Objective In order to solve the problem of difficult extraction of feature points and low robustness in the model of blood pressure measurement with pulse waveform, and improve the universality and accuracy of the blood pressure model, we present a new convolutional recurrent neural network-blood pressure (CRNN-BP) . Methods Firstly, we use convolutional network layer to extract the waveform features of the pulse wave automatically; Secondly, recurrent network layer is used to correct the features of waveform according to the relationship of the change of the blood pressure in the continuous cardiac cycle; Finally, full connected layer is used to predict the current blood pressure value. Results The model is validated using real human physiological information in the MIMIC data set. The mean absolute error (MAE) of systolic and diastolic blood pressure are 2.71 mm Hg and 1.41 mm Hg, respectively. Other than the accuracy of CRNN-BP is consistent with the standards of AAMI and BHS, it's superior to CNN-BP which does not use the recurrent network layer and traditional blood pressure regression models which use all pulse wave shape points.Conclusions CRNN-BP effectively extracts the waveform features of pulse wave and improves the accuracy and robustness of the model.