Machine learning-assisted rapid determination for traditional Chinese Medicine Constitution

Wen Sun, Minghua Bai, Ji Wang, Wang Bei, Yixing Liu, Qi Wang, Dongran HanChinese 2024 · DOI

TL;DR

Researchers have developed a machine learning-based approach to drastically shorten the standard 60-item Traditional Chinese Medicine (TCM) constitution questionnaire. By leveraging Automated Machine Learning (AutoML), the study identifies small subsets of "core" questions that can predict an individual's body constitution with up to 93.6% accuracy, paving the way for rapid, data-driven health assessments.

What Problem Does This Paper Address?

In Traditional Chinese Medicine (TCM), "Body Constitution" (BC) is a fundamental framework used to describe an individual's state of subhealth—the "third state" between health and disease. Identifying one's BC (e.g., "Qi-deficiency" or "Damp-heat") allows for personalized preventative care.

The gold standard for this is the Constitution in Chinese Medicine Questionnaire (CCMQ). However, the CCMQ consists of 60 items, which can be time-consuming for patients and practitioners. Furthermore, traditional scoring relies on simple linear addition, which fails to account for the fact that some symptoms are more "informative" than others or that relationships between symptoms are often non-linear. This paper seeks to use machine learning to identify the minimum number of questions needed to accurately determine a person's BC type and severity.

Methods

The researchers compared two distinct machine learning strategies for item selection:

  1. Supervised Learning (AutoML via TPOT): They used the Tree-based Pipeline Optimization Tool (TPOT), which uses genetic programming to automatically design and optimize the best machine learning pipelines (including model selection and parameter tuning). They tested every possible combination of items to find which ones best predicted the original 60-item scores.
  2. Unsupervised Learning (Variable Clustering): They used the varclus algorithm to group similar items into clusters and select the most representative item from each, reducing data complexity without requiring a predefined "target" score.

To find the "sweet spot" between brevity and accuracy, the team used elbow point analysis, identifying the point where adding more questions yielded diminishing returns in predictive power (AUC or $R^2$).

Datasets

The study utilized a massive dataset of 91,145 valid questionnaires collected via a web-based platform in China between 2015 and 2017. Participants were aged 15–64. This large-scale dataset provided the statistical power necessary to train robust ML models across all nine TCM constitution types.

Key Insights and Ideas

  • AutoML Outperforms Manual Clustering: The automated supervised approach (TPOT) was generally superior to unsupervised clustering for item selection. This suggests that when the goal is to replicate a specific diagnostic outcome, targeting that outcome directly during feature selection is more effective than looking for intrinsic data patterns.
  • Information-Dense Symptoms: Certain items were "consistently chosen" by all algorithms, acting as core predictors. For example, tiredness (item 2), cold aversion (item 19), and oily skin (item 39) were found to be highly representative of their respective constitution types.
  • Non-Linearity Matters: The study highlights that TCM diagnosis is inherently non-linear. By using algorithms like XGBoost and Random Forests within the TPOT pipelines, the researchers captured complex interactions between symptoms that traditional linear scoring misses.
  • The "Elbow" of Efficiency: For most constitution types, the researchers found that just 2 to 4 items could achieve an AUC (Area Under the Curve) greater than 0.80, which is considered excellent predictive performance.

Key Results

  • Classification Accuracy: The ML models predicted the correct BC classification with accuracies ranging from 0.819 to 0.936.
  • Predictive Power: AUC values for classification were high, ranging from 0.857 to 0.946.
  • Score Estimation: For predicting the actual numerical BC scores (severity), the models achieved $R^2$ values between 0.785 and 0.879, with Root Mean Square Errors (RMSE) stabilizing between 6.241 and 9.877.
  • Brevity: The researchers provided a "top three" list of item combinations for various subscale lengths, allowing practitioners to choose a version of the test that fits their specific time constraints.

Limitations and Future Work

The authors acknowledge a few key hurdles:

  • Computational Cost: The complexity of the TPOT algorithm combined with the massive dataset resulted in high computational demands.
  • Subjectivity: The study still relies on self-reported questionnaire data. While ML improves the processing of this data, it doesn't eliminate the inherent bias of self-reporting.
  • Missing Modalities: Traditional TCM diagnosis involves "four examinations," including tongue and pulse analysis. The authors suggest that future work should integrate multimodal data (e.g., image recognition for tongue analysis) to create a truly comprehensive rapid diagnostic tool.

Summary updated

Abstract

The aim of this study was to develop a machine learning-assisted rapid determination methodology for traditional Chinese Medicine Constitution. Based on the Constitution in Chinese Medicine Questionnaire (CCMQ), the most applied diagnostic instrument for assessing individuals' constitutions, we employed automated supervised machine learning algorithms (i.e., Tree-based Pipeline Optimization Tool; TPOT) on all the possible item combinations for each subscale and an unsupervised machine learning algorithm (i.e., variable clustering; varclus) on the whole scale to select items that can best predict body constitution (BC) classifications or BC scores. By utilizing subsets of items selected based on TPOT and corresponding machine learning algorithms, the accuracies of BC classifications prediction ranged from 0.819 to 0.936, with the root mean square errors of BC scores prediction stabilizing between 6.241 and 9.877. Overall, the results suggested that the automated machine learning algorithms performed better than the varclus algorithm for item selection. Additionally, based on an automated machine learning item selection procedure, we provided the top three ranked item combinations with each possible subscale length, along with their corresponding algorithms for predicting BC classification and severity. This approach could accommodate the needs of different practitioners in traditional Chinese medicine for rapid constitution determination.
BibTeX
@inproceedings{sun2024machine,
  author = {Wen Sun and Minghua Bai and Ji Wang and Wang Bei and Yixing Liu and Qi Wang and Dongran Han},
  title = {{Machine learning-assisted rapid determination for traditional Chinese Medicine Constitution}},
  booktitle = {Chinese},
  year = {2024},
  doi = {10.1186/s13020-024-00992-0},
}
Figure