Research On Blood Pressure Detection Method Based On Deep Learning

In this paper,two methods of blood pressure detection are studied based on photoplethysmography and historical blood pressure data.The first is blood pressure prediction based on Xgboost integrated model,and the second is blood pressure prediction based on one-dimensional convolutional neural network.In this paper,from the results of the two methods,they are both suitable for blood pressure detection tasks,and the effect of the convolutional neural network model is significantly better than the machine learning model.In addition,the accuracy of the results of the one-dimensional convolutional neural network model can reach the American National Hypertension Standard,and the British Hypertension Association standard A level.The research content of this article has a certain reference value to the blood pressure detection task,especially for wearable devices,personalized prediction and other applications.First,based on the MIMIC-III database,this article selects the pulse wave PPG signal and blood pressure BP signal as the research object.Starting from data preprocessing,through the noise reduction of signal data,the data samples of this paper are selected;then in order to construct the training data of this paper,that is,the input data and output target values of the model,this paper has designed two sliding window searches for waveform detection The algorithm is mainly used to detect the peak point and the trough point in the signal.The detection results show that the two detection algorithms have the same effect.Detecting the peak and trough points of the signal is the key to construct training data in this paper.For the machine learning part,this paper takes a certain period on the PPG signal as the feature extraction interval and extracts 13 statistical features among them;and after this feature interval corresponds to the BP signal,the maximum value in the corresponding interval will be used as the systolic pressure,and the first minimum value on the left of the systolic pressure will be used as the diastolic pressure,so as to obtain the two target values corresponding to the feature interval.For the deep learning part,the last peak(trough)point on the BP signal is used as the candidate target value of systolic pressure(diastolic pressure),and the historical PPG signal and BP signal before this point are the features interval corresponding to the candidate target value.The size of the feature interval is determined by the size of the model input dimension.In this paper,the size of the model input dimension is 512.For the model training part of this article,this article first uses machine learning models for training,including linear models,decision tree regression models,neighbor regression models,support vector regression models,random forests,gradient boosting decision trees,Catboost models,Light GBM models,Xgboost models.From the model training results,the Xgboost model has the highest accuracy in predicting systolic blood pressure,which can reach a maximum of 20.9391 mm Hg;while the accuracy of the diastolic blood pressure is not much different from the Light GBM model,which can reach a maximum of 11.8612 mm Hg,and both are superior other machine learning models.In addition,in the case of adding historical blood pressure as a feature,the accuracy of the prediction result of the diastolic blood pressure of the Xgboost model reaches 4.3946 mm Hg,which can reach the A level in the National Hypertension Standard and the British Hypertension Association Standard;while in the systolic blood,the accuracy of the blood pressure prediction result only reached 9.0330 mm Hg,which still has a certain gap with the United States National Hypertension Standard and the British Hypertension Association Standard.Again,the deep learning research part of this article uses a one-dimensional convolutional neural network model.From the network model training results,it can be seen that the accuracy of the results of the deep learning method is better than the machine learning model for the presence or absence of historical blood pressure.For the case of training using PPG signal data alone as input data,the accuracy of the one-dimensional convolutional neural network model reached 19.3215 mm Hg in the systolic blood pressure result,which was about 1.614 mm Hg higher than the machine learning model;and the accuracy of the diastolic blood pressure result reached 11.4862 mm Hg,about 0.375 mm Hg higher than the machine learning model.For the case of training using historical blood pressure and PPG signal data as input data at the same time,the accuracy of the one-dimensional convolutional neural network model reached 5.3177 mm Hg in the systolic blood pressure result,which is about 3.7153 mm Hg higher than the machine learning model;reached 3.0524 mm Hg in the diastolic blood pressure result,which is about 1.3422 mm Hg higher than the machine learning model;the accuracy of its prediction results have reached the the American National Hypertension Standard,and the A level of the British Hypertension Association standard.Finally,this paper summarizes the research of this topic,and looks forward to the future research direction of this topic,and expounds the feasibility of the regression task in this research can be transformed into a classification task.