Bayesian networks and chained classifiers based on SVM for traditional chinese medical prescription generation

Traditional Chinese Medicine(TCM) is playing an increasingly prominent role in lung cancer treatment, as it can prolong patients’ survival, improve their quality of life, and reduce the adverse effects of radiotherapy and chemotherapy. However, the effectiveness of TCM treatment depends more on the personal experience of doctors, and the standardization of TCM prescriptions needs to be strengthened. In this study, we use TCM clinical prescriptions to train a standardized TCM prescription generation model to provide an auxiliary prescription reference for physicians. However, in our initial experiments, we found two severe problems in the dataset. The first problem is a strong correlation between each herb; for instance, some herbs often appear together to treat specific symptoms. The second is a severe class imbalance within each label, a few herbs always appear in most prescriptions, but most herbs have a low frequency of occurrence in the total dataset. To solve the correlation between each herb label, we adopt the Bayes Classifier Chain(BCC) algorithm, whose basic classifier is Cost-Sensitive SVM targeted to the class imbalance of the label. Based on this, we also improve the BCC method according to the characteristics of TCM prescription dataset. In our BCC classifier, the Directed Acyclic Graph (DAG) construction method has high interpretability in the scenario of TCM prescription. After combining multi-label learning algorithms with several SVM algorithms and comparing their performance in detail, we find that BBC+CS-SVM best deals with class imbalance within the label in multi-label classification problems.