End-to-end deep learning and sensor fusion for non-invasive BP monitoring using multivariate physiological signals

Pietro Cerveri1, Mattia Sarti1, Matteo Rossi1, Giulia Alessandrelli1, Carolina Lombardi2, Luca Mainardi1
1Politecnico di MILANO, 2Istituto Auxologico Italiano


End-to-end deep learning may process in bundle raw physiological signals to output explicitly quantities of clinical interest, therefore avoiding fiducial point identification in the signal waveforms, disregarding any a-priori model linking intermediate variables to clinical quantities, removing the need for model parameter calibration, enabling data redundancy by sensor fusion and artifact mitigation for increasing resiliency of the measure. This speech focuses on deep neural networks that can be specialized in the estimation of the arterial blood pressure leveraging the bundle of signals like electrocardiogram (ECG), photoplethysmogram (PPG), phonocardiogram (PCG), seismocardiogram (SCG) and bio-impedance (BI) concurrently acquired using a chest-worn apparatus, called Soundi (Biocubica Srl, Milan - Italy). Signal pre-processing, network features, and training process are described, along with potential architectural updates paving the way to AI explainability and trustworthiness. We present preliminary tests of the application of end-to-end deep networks to the estimation of the ABP on a cohort of healthy subjects undergoing an acquisition protocol involving mild physical exercises. During the signal acquisition, we recorded a number of ABP measures (both diastolic and systolic pressures were stored) by means of the medical certified Holter pressure system (GIMA ABPM, Italy), which serve as the labels for the network training. Performances are compared with state-of-the-art methods, including regression models using pulse transit time (PTT). Technical features and quantitative outcomes are discussed providing some potential pathways for the clinical translation of this approach.