Mining Symptom-Herb Patterns from Patient Records Using Tripartite Graph

Jinpeng Chen, Josiah Poon, Simon Poon, Ling Xu, Daniel SzeEvidence-based 2015 · DOI

TL;DR

This paper introduces a novel three-step approach to mine symptom-herb relationships in Traditional Chinese Medicine (TCM) by constructing a heterogeneous tripartite graph. By leveraging the crucial intermediate step of "syndrome differentiation" in TCM, their method significantly improves the accuracy of predicting effective herb combinations compared to direct symptom-herb mappings.

What Problem Does This Paper Address?

Traditional Western medicine often treats specific symptoms with targeted drugs. In contrast, Traditional Chinese Medicine (TCM) employs a unique process called "syndrome differentiation" (SD). This involves classifying a patient's condition into a specific syndrome before selecting the most appropriate herbal formula. This SD step is crucial for personalized treatment but makes directly mapping symptoms to herbs challenging. The paper aims to bridge this gap by developing a data-driven method to uncover these complex symptom-herb relationships, acknowledging that these relationships are often mediated by syndromes. This is important for improving the efficacy of TCM prescriptions and for understanding the underlying principles of TCM.

Methods

The authors propose a three-step prediction approach (Tri-TSPA) built upon a heterogeneous tripartite information network.

  1. Network Construction: They first construct a tripartite graph representing relationships between Herbs (H), Syndromes (D), and Symptoms (P). This network captures richer information than simpler, homogeneous networks.
  2. Path Extraction: Two types of paths are extracted from this tripartite graph:
  3. PH_Path (Symptom → Herb): Represents a direct symptom-herb relationship, akin to Western medicine.
  4. PDH_Path (Symptom → Syndrome → Herb): Represents the indirect relationship, reflecting the TCM approach where syndromes mediate the symptom-herb connection.
  5. Matrix Construction and Similarity Measurement: Based on these paths, matrices are built to represent co-occurrence relationships. Several measurement strategies (Frequency, Jaccard Coefficient, Asymmetric Measure, TfIdf) are used to quantify these relationships. Similarity measures (simPH_Path and simPDH_Path) are then defined to quantify the similarity between entities of the same type, incorporating both direct and indirect relationships.
  6. Relevance Function and Prediction: A relevance function is introduced to filter out patterns associated with "bad outcomes" (ineffective treatments). Finally, an unsupervised prediction method (Tri-TSPA) is employed to discover symptom-herb patterns by identifying nearest entities based on similarity measures and then ranking potential herb predictions.

Datasets

The experiments were conducted on four real-world TCM datasets provided by Guang'anmen Hospital, China Academy of Chinese Medical Sciences:

  • Insomnia
  • Infertility
  • Diabetes
  • Tourette Syndrome

These datasets contain information on symptoms, syndromes, and prescription details of outpatients. The paper does not introduce new benchmark datasets but utilizes existing clinical data.

Key Insights and Ideas

  • Syndrome Differentiation is Key: The most significant insight is that incorporating the "syndrome differentiation" step (Symptom → Syndrome → Herb) leads to more accurate and effective symptom-herb pattern discovery than direct Symptom → Herb mapping. This validates a core principle of TCM in a data-driven manner.
  • Heterogeneous Networks Capture Richer Information: Using a tripartite graph (symptoms, syndromes, herbs) provides a more comprehensive view than homogeneous networks (e.g., just herb-herb interactions).
  • Power Law Distributions are Prevalent: The frequency of herbs, symptoms, syndromes, and their co-occurrence patterns all follow power-law distributions, indicating a few highly frequent entities and many rare ones. This is a common observation in complex systems.
  • TfIdf Outperforms Other Measures: The TfIdf (Term Frequency-Inverse Document Frequency) measurement strategy consistently yielded the best results for quantifying relationships within the matrices, suggesting its effectiveness in capturing important symptom-herb associations.

Key Results

  • Improved Prediction Accuracy: The proposed Tri-TSPA method, particularly using the PDH_Path (Symptom → Syndrome → Herb), significantly improved prediction performance.
  • Precision@30 was improved by 10.8% compared to the PH_Path (Symptom → Herb) strategy.
  • Recall@30 was improved by 11% compared to the PH_Path strategy.
  • Optimal Parameters Identified: Through parameter tuning, optimal values for neighborhood size (K=30) and various similarity weighting parameters were determined, leading to the best experimental results.
  • Qualitative Validation: The paper demonstrates the method's ability to generate a Top-30 list of herbs for a given symptom (e.g., Stomachache), and these lists contain herbs with known properties for relieving pain and stomach-related issues, aligning with TCM principles.

Limitations and Future Work

  • Limitations:
    • The paper notes that the Insomnia and Tourette datasets lacked syndrome information, requiring the creation of "virtual objects" to construct the tripartite graph. This might introduce some artificiality.
  • The study acknowledges that while the method identifies potential symptom-herb patterns, further work is needed to discover "effective combinations of interacting herbs" for a given symptom.
  • Future Work:
    • Explore new measures for evaluating performance.
  • Develop novel similarity measures to better capture rich topological features.
  • Investigate new matching functions to improve the predictive performance and identify synergistic herb combinations.
  • Further explore the use of the relevance function and matching function to discover effective herb combinations.

Summary updated

Abstract

Unlike the western medical approach where a drug is prescribed against specific symptoms of patients, traditional Chinese medicine (TCM) treatment has a unique step, which is called syndrome differentiation (SD). It is argued that SD is considered as patient classification because prior to the selection of the most appropriate formula from a set of relevant formulae for personalization, a practitioner has to label a patient belonging to a particular class (syndrome) first. Hence, to detect the patterns between herbs and symptoms via syndrome is a challenging problem; finding these patterns can help prepare a prescription that contributes to the efficacy of a treatment. In order to highlight this unique triangular relationship of symptom, syndrome, and herb, we propose a novel three-step mining approach. It first starts with the construction of a heterogeneous tripartite information network, which carries richer information. The second step is to systematically extract path-based topological features from this tripartite network. Finally, an unsupervised method is used to learn the best parameters associated with different features in deciding the symptom-herb relationships. Experiments have been carried out on four real-world patient records (Insomnia, Diabetes, Infertility, and Tourette syndrome) with comprehensive measurements. Interesting and insightful experimental results are noted and discussed.
BibTeX
@inproceedings{chen2015mining,
  author = {Jinpeng Chen and Josiah Poon and Simon Poon and Ling Xu and Daniel Sze},
  title = {{Mining Symptom-Herb Patterns from Patient Records Using Tripartite Graph}},
  booktitle = {Evidence-based},
  year = {2015},
  doi = {10.1155/2015/435085},
}
Figure