with temperature set to 0.1 for output stability. Triple extraction followed three structured prompts (entity recognition, JSON triple extraction, and answer synthesis), with outputs processed through a four-stage data cleaning pipeline: (1) synonym normalization using a curated 200-entry synonym dictionary; (2) entity uniqueness constraints by (normalized_name, type) tuples; (3) supplementation of entities referenced in relations but missing from the entity list; and (4) triple deduplication by (subject, relation, object) tuples. The knowledge graph was deployed on Neo4j 5.x with a hybrid retrieval-and-LLM-synthesis question-answering engine. An internal 150-query evaluation across the four core TCM relation types (TREATS, COMPOSED_OF, USES, CAUSES) assessed system accuracy.

Results:

The final knowledge graph comprised 1,033 entity nodes across 25 entity types and 312 directed relationships. The six core TCM clinical entity types—those that participate in the four core relations (TREATS, COMPOSED_OF, CAUSES, USES)—were Treatment Method (159 nodes, 15.4%), Formula (122 nodes, 11.8%), Symptom (108 nodes, 10.5%), Disease (90 nodes, 8.7%), Herb (66 nodes, 6.4%) and Pathogenesis (53 nodes, 5.1%), together comprising 598 nodes (57.9%). The remaining 435 nodes spanned 19 secondary categories, predominantly measurement tools (159), clinical scales (81), Western pharmaceutical agents (45) and TCM pattern types (39). Relations distributed as TREATS 171 (54.8%), COMPOSED_OF 84 (26.9%), CAUSES 32 (10.3%) and USES 25 (8.0%). Internal question-answering accuracy was 92.67% (139/150) overall; per-relation accuracy ranged from 90.4% to 100%: CAUSES 100% (15/15), COMPOSED_OF 95.0% (38/40), USES 91.7% (11/12), and TREATS 90.4% (75/83), with all four core relation types thus achieving ≥ 90% accuracy. External validation on a stratified sample of 100 triples by two independent TCM clinicians yielded 91.0% observed inter-rater agreement (Gwet AC1 = 0.90, 95% CI 0.84–0.96) and consensus-resolved triple precision of 96.0% (95% Wilson CI 90.2–98.4; n = 96/100).

Conclusions:

MSI-HBP-KG is the first open-source TCM knowledge graph dedicated to mental stress-induced hypertension, constructed from contemporary peer-reviewed literature. Its construction pipeline—built on a small open-source 7-billion-parameter LLM with explicit synonym normalization, entity uniqueness enforcement, and missing-entity supplementation—supports reproducibility on consumer-grade hardware without commercial API dependency. The system is positioned as a clinician-facing literature-retrieval and decision-support aid rather than an autonomous prescribing tool. The codebase, knowledge base, prompt templates and deployment guide are openly available15 under the MIT License.