Graph Neural Network Models Symptom Interactions for Intelligent Acupoint Recommendation

A computational framework combining graph neural networks, contextual transformer representations, and hierarchical attention mechanisms has achieved improved predictive performance in modeling symptom-to-acupoint compatibility for functional gastrointestinal disorders, according to an evaluation published in the journal Chinese Medicine. Formulating acupuncture point selection as a multi-label text classification problem, researchers introduced an architecture designated as GNN-BERT-Attention to map complex multi-symptom presentations directly into clinically coherent sets of therapeutic acupoints. Functional gastrointestinal disorders, such as irritable bowel syndrome, functional dyspepsia, and functional constipation, represent chronic and recurrent conditions characterized by abdominal discomfort, nausea, bloating, and altered bowel habits without identifiable organic pathology. While clinical trials indicate that traditional acupuncture can alleviate these gastrointestinal symptoms and improve patient quality of life, real-world clinical implementation faces a lack of standardized treatment protocols because individual acupoints possess multiple therapeutic functions and clinical prescriptions routinely require synergistic point combinations tailored to multifaceted syndrome patterns.

Computational recommendation systems for traditional Chinese medicine have previously struggled to capture the intricate compatibility mechanisms inherent in historical and contemporary clinical practice. Earlier algorithmic strategies, including rule-based expert systems, static knowledge graphs such as AcuKG, and basic machine learning classifiers, frequently failed to process the rich semantic variations present in unstructured symptom descriptions or capture higher-order topological interactions connecting diverse acupoint combinations. Conventional natural language processing models frequently handle multi-label outputs as independent targets, overlooking synergistic clinical pairing principles such as classic point pairs, front-mu and back-shu combinations, or local-distant associations. To bridge the divide between computational modeling and clinical diagnostic logic, the research team designed an integrated framework that addresses extreme data sparsity, balances core primary points with supplementary auxiliary points, and accounts for severe class imbalances caused by the long-tailed frequency distribution of specialized acupuncture points across patient records.

To train and evaluate the system, investigators compiled a specialized dataset derived from 1,086 historical medical excerpts extracted from 68 classical acupuncture treatises, including the Miraculous Pivot and the A-B Classic of Acupuncture and Moxibustion, preserved in digital archives and institutional collections at the Chengdu University of Traditional Chinese Medicine. Guided by Rome IV diagnostic criteria and recognized traditional Chinese medicine internal guidelines, the investigators identified eight foundational symptom categories associated with functional gastrointestinal disorders. Following rigorous deduplication and filtering of incomplete records, the team curated a standardized corpus consisting of 566 distinct symptom-acupoint pairs encompassing 154 specific symptom descriptors and 89 unique acupoints, such as Zusanli (ST36), Zhongwan (CV12), and Sanyinjiao (SP6). Symptom terminologies were normalized into uniform anatomical and sensory formats in compliance with World Health Organization international standard terminologies and national standard symptom guidelines, establishing a structured baseline for computational modeling.

The feature extraction pipeline begins at the input encoding layer, where raw symptom descriptions are converted into structured representations using the BERT-base-uncased tokenizer and bidirectional transformer encoder. The symptom text sequences are padded to a standardized length of 128 tokens, and context-aware hidden state vectors of 768 dimensions are generated alongside attention mask matrices. To extract salient diagnostic signals, the hidden representations are processed through a custom attention mechanism comprising learnable weight matrices, projection vectors, and hyperbolic tangent non-linear transformations. This attention layer dynamically calculates normalized scalar weights across sequence tokens, focusing computational attention on diagnostically critical terms, such as abdominal distension or dull gastric pain, while filtering out non-specific linguistic noise. The resulting context vector synthesizes the token-level embeddings into an aggregated 768-dimensional symptom representation that reflects the diagnostic emphasis that traditional clinicians place on dominant complaints during clinical intake.

Relational reasoning is executed in the subsequent graph neural network feature interaction layer, where structural dependencies among acupoint targets are formalized as an adjacency matrix. Within this clinically informed network, acupoints serve as nodes initialized with 256-dimensional random embedding vectors, while edges represent verified synergistic pairings, anatomical adjacency, and shared meridian pathways documented in clinical literature. The global 768-dimensional symptom vector derived from the attention layer is linearly projected into a 256-dimensional subspace and broadcast across all target acupoint nodes. By concatenating the projected symptom representation with the initial node features, the framework constructs a 512-dimensional composite feature matrix. A graph convolutional transformation then performs neighborhood feature aggregation and propagation across the graph topology using normalized degree matrices, learnable weight tensors, and rectified linear unit activations, generating updated 256-dimensional node representations that capture latent relational affinities.

The multi-label prediction layer maps the updated acupoint node embeddings into independent binary decision outputs, mirroring the traditional acupuncture prescription structure of primary core points supported by auxiliary secondary points. Unnormalized logit values are computed for each acupoint node via dedicated fully connected layers and transformed into continuous activation probabilities through individual sigmoid functions without constraining their cumulative sum to unity. To resolve severe label sparsity and frequency imbalances across the 89 potential targets, the researchers implemented a dynamic composite loss function incorporating Focal Loss alongside standard binary cross-entropy. By introducing a modulating focusing parameter and a balancing weight coefficient, Focal Loss penalizes trivial negative background samples and concentrates gradient updates on challenging, hard-to-classify minority acupoint classes. Hyperparameter grid searches established optimal performance on validation splits when setting the balance coefficient alpha to 1.0 and the focusing factor gamma to 2.0.

Empirical evaluations conducted using an eight-to-two train-test split under fixed random seeds demonstrated that the GNN-BERT-Attention framework outperformed conventional machine learning techniques and standard deep learning baselines across key ranking and precision metrics. In evaluating average precision at K, the proposed model maintained exceptionally stable performance, scoring 0.743 at K = 1, 0.757 at K = 3, and 0.744 at K = 5. While baseline architectures such as convolutional neural networks, recurrent neural networks, and term frequency-inverse document frequency support vector machines exhibited steep declines in precision as the retrieval threshold expanded, the hybrid graph model preserved its predictive stability. Ranking evaluations utilizing normalized discounted cumulative gain confirmed the framework’s effectiveness, yielding scores of 0.688 at K = 1 and 0.740 at K = 3, demonstrating that the combination of semantic token encoding and topological relational modeling consistently prioritized clinically viable point combinations.

Comprehensive multi-label evaluation metrics detailed in the study further highlighted the advantages of the integrated design. The GNN-BERT-Attention architecture achieved an overall label average accuracy of 0.960, a micro-averaged F1 score of 0.603, a sample-averaged F1 score of 0.559, and an overall precision of 0.799. By comparison, traditional methods such as rule-based reasoning, co-occurrence frequency heuristics, and linear support vector classifiers achieved substantially lower F1 scores. Receiver operating characteristic analyses revealed steady discriminative gains over iterative training cycles, with the mean area under the curve rising from approximately 0.77 at epoch 1 to roughly 0.90 by epoch 20. Distributional calibration assessments via information-theoretic metrics indicated that the model achieved a test cross-entropy of 3.543 and a Kullback-Leibler divergence of 0.076, demonstrating close mathematical alignment between the predicted output distribution and empirical ground-truth label distributions.

Ablation experiments confirmed that each structural component of the pipeline contributed significantly to overall predictive capability and stability. When the graph neural network interaction layer was systematically replaced with non-graph sequence architectures, such as multi-layer perceptrons or long short-term memory networks, label average accuracy plummeted from 0.960 to 0.062, accompanied by substantial drops in micro-F1 and precision. Similarly, substituting the dynamic Focal Loss with conventional binary cross-entropy reduced label accuracy to 0.412 and micro-F1 to 0.464, while stripping the input pipeline of its random token deletion noise augmentation resulted in a sharp drop in label accuracy to 0.050 and micro-F1 to 0.292. These ablation findings verified that explicit graph topology, noise-resilient data augmentation, and imbalance-aware loss functions operate synergistically to prevent overfitting on sparse medical corpora and maintain robust multi-label classification accuracy.

To evaluate practical clinical utility, the research team deployed the trained PyTorch model within an interactive web application built on the Flask microframework. Hosted on an open-access demonstration server, the web platform processes user-selected gastrointestinal symptom combinations, generates instant real-time acupoint inference, and presents prioritized recommendations with attention-weighted rationales. Despite these promising operational capabilities, the authors noted several methodological limitations, emphasizing that the system is intended as an auxiliary clinical decision-support tool rather than an autonomous diagnostic replacement. The initial training corpus was restricted to 566 classical excerpts, and external multi-center clinical validations remain necessary to confirm generalizability across modern electronic health records. Furthermore, the model’s static relational graph does not yet capture dynamic temporal dimensions, such as five-phase cycles or holistic constitution tracking, pointing toward future opportunities in knowledge graph expansion and causal reasoning.

Subject of Research: Multi-label deep learning framework combining graph neural networks and language models for clinical acupoint recommendation

Article Title: Modeling symptom-acupoint interactions via a heterogeneous graph learning framework for intelligent acupoint recommendation

Article References: Lin, W., Chen, J., Chen, Z., Yin, T., Guo, J., & Zeng, F. (2026). Modeling symptom-acupoint interactions via a heterogeneous graph learning framework for intelligent acupoint recommendation. Chinese Medicine, 21(1), Article 195. https://doi.org/10.1186/s13020-026-01458-1

Image Credits: AI Generated

DOI: 10.1186/s13020-026-01458-1

Keywords: Graph Neural Networks, Natural Language Processing, Multi-Label Text Classification, Acupuncture Point Recommendation, Functional Gastrointestinal Disorders, BERT Language Model, Attention Mechanism, Heterogeneous Graph Learning, Focal Loss, Traditional Chinese Medicine, Acupoint Compatibility, Data Augmentation, Clinical Decision Support, Model Interpretability, Label Co-occurrence, Biomedical Informatics