Use of Recurrent Neural Networks for Mean Blood Pressure Prediction Based on Impedance Cardiography Measurements.

In this paper we tried to predict value of mean blood pressure (MAP) basing on time series of measurements: maximum of Z first time derivative, (dZ/dtma), ejection time (ET), basal impedance Z (ZO) stroke volume (SV) and heart rate (HR) obtained non-invasively for every heartbeat with impedance cardiography (ICG). Continuous non-invasive measurements of MAP was accomplished with blood pressure monitor that uses the volume-clamp method. We used signals recorded in 10 young, healthy subjects, when performing three minutes handgrip test followed by two minutes of recovery. We used simple neural network model being a combination of the long short-term memory layer with the dense output layer. We divided data into sequences (data from 10 cardiac cycles and subsequent single MAP value as desired output). The best result, expressed as the mean absolute error, obtained with the neural network used by us was 0.4 after normalization of data. This result may indicate that the neural network used in this study was unable to predicted MAP value from ICG signal or that this signal do not provide sufficient information for correct estimation of MAP.