Blood pressure is a physiological indicator of human body. Continuous measurement of arterial blood pressure in each cardiac cycle is an important basis for real time diagnoses. Most of the cuffless continuous blood pressure measurements are performed according to the predictive models based on the pulse wave and electrocardiogram signals. However, they may produce errors due to the limited measurements. In this paper, multiple physical signs, such as impedance cardiogram, are explored to improve the measured accuracy of blood pressure. Experiments were conducted upon 55 volunteers, and results show that the random forest model based on multi-parameter feature fusion outperformed the linear model based on a single feature, with mean absolute errors of 2.56 mmHg and 1.91 mmHg for the prediction of systolic and diastolic blood pressure, respectively. It proves that the proposed cuffless blood pressure prediction model based on the multi-feature fusion could improve the accuracy of blood pressure prediction.