A Novel Blood Pressure Estimation Method with the Combination of Long Short Term Memory Neural Network and Principal Component Analysis Based on PPG Signals

Abstract

The worldwide high blood pressure-related mortality rate is increasing. Alternative measurement methods are required for blood pressure measurement. There are similarities between blood pressure and photoplethysmography (PPG) signals. In this study, a novel blood pressure estimation methods based on the feature extracted from the PPG signals have been proposed. First of all, 12-time domain features have extracted from the raw PPG signal. Secondly, raw PPG signals have been applied to Principal Component Analysis (PCA) to obtain 10 new features. The resulting features have been combined to form a hybrid feature set consisting of 22 features. After features extraction, blood pressure values have automatically been predicted by using the Long Short Term Memory Neural Network (LSTM-NN) model. The prediction performance measures including MAE, MAPE, RMSE, and IA values have been used. While the combination of 12-time domain features from PPG signals and LSTM has obtained the MAPE values of 0,0547 in the prediction of blood pressures, the combination of 10-PCA coefficients and LSTM has achieved the MAPE value of 0,0559. The combination model of all features (22) and LSTM has obtained the MAPE values of 0,0488 in the prediction of blood pressures. The achieved results have shown that the proposed hybrid model based on combining PCA and LSTM is very promising in the prediction of blood pressure from PPG signals. In the future, the proposed hybrid method can be used as a wearable device in the measurement of blood pressure without any calibration.