Table 2 Applications of machine learning in TCM research

Natural product development

Deep neurol network [47, 50], RF [53, 54, 58, 59, 93], SVM [51, 53, 54, 57, 59, 93], DT [59, 93], neural network [53, 59]

RF was better than SVM, neurol network and DT in screening hepatotoxic compounds [59]. RF model is more accurate than SVM and DT in identifying molecular characteristics of natural product compounds with the meridians of TCM [93]

Disease diagnosis

SVM [10, 61, 66, 81,82,83], DT [68, 81,82,83], neural network [45, 61,62,63, 65, 82, 83], RF [61, 64, 67, 82, 83], CNN [64, 67, 70,71,72,73,74,75,76,77,78, 81], RNN [80], LSTM [64], MLP [68], residual CNN [81]

Compared with SVM, DT, traditional neural network and RF, CNN achieves higher accuracy in tongue shape classification [71]; Residual CNN model with highest accuracy and sensitivity in sounds auscultation [81]; SVM usually achieves better performance than other traditional algorithms in pulse diagnosis [82, 83]; CNN perform better than traditional methods in pulse diagnosis, with accuracy above 90% [71]

Disease treatment and effect evaluation

SVM [84, 87, 88], RF [84], GAN [85], CNN [86, 88], DT [88]

CNN performs better compared with SVM and DT [88]

Prediction of biomarkers in TCM

RF [89], SVM [90, 92], Residual CNN [48], DT [92], neurol network [92]

SVM model is better (accuracy:98.45%) than DT, neurol network in predicting tumor necrosis factor-alpha converting enzyme inhibitors [92]