Blood pressure prediction with deep recurrent network and contextual information

In recent years, the proportion of patients with hypertension has been increasing. How to give an alarm before abnormal blood pressure value and early treatment has become a widely concerned research topic. The structure of bidirectional LSTM is used to add the influence of the timing information from the negative time direction on the current state, and LSTM with residual connection is added to solve the gradient disappearance and gradient explosion caused by multi-layer network. Before the output layer, an additional layer is added to add the basic information data of the user and the activation function of the extra layer is modified linear unit (ReLU), multiple time series data are used to predict the blood pressure of different periods. The experimental results show that the best results using 24 time series data sets. On the dataset of 24 time series, blood pressure prediction of 1 h, 6 h, 12 h and 24 h at different time periods was performed. The prediction error and the prediction deviation for systolic blood pressure are respectively 0.002644, 0.003952, 0.004216, 0.005528 and 0.037796, 0.047931, 0.049879, 0.057454, for diastolic blood pressure are respectively 0.001226, 0.001554, 0.001706, 0.001955 and 0.024293, 0.028369, 0.030190, 0.032668. Experimental error compared with other models, the prediction error and prediction bias of the proposed model are reduced.