Research On Multi-feature Fusion For Blood Pressure Prediction

With the development of society and the improvement of people’s living standard,health problem has attracted people’s attention.With the changes of people’s living habits and dietary structure,the blood pressure problem has become a major problem of people’s lives.At present,there are so many methods to monitor blood pressure.In Hospital,doctor could make a diagnosis for a person who is suffering from high blood pressure,by ambulatory blood pressure monitor.Electronic blood pressure monitors are usually used to test blood pressure at home.The former requires the user to have a wealth of medical experience.The latter is easy to operate,but the accuracy cannot be guaranteed.These two methods both have the difficulties in the noninvasive and continuous detection of blood pressure.With the continuous development and improvement of hardware facilities,there have been non-invasive prediction of blood pressure by using two kinds of signals,Photoplethysmograph and electrocardiogram,which are collected by sensors.However,the limitation of wearable devices is still a difficult problem for non-invasive continuous blood pressure monitoring.However,most of the current studies on non-invasive blood pressure prediction using Photoplethysmograph are focused on the use of characteristic parameters in the Photoplethysmograph to predict blood pressure.What’s more,more work focused on the discovery of the transform characteristic parameters of Photoplethysmograph and the selection of the new prediction model.However,there is a lack of relevant studies on Photoplethysmograph waveforms and feature fusion for existing features to optimize blood pressure prediction.To solve this problem,this paper focuses on the related methods of feature fusion for blood pressure prediction.The main research work and innovation of this research include:Firstly,a statistical method is proposed to classify Photoplethysmograph signals.In this method,the autocorrelation function is used to determine the period of Photoplethysmograph,and the waveform type of Photoplethysmograph is determined by the pair number of the maximum and minimum in a period,which lays the foundation for further research on the optimization method of feature fusion.Secondly,Then,a feature selection method for blood pressure prediction is studied and implemented.The two characteristics of time domain and frequency domain are fused by Top-k and SVD methods,and the characteristics of high correlation with blood pressure are selected.Based on the related definitions of feature selection,the feature parameters of Photoplethysmograph in previous researches are selected in different ways,and the feature fusion is combined with the characteristics of Photoplethysmograph in time and frequency domain.The feature selection method is divided into two parts: 1)The importance of each variable in predicting blood pressure is evaluated by filtering feature selection methods.2)Based on learned relevant features,get the best feature subset by wrapping feature selection method.Three different prediction models,linear regression,random forest and neural network,were used to observe the changes of blood pressure prediction results and the algorithm was analyzed.Finally,experiments are designed and implemented to verify the effectiveness of the proposed method.The experiment results on MIMIC datasets demonstrate the feasibility and correctness of Photoplethysmograph waveform classification and feature fusion methods.