A Support Vector Machine Learning for the Upward and Downward Tendency Theory of Traditional Chinese Medicine

Objective: In order to study the traditional Chinese medicine (TCM) drug property theory based on machine learning (ML), the support vector machine (SVM) as a powerful model in ML is worthy of exploring for the distinguish on the TCM drug Upward and Downward Tendency in this theory. Methods: 1. To select and include the research materials and objects. From a TCM drug textbook applying for the state TCM university education (TBTCMUE), a total (t) of 135 TCM drug was selected, containing necessary features such as the main chemical structure among active ingredients, the botany Family, the Medicinal Part, and the TCM drug's features of Four-Qi or Cold-Heat, Five-Taste or Flavor, and Tendency, a known classification feature in the TBTCMUE. 2. To establish the TCM drug features' coding rules. TCM drug's features were digitized and coded with domain codevalues in different levels of extent. 3. To build template, training, and pattern vector data sets from TCM drug features. Each TCM drug's Tendency feature can suppose as a column vector in matrix T(:, 1), whereas the other features or their combinations could suppose as corresponding numbers column vectors in another matrix L(:, c). Meanwhile, extract r numbers (r < t) of drugs from the 135 TCM drugs (t=135) to create r rows matrix T (r, 1) for template set and the same rows matrix L (r, c) for a learning or training set. The left p (p=t-r) numbers of TCM drugs after the extraction could form a matrix P (p, c) for a pattern or testing set. 4. To create an SVM model for recognizing the TCM drug Tendency. By matching the template set T (r, 1), each TCM drug of the pattern set P (p, c), was recognized by the SVM model and the trained SVM rule from its the learning set L (r, c). Then, the matched rate expected, a value counted from the matched results, was counted divided by the total matching count and greater than a supposing threshold value (THV) of 0.75, referring to an acceptable result in the pattern recognition. Results: Based on the recognition with SVM and the SVM algorithm rule relative to each different pattern when parameters were specified by p=1 and r=tp=134, the Medicinal Parts, as one single feature vector of the TCM drug's features, showed that its matched rate of the Downward Tendency was 0.8 and referred to be an acceptable outcome. Another accepted at the matched rate of 0.75 was the Family-Flavor-Benzene-atom combination group of four-feature combination vectors for Downward Tendency. For Upward Tendency and Dual Tendency, the features had unacceptable results. Conclusion: The TCM drug's feature vectors composed of TCM drug Medical Parts, or TCM drug's features combination of the family, flavor, Benzene, and atom, can be helpfully utilized to reveal the contributing factors for the TCM drug's Downward Tendency.