acceptodds
Under review as a conference paper at ICLR 2027

KnowAcu: A Knowledge-Guided Heterogeneous Multi-Graph Learning Framework for Acupoint Prescription

Abstract

As an indispensable component of Traditional Chinese Medicine (TCM), acupuncture therapy plays a pivotal role in both clinical treatment and preventive healthcare. However, generating appropriate acupoint prescriptions—structured treatment schemes tailored to complex symptoms remains a formidable challenge in automated medical reasoning. Existing studies struggle to bridge the semantic knowledge gap between diverse symptom manifestations and rigorous TCM laws, often failing to capture the inherent structural heterogeneity of symptom-acupoint interactions. In this paper, we propose KnowAcu, a Knowledge-Guided framework for precise Acupoint prescription that synergizes TCM-driven heterogeneous multi-graph structures with the semantic expertise of Large Language Models (LLMs). Specifically, we construct five distinct graphs to explicitly model the multi-faceted relational dependencies defined by TCM principles and historical prescriptions. These structures are integrated through a Multi-Graph Joint Learning (MGJL) module, which extracts multi-granular representations of symptoms and acupoints. To align the model with professional clinical logic, we leverage LLMs for auxiliary supervision and knowledge distillation, incorporate the functional hierarchies of acupoints and employ a acupoint-symptom reconstruction task for semantic consistency verification. Extensive experiments on a curated real-world clinical dataset and a large-scale simulated benchmark demonstrate that KnowAcu significantly outperforms state-of-the-art methods in both clinical applicability and model robustness.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Acupoint Recommendation, Graph Neural Network, Knowledge Distillation, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3c5POpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes KnowAcu, a graph-based model for recommending acupuncture points from patient symptoms. It combines five graphs built from prescription co-occurrence patterns and traditional Chinese medicine knowledge about organs and meridians. A gated graph attention network learns symptom and acupoint representations, which are used to rank candidate acupoints. The model also uses LLM-generated supervision to distinguish primary from supporting acupoints and encourage consistency between the recommendations and input symptoms. Experiments on a small dataset drawn from clinical records and literature, along with a larger synthetic dataset, show improvements over several recommendation and LLM baselines.

Critical strengths

Organ and meridian relationships add useful domain knowledge beyond the patterns observed in prescriptions.

Reconstructing symptoms from the recommended acupoints is a useful consistency check, although the paper needs to explain its implementation more clearly.

The ablation results show lower performance when the theory graphs, joint graph learning, or LLM losses are removed, suggesting that each component contributes to the results.

Critical weaknesses

The real dataset has only 546 prescriptions but 2,174 distinct symptoms. Given the number of components in the model, overfitting is a concern. The authors use five-fold cross-validation, but without standard deviations or confidence intervals, it is hard to tell how consistent the gains are.

The LLM scores the importance of acupoints and links them to symptoms, but the paper does not show whether experts checked these judgments systematically. If these associations are incorrect, the reconstruction loss could push the model to learn them anyway.

Symptoms and acupoints are connected whenever they occur in the same prescription. This treats frequent and occasional co-occurrences alike. It also ignores that a relationship may be useful for one clinical presentation but not another.

LINKO, “Multi-Ontology Integration with Dual-Axis Propagation for Medical Concept Representation” (CIKM 2025), has a similar motivation: combining medical knowledge with co-occurrence patterns through graph learning. It would be useful to discuss this work and consider it as a baseline. Since LINKO uses hierarchical ontologies, the authors would need to explain how it could be adapted for a fair comparison.

Questions for authors

sea weaknesses

Reviewer_Z5D9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes KnowAcu, a knowledge-guided framework for acupoint recommendation. It builds five relational graphs from historical prescriptions, zang-fu theory, and meridian theory, and integrates them through graph filtering, gated attention, and layer-wise representation fusion. Training is further guided by LLM-generated acupoint importance scores and an auxiliary symptom reconstruction loss. On a real dataset of 546 prescriptions and a synthetic dataset of 26,360 prescriptions, KnowAcu outperforms several recommendation methods and LLM baselines.

Critical strengths
  1. The framework integrates prescription co-occurrence with zang-fu and meridian relations, offering a principled way to bring domain knowledge into symptom-acupoint modeling.

  2. The gains on the real dataset are substantial, with P@5 rising from 0.53 for PEGCN to 0.69 for KnowAcu, which suggests clear promise for this application.

Critical weaknesses
  1. The "w/o LLM loss" ablation removes all task losses in Eq. (23) without specifying a replacement objective, which conflates the benefit of LLM guidance with that of basic task supervision.

  2. The paper does not state whether the co-occurrence graphs are built only from the training prescriptions in each fold, nor which data are used for hyperparameter selection.

  3. The synthetic benchmark is circular, since both its target prescriptions and KnowAcu's auxiliary supervision come from LLMs, which may favor the proposed method. The main text also attributes the synthetic data to Sunsimiao, whereas the appendix names Qwen3.

Questions for authors
  1. Are the co-occurrence graphs rebuilt from the training prescriptions only in each fold, and which split is used for hyperparameter selection?

  2. What objective is used in the "w/o LLM loss" ablation? Could the authors add a variant that keeps a standard recommendation loss and removes only the LLM-derived scores and the symptom reconstruction loss?

Reviewer_47m8OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces a multi-graph learning framework for recommending acupuncture prescriptions from patient symptom sets. The authors construct five relational graphs combining empirical prescription co-occurrences with Traditional Chinese Medicine (TCM) domain priors, specifically zang-fu organ associations for symptoms and meridian pathways for acupoints. These graphs are merged and pruned via topological cosine similarity, encoded through a gated multi-layer GAT, and passed to a transformer module for acupoint prediction. Experiments show the effectiveness of the proposed method.

Critical strengths
  1. Integrating empirical co-occurrence graphs with zang-fu organ and meridian graphs provides a sensible inductive bias for sparse clinical data, where pure symptom-acupoint co-occurrence is often noisy or incomplete. Ablations on the real dataset confirm that both theoretical graphs and multi-graph joint learning contribute to retrieval and ranking accuracy.

  2. Using an LLM offline to differentiate primary from auxiliary acupoints during training is a practical way to inject clinical hierarchy into a compact graph model without incurring LLM inference latency at test time.

Critical weaknesses
  1. In Section 3.4, the counter-prediction loss requires passing predicted acupoints through an LLM during training, which is non-differentiable and blocks gradient back-propagation unless an unexplained approximation is used. Additionally, the positive-negative ranking loss leaves the negative acupoint index undefined, and the "w/o LLM loss" ablation removes all three training losses without explaining what objective actually trains the network.

  2. Section 4.1 claims the synthetic dataset was generated by Sunsimiao, whereas Appendix H states it was generated by Qwen3. Because Qwen3 also generates the training supervision for KnowAcu, evaluating on Qwen3-generated data measures how well the student mimics its teacher rather than true clinical generalization.

  3. The paper does not clarify whether the co-occurrence graphs and similarity pruning are computed strictly on the training split within each cross-validation fold or across the entire dataset beforehand. Furthermore, the prompt in Table 9 mistakenly copies instructions from Table 8 (asking for symptom json outputs instead of acupoint scores), and Appendix F.2 cites an obstetrics fetal MRI study ("The MERIDIAN trial") as a TCM meridian reference.

Questions for authors
  1. How are gradients computed for the counter-prediction loss during training? Specifically, how is the continuous predicted acupoint probability vector combined with the offline LLM outputs from Table 8 to produce a differentiable symptom prediction?

  2. In Eq. 20, how are negative acupoints selected or summed over? Additionally, when all three loss terms are removed in the "w/o LLM loss" and "w/o all components" rows of Table 5, what training loss is used?

  3. Was the synthetic dataset generated by Sunsimiao (as stated in Section 4.1) or Qwen3 (as stated in Appendix H)?

  4. Are the prescription co-occurrence matrices (symptom-symptom, acupoint-acupoint, and symptom-acupoint) built strictly from the training fold in each of the 5 cross-validation splits, and how are unseen test symptoms handled when 2174 unique symptoms are distributed across only 546 prescriptions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.