| In recent years,asthma has become the biggest killer of human health worldwide because of its high numbers of cases and long period in therapeutic time.In the field of traditional Chinese medicine（TCM）,asthma should be treated based on syndrome differentiation.Generally,TCM physicians need to divide asthma patients into five syndrome types according to their symptoms and physical signs before making prescriptions for different syndrome.With the development of smart healthcare,plenty of medical diagnosis work is inseparable from cutting-edge technologies such as machine learning and data mining.Thanks to the support of national policies,the development of intelligent diagnosis and treatment of TCM is being promoted.Aiming at the current situation that the subjectivity of TCM asthma syndrome differentiation is too strong,this thesis uses machine learning method to establish the standardized processing process of TCM asthma medical case data,and establishes a deep learning model to realize the prediction of TCM asthma syndrome differentiation of five syndromes,and realizes the objective auxiliary diagnosis and application of TCM asthma syndrome differentiation through informatization.The specific research work of this thesis is as follows:（1）The standardized processing flow of asthma medical record data in the respiratory department of TCM in a hospital in Tsingtao is established by using machine learning algorithm.In view of the non-standardization,missing and outlier of asthma medical records,in order to better train the model to realize the task of syndrome differentiation and typing of asthma,this thesis deletes the redundant value of patient characteristic information,completes the missing value of medical record information according to the two dimensions of characteristic importance and missing degree,quantifies the data of medical record characteristic text information,and normalizes the data through Min-Max algorithm.The pre-processed data are used for feature selection,the candidate feature sets of medical records are determined by m RMR algorithm and Pearson correlation coefficient threshold respectively,and the feature contribution is calculated according to XGBoost algorithm,so as to lock the simplified feature set as the input data during the training of pattern recognition network model.（2）To establish a deep learning model of TCM asthma syndrome differentiation.Aiming at the structured tabular data of asthma medical records,in order to avoid the problem that small sample data set is easy to lead to poor fitting ability of discriminating model and difficult to be used in practice,this thesis proposes a TabNet network model based on directional regularization-TD-TabNet.The targeted dropout directed regularization mechanism is applied to the traditional TabNet model to improve the prediction performance of the model for syndrome differentiation and classification of asthma,and avoid the problem of model over fitting in the context of small-scale asthma medical records.In order to verify the effectiveness of the optimized asthma discrimination model,it is compared with traditional TabNet and other algorithms.The experimental results show that TD-TabNet discrimination model is superior in accuracy,accuracy,recall,F1 value and so on.（3）Establish an auxiliary diagnosis platform of TCM asthma.In order to better realize the application of TCM asthma syndrome differentiation model,this thesis establishes a TCM asthma auxiliary diagnosis platform.Its main function is to visually predict and analyze the results of asthma syndrome differentiation.At the same time,it can intuitively monitor the state of the syndrome differentiation model and optimize its parameters,re train the data,and further serve TCM doctors to realize the auxiliary diagnosis task of asthma syndrome differentiation. |
