Continuous Blood Pressure Prediction Based On Deep Learning

Obtaining continuous blood pressure waveform is of great significance for monitoring human health.Traditional blood pressure prediction methods mostly predict point by point or only predict high pressure and low pressure through parameters such as pulse signals,and cannot quickly obtain complete and accurate continuous blood pressure waveforms.Therefore,how to obtain continuous blood pressure waveform from pulse signal is an urgent problem to be solved.Focusin g on this problem,this paper proposes three continuous blood pressure prediction methods by improving and innovating the existing depth learning model.Through detailed experimental comparison and result analysis,it is proved that these prediction method s based on depth learning can accurately predict continuous blood pressure waveforms from pulse signals.The main work of this paper is as follows:(1)A continuous blood pressure prediction method based on multiple gated recurrent unit(GRU)embedded in SENet(Sequeze-and-Excitation Network)is proposed.Firstly,multiple GRU channels are used to extract the features in pulse,then SENet module is embedded to learn the interdependence between multiple channels,thus obtaining the weight of each channel.Finally,the weight is added to each channel,and the predicted continuous blood pressure value is obtained through integration of two Linear layers.The experimental results show that embedding SENet can effectively increase the prediction ability of multi-GRU structure and obtain good continuous blood pressure waveform.Compared with Long Short-Term Memory(LSTM)network and GRU model without embedding SENet,the MSE error of the proposed prediction method is reduced by 31.4% and 27.3% respectively,and the training time is shortened by 69.8% and 68.7% respectively.(2)A continuous blood pressure prediction method based on improved SENet convolution neural network and self-learning parameter filter is proposed.Firstly,according to the timing characteristics of pulse-blood pressure data,the modeling process of internal channels of the original SENET is improved,that is,the internal channel dependency relationship of the original SENET is changed from Linear layer to GRU layer,which can utilize t iming data.The improved SENet inputs the descriptor of each channel in the convolution layer into the GRU,so that the weight output by the GRU can be weighted to each channel,thus enabling the convolution neural network to have the ability to utilize time sequence information.The experimental results show that,The improved SENet can effectively increase the prediction ability of the simple convolution neural network to time series data,and the prediction accuracy is improved by 34.8%,23.5% and 36.0% compared with the simple convolution neural network when the number of convolution layers is 2,3 and 4.On this basis,the self-learning parameter filter is used to eliminate burrs in the blood pressure prediction waveform,and finally a smooth continuous blood pressure prediction result is obtained.(3).A continuous blood pressure prediction method with convolution multi-head attention is proposed.The method replaces all the connection layers in the classical multi-head attention with convolution layers,which solves the problem of low accuracy of the classical multi-head attention in continuous blood pressure prediction tasks and the problem of too complicated dimensional operation when the classical multi-head attention performs separate operations.The experimental results show that convolution multi-head attention improves the accuracy of continuous blood pressure prediction by about 66% compared with classical multi-head attention,and convolution multi-head attention has the best comprehensive perfo rmance in continuous blood pressure prediction tasks when sparse attention is selected.