Research On Cuffless Blood Pressure Modeling Method Based On Deep Neural Network

Continuous blood pressure monitoring is of great significance for the diagnosis and management of people with hypertension.It also has the role of regulation and early warning for the risk of major diseases in certain specific populations such as patients with arrhythmia.The intermittent measurement method of the cuff-type sphygmomanometer cannot meet the needs of continuous and long-term blood pressure monitoring.Existing non-invasive blood pressure methods require manual feature extraction.Due to the large individual differences,the accuracy is often insufficient or the scope of application is limited,which makes it difficult to meet the needs of blood pressure monitoring of different populations.A small number of existing articles have shown the potential and feasibility of deep learning in this field.With its automated feature learning and expression capabilities,it is expected to achieve accurate estimation of blood pressure for a wider population.It can also provide a theoretical basis and prototype for the blood pressure measurement function of wearable devices.The DNN's(Deep Neural Network)different dimensions of feature expression can also provide a new possibility for revealing the internal relationship between body surface physiological signals and internal invasive blood pressure from the perspective of models.Regarding the issues above,this thesis studies the blood pressure estimation method based on deep neural network,and establishes the blood pressure estimation model according to the MIMIC-III(Medical Information Mart for Intensive Care III)public data set and the data collected before surgery in patients with arrhythmia in Fuwai Hospital.The main content of this paper includes the following aspects:1.Based on the DNN,a wearable continuous blood pressure modeling method is proposed.Using body surface physiological signals collected by the wearable device,mainly ECG(electrocardiogram)and PPG(photoplethysmogram),a relationship model between signals and continuous blood pressure is established based on MIMIC-III public database.The model's error for SBP (Systolic Blood Pressure)and DBP(Diastolic Blood Pressure)are-0.08 ± 9.0 and-1.6 ± 5.0,respectively,which is close to AAMI(Association for the Advancement of Medical Instrumentation)standards.2.Based on deep transfer learning,a modeling method for small-scale arrhythmia data sets is proposed.The establishment of the model combines a strategy network,adaptive layers,and a domain loss function.The model is verified on the invasive continuous blood pressure data set collected from patients with arrhythmia before surgery.The test errors of SBP and DBP are 0.49±7.69 and 0.17±4.85,respectively,which meet the AAMI standard.The model also shows good tracking ability for large fluctuations in blood pressure.3.Based on the above methods,a wearable blood pressure monitoring system is developed Only with the input of ECG and PPG signals collected by the wearable device,the system can output the dynamic blood pressure in real time.The analysis of the model complexity also reflects the good real-time performance and ease of use of the model.The relevant prototype has been demonstrated at the 20 th China Hi-Tech Fair.This thesis relies on wearable devices,proposes an end-to-end blood pressure modeling method based on DNN and deep transfer learning.It establishes a direct connection between invasive blood pressure and signals collected by wearable devices without manual extraction of features.It achieves cuffless high-precision continuous blood pressure monitoring,and the accuracy meets AAMI standards.It also provides a theoretical basis and prototype for wearable blood pressure monitoring equipment.