ABSTRACT
Pulse recognition is an important part of the objectification and intelligence of TCM.This non-invasive and fast diagnostic method has great clinical value,however,data imbalance and cumbersome feature extraction are still challenging problems.The feature vectors were extracted from the one-dimensional pulse signal obtained after the Butterworth bandpass filter using the tsfresh library.And 9 columns of medical auxiliary features selected by exploratory data analysis were added.The feature filtering is performed jointly to derive 21 columns of feature vectors,which are used as input to the weighted soft voting fusion model.The data imbalance problem is solved by Borderline SMOTE algorithm.Construct a weighted soft-voting fusion model based on four types of machine learning:XGBoost,RF,LGBM,and GBDT.Eventually,the models will output specific pulse categories and demonstrate the performance by evaluating the metrics accuracy,precision,recall and F1 score.The experimental results show that the screened 21 feature vectors for a total of six types of pulse signal test sets achieve an accuracy of 90.04%in the five-fold cross-validation and take only 65.9466 seconds.It can provide a more accurate and intelligent auxiliary reference for pulse signal recognition,with lower operational complexity and higher accuracy compared to commonly used pulse recognition methods.The shorter training time also makes it more clinically useful in the recognition of multiple pulse signals.
ABSTRACT
Objective To analyze the factors of errors in the pulse recognition; To improve the speed of processing massive data; To explore the method of reducing the subjective errors in pulse recognition. Methods BP algorithm based on distributed MapReduce in Hadoop environment was optimized. Optimized BP algorithm was used to self-learn pulse-sequence data to reduce fitting errors. The pulse-counting data collected by TCM electronic pulse diagnosis instrument were used as input layer of neural network. Momentum-learning rate adaptive fast BP algorithm was adopted to train neural network. Results In the training set (75%) of 768 M, a total of 35 890 data were collected, and 29 150 items were correctly predicted in stand-alone mode, with the correct rate of 81.22%. MapRedece parallel improved BP algorithm model correctly predicted 35 841 items, with the correct rate of 99.86%. Conclusion Compared with traditional BP algorithm, BP algorithm based on distributed MapReduce in Hadoop environment has smaller fitting errors, with higher accuracy.