A home triage model for TCM gynecology: MAGR-TCM separates emergencies from routine cases in all 30 safety-test records

Home-based TCM gynecology consultation can now return advice along a traceable reasoning chain rather than a single generated passage. The system maps a user's colloquial description onto symptom nodes in the knowledge graph and walks hop by hop to pattern, therapeutic principle, formula and medicinal, so each recommendation points back to case evidence. General-purpose large language models answering directly tend to produce medical content without reliable grounding, while vector-similarity Naive RAG, summary-oriented GraphRAG and lightweight LightRAG all struggle to reproduce the full disease-pattern-principle-formula-medicinal derivation consistently. On 30 test cases disjoint from the training set, RAGAS faithfulness is 0.821 for MAGR-TCM, 0.766 for GraphRAG, 0.651 for LightRAG and 0.590 for Naive RAG; the lower bound of MAGR-TCM's faithfulness interval, 0.814, exceeds the upper bounds of all three baselines.

A consultation is split across four agents in sequence: a question-analysis agent handles privacy filtering, emergency grading, typo correction and entity extraction; a routing and triage agent chooses emergency blocking, web search or graph retrieval; a graph-reasoning agent searches four hops; and an evaluation agent scores the answer and triggers dual-model rewriting when it falls short. A single model generating in one pass is prone to decision errors from long contexts and dispersed task focus; dividing the task among dedicated agents gives each step a defined input and a checkable output. In double-blind expert scoring, MAGR-TCM reaches 0.898 for pattern identification accuracy, 0.884 for consistency of principle, method, formula and medicinal, and 0.889 for clinical utility, above LightRAG's 0.734, 0.700 and 0.732; all 30 paired comparisons remain significant after Holm-Bonferroni correction.

Safety triage is a separate hard-wired path: the question-analysis agent first scans for red-flag terms such as shock, massive hemorrhage and severe pain, and a hit terminates further reasoning and returns offline emergency guidance instead of a formula recommendation. The three baselines have no dedicated emergency detection module, so emergency cases are handled as routine TCM consultations or depend on the underlying model's unpredictable refusals. On the 30-case safety stress set of 10 emergency and 20 routine cases, MAGR-TCM records sensitivity 1.000, specificity 1.000 and false negative rate 0.000.

Evaluation runs on expert-annotated retrospective records: 30 routine test cases, 30 safety stress-test cases and the 741 clinical cases used to build the graph, all from the Ancient and Modern Medical Case Cloud Platform. The system targets home-user TCM gynecology consultation and preliminary triage; four-hop graph reasoning covers all five entity types on the 30-case test set, and a fifth hop adds no coverage.

The 741-case dataset may limit generalization to rare patterns or atypical symptom combinations; the graph is text-based and does not yet integrate modern indicators such as ultrasound or serology; training on consensus logic and standard textbooks may not capture the individualized modifications of different TCM schools. All metrics are upper-bound estimates from retrospective evaluation, the system has not been prospectively validated with real home users, and the red-flag mechanism has not been formally validated against clinical triage standards such as the Emergency Severity Index.