The PPG-Based Noninvasive Blood Pressure Measurement Model

In the light of the threat of cardiovascular diseases,biomedical wearable devices used for the diagnosis of cardiovascular disease have attracted growing attention.Photoplethysmography(PPG)signal contains a large amount of physiological information,and thus could be utilized to aid the measurement of blood pressure on wearable monitoring devices.However,since current researches on blood pressure measurement models requires the extraction of artificial features,the model will lack generalizability and robustness.To address the problem above,a deep learning-based blood pressure measurement model is proposed and realized in this thesis.The model automatically extracts features from the PPG signal directly and predicts the systolic and diastolic blood pressure.In order to meet the requirements of deep learning neural network for data quality and data quantity,the thesis makes improvements to the data processing methods in the following aspects.First of all,in order to enrich the change of training data and add personal information that is currently lacking in open source databases,experiments are designed based on NIBP100 D,a blood pressure measurement equipment,and PPG100 C,a PPG sensor,all manufactured by BIOPAC(USA)in this thesis.Secondly,the process of quality assessment of the PPG signal for each pulse period is added to ensure data quality.Finally,data normalization methods and data enhancement methods get improved and tried.This thesis uses the method mentioned above to process the self-acquired BiCASL1.0 dataset and MIMIC II(Multiparameter Intelligent Monitoring in Intensive Care II,MIMIC II)dataset.The BiCASL1.0 was divided into training dataset,test dataset andvalidation dataset.The training and validation dataset are used to establish the blood pressure measurement model based on deep recurrent neural network with a contextual layer,meanwhile,the test dataset of BiCASL1.0and MIMIC II dataset are used for model evaluation.Meanwhile,MIMIC II dataset is used for testing and evaluation analysis.The model used in this thesis achieves an accuracy of 4.87±7.80 mmHg for systolic blood pressure and 4.60±6.09 mmHg for diastolic blood pressure on the test dataset of BiCASL1.0,and get an acceptable results on MIMIC II dataset which is 5.01 ± 6.73 mmHg for systolic blood pressure and 4.69 ± 4.65 mmHg for diastolic blood pressure.Different with the traditional machine learning method and without using complex characteristic engineering,this model acquires a good accuracy even with the raw data input directly,providing a reference and inspiration for blood pressure prediction with raw PPG signal in the future.Since the feature points are not required for automatic feature extraction,this thesis also provides some reference for solving the precision falling problem of the model,due to hardware differences and inaccurate detection of feature points.