Research On Noninvasive Blood Pressure Prediction Based On Deep Learning

As technology advances and living standards continue to improve,people are more concerned about their health.In daily life,it is not easy to detect whether an individual is suffering from hypertension disease,but the harm it brings is innumerable.Therefore,how to efficiently detect blood pressure and prevent the harm of hypertension in advance is a problem that people are constantly studying.Manual detection of blood pressure requires professional operators and specialized maneuvers to perform the test,and the blood pressure value cannot be detected every time.The development of deep learning techniques in recent years has made possible an efficient method for detecting blood pressure based on convolutional neural networks for noninvasive blood pressure prediction.However,the shortcomings of this method are the large number of model parameters designed and the slightly poor prediction accuracy.To address these problems,the following two main works are carried out in this thesis.1.A method DUnet based on Unet convolutional neural network is proposed to carry out the prediction of noninvasive blood pressure for the problem of blood pressure prediction in small depth models.The DUnet model contains two branching networks,the main network uses the designed large convolutional kernel module to extract as many features as possible and is mainly responsible for the prediction work of the model;the subnetworks use the small convolutional kernel module to extract different feature information and are used to enhance the prediction ability of the model.The experimental results demonstrate the prediction capability of the DUnet model,which achieves 91.23% prediction accuracy in the 15 mm Hg error range and 5.9% prediction accuracy improvement in the 5 mm Hg error range under the BHS criterion with only 2.6 ms prediction time when the parameters of the DUnet model are reduced by 27% compared with those of the original model.At the same time,DUnet method provides an idea to apply the model to embedded devices by breaking the model into several small models in an embedded scenario and performing prediction step by step until the model gets satisfactory results.2.A WSeg Net method based on the Seg Net model is proposed for the problem of blood pressure prediction by a high-precision depth model.Compared with the first work,WSeg Net not only predicts blood pressure in a shorter time,but also predicts blood pressure with the highest level of accuracy.WSeg Net contains only one backbone network,and the improved structure using the main framework is able to extract information of different sizes and use attention mechanisms to quickly extract important features from the perspective of the channels,enhancing the model prediction performance while adding little additional prediction time.To further improve the predictive performance of the model,the model uses modules for different feature fusions after the decoding is finished.The experimental results show that WSeg Net achieves 96.27% accuracy within the BHS standard 15 mm Hg error range,and the method achieves the highest prediction accuracy within 2.2 ms prediction time without significantly increasing the parameters.At the same time,WSeg Net’s approach only needs to be fine-tuned on its basis to obtain a model for blood pressure prediction projects,providing a base model for future blood pressure prediction using neural network methods.