Abstract
Cardiovascular and respiratory diseases (CVRD) are the leading causes of death worldwide. The construction of health digital twins for patient monitoring is becoming a fundamental tool to reduce invasive procedures, lower healthcare costs, minimize patient hospitalization, design clinical trials and personalize therapies. The aim of this study is to investigate the feasibility of machine learning-based monitoring of healthy subjects and CVRD patients in an in silico context. A population of virtual subjects, both healthy and with CVRD, was created using a comprehensive zero-dimensional global closed-loop model. In particular, the most relevant model parameters were varied within physiologically and pathologically plausible ranges, using local sensitivity analysis to guide the parameter selection. Then, we trained Gaussian process regression (GPR) models, informed by wearable-acquired data (e.g., heart rate, peripheral pressures and oxygen saturation), to predict variables normally acquired with invasive or operator-dependent methods (e.g., central venous pressure, stroke volume, cardiac output, left ventricular ejection fraction, arterial partial pressure of O2, arterial partial pressure of CO2). We also evaluated GPR models performance under simulated wearable signal acquisition errors via an error propagation analysis. Presented results demonstrate the feasibility of predicting in-hospital variables from wearable-derived indices using GPR models under the controlled conditions and assumptions of the adopted modeling approach.