Cuffless Blood Pressure Estimation Using Features Extracted from Carotid Dual-Diameter Waveforms

The major challenges in deep learning approaches to cuffless blood pressure estimation is selecting the most appropriate representative of the blood pulse waveform and extraction of relevant features for data collection. This paper performs an analysis of a novel dataset consisting of 71 features from the carotid dual-diameter waveforms and 4 blood pressure parameters. In particular, the analysis uses gradient boosting and graph-theoretic algorithms to determine (1) features with high predictive power and (2) potential to be pruned. Identifying such features and understanding their physiological significance is important for building blood pressure estimation models using machine learning that is robust across diverse clinical environments and patient sets.