Non-invasive Continuous Blood Pressure Measurement Based On Deep Neural Network

Blood pressure is an important physiological indicator of the human body.Many diseases are closely related to blood pressure.Among them,hypertension has become one of the major diseases that modern humans are troubled.According to the World Health Organization,one fifth of adults worldwide have hypertension till 2015.In China,the incidence of hypertension in adults is even more than a quarter.Therefore,the prevention and monitoring of cardiovascular diseases such as hypertension has become the focus of research in various fields.Among them,accurate measurement of blood pressure is a hotspot and the difficulty of research.At present,the most accurate measurement used clinically is the invasive direct measurement,but it is only suitable for special patients due to the traumatic injury.The most commonly used method for measuring blood pressure is the cuff-barrier sphygmomanometer based on Koch’s auscultation.However,this method may has a large error and it will also cause discomfort to the patient during the measurement,and it is an intermittent measurement method that cannot meet the requirements of disease monitoring.Therefore,the study of non-invasive continuous blood pressure measurement has very important practical significance.In recent years,machine learning methods,especially deep neural networks,have become a research hotspot in various fields.Many researchers have begun to try to apply it to medical problems.At the same time,studies of blood pressure measurements have found a correlation between pulse wave and blood pressure,which can be measured using a single photoplethysmographic signal.Therefore,applying a deep neural network to process pulse waves to measure blood pressure has become a feasible solution.In this research,applying deep learning theory to construct blood pressure measurement model for the problem of non-invasive continuous blood pressure measurement and applying transfer learning to solve the difference of sample distribution is studied.A new data preprocessing method is proposed to achieve high precision blood pressure.The application performance of transfer learning in the blood pressure measurement model is obtained.The main work and achievements of this thesis are as follows:(1)A photoplethysmographic signal preprocessing algorithm for deep neural networks is proposed.The algorithm eliminates the invalid data and noise data,preserves the effective information of the pulse wave as much as possible,and processes the data into equal-length single-cycle structured sample data.The experiment proves that the algorithm is low in complexity,easy to run,and the sample information after processing is rich,which satisfies the learning requirements of the model.(2)The depth range of the model suitable for the experimental sample data is obtained through a large number of comparative experiments.Then,in this range,different neural network models with different structures are designed,and the best model on performance is obtained through experiments.Finally,the model’s structure on different individuals proves that the model structure has good robustness,and the average error is compared with the results of similar research,the result shows that the model has better performance in accuracy.(3)Transfer learning is applied to the blood pressure measurement model for the first time.By using the data transfer method based on probability distribution adaptation and the feature transfer method based on CNN model,the performance of the two methods in solving the problem of data distribution difference is obtained.The results show that both methods can improve the distribution of data to some extent,but there all have the disadvantage of reducing the accuracy of measurement.Finally,through summarization,the problems in the practical use of blood pressure measurement are pointed out and the future work is summarized.