Objectives: To research a better method for reducing the dimensions of Traditional Chinese Medicine(TCM) symptom information by comparing Principal Component Analysis (PCA), Logistics Principal Component Analysis (LPCA) and Deep Belief network Autoencoder(DBA). All methods were applied to the pneumonic symptom dataset. Method: Mean-Square Error(MSE) was used to measure the performance of models. Heat maps were used to show the weight distributions of models which contained 10, 20 or 30 features. And Scatter plots were used to show the sample distributions which are described by 2 features. Results: In the 5-layer DBA network, MSE do not always decrease as the neuronal number of the 2nd layer increase, if the number is larger than 40. 5-layer DBA models have the best performance if the number of feature < 8. Else LPCA models have the minimal MSE. If the number of feature > 16, 3-layer DBA models are similar to 5-layer DBA models at MSE. Compared to PCA and LPCA, 3-layer DBA has the maximum range of weights and only a few variables in each feature have weights which significantly higher or lower than the other. In samples distributions, the features of 3-layer DBA and 5-layer DBA can divide the samples into different clusters more clearly than PCA and LPCA. Conclusion: LPCA and DBA can instead of PCA on dimensionality reduction for TCM symptom information. Compared with LPCA, DBAs features represent better with the same number of features and DBAs performance is better by building a multi-layer network when the feature number is small.