Integrating Transfer Learning with Scalogram Analysis for Blood Pressure Estimation from PPG Signals

Abstract

The blood pressure estimation plays a crucial role in assessing cardiovascular health and preventing related complications. One of early warning indicators for heart disorders is elevated blood pressure. Thus, monitoring of blood pressure continuously is needed. This paper provides a novel transfer learning approach for blood pressure estimation using photoplethysmography from the publicly available database namely MIMIC-II. The Continuous Wavelet Transform was used to transform the PPG signals into scalograms, which were then input into six different deep learning models: VGG16, ResNet50, InceptionV3, NASNetLarge, InceptionResNetV2 and ConvNeXtTiny. The obtained deep features from each one of these models were employed to estimate BP values using Random Forest. The models were assessed using mean absolute error and standard deviation in estimating the systolic and diastolic blood pressure values. Out of six models, ConvNeXtTiny and VGG16 proved to be particularly challenging, resulting in the mean absolute error (MAE) of 2.95 mmHg and 4.11 mmHg for systolic blood pressure respectively, and standard deviation of 1.66 mmHg and 2.60 mmHg for diastolic blood pressure, respectively. The achieved result complies with the clinical standards set by Advancement of Medical Instrumentation Standard and the British Hypertension Society standard. The suggested method shows that reliable Blood Pressure estimation from photoplethysmography signals is possible with the use of deep learning and transfer learning. Above all, ConvNeXtTiny offers a dependable method for continuous blood pressure monitoring that satisfies clinical requirements and may help in the early identification of cardiovascular problems.