Cuff-less wearable blood pressure estimation system with photoplethysmography signal

Early detection of abnormalities in blood pressure (BP) helps to significantly reduce the risk of diseases such as chronic heart failure, heart attack and stroke. In current clinical practice, most of the BP measurement techniques consist of cuff-based, invasive, or multi-sensor systems. These techniques are impractical as they are inconvenient for patients and require frequent calibration. In this study, continuous, cuff-free, non-invasive and calibration-free estimation of BP with only photoplethysmogram (PPG) signal was aimed. PPG and Aortic BP signals were obtained from the available to access MIMIC-II dataset provided by PhysioNet. Deep features were extracted from the PPG signal and its derivatives using pre-trained ResNet50, VGG16, DenseNet121, MobileNetV2 and Xception models. Recursive feature elimination algorithm was applied to all deep features and reduced selected feature set was obtained. Selected deep features were used with classical machine learning and deep learning algorithms and BP prediction models were obtained. Statistical metrics and visual analytical tools were used to evaluate the models. In addition, BP prediction models were graded using the AAMI and BHS protocols. The results show that the deep features obtained with the DenseNet121 model have a higher performance than the other models. In addition, it was concluded that deep learning algorithms performed more effectively than classical machine learning methods. The MAE±SD error values of the proposed BP prediction models were calculated as 5.63±8.83, 2.82±5.60, 4.39±5.95 for SBP, DBP and MBP, respectively. The results show that the BP prediction models are level-A according to the BHS protocol and meet the AAMI criterion.