Continuous Venous Oxygen Saturation Estimation via Population-Informed Personalized Gaussian Sum Extended Kalman Filtering

Abstract

Mixed venous oxygen saturation (SvO2) can play a pivotal role for patient monitoring and treatment in critical care and cardiopulmonary medicine. Unfortunately, its continuous measurement requires the use of invasive pulmonary artery catheters. This letter presents a novel population-informed personalized Gaussian sum extended Kalman filtering (PI-P-GSEKF) approach to continuous ${\mathrm { SvO}}_{2}$ estimation from arterial oxygen saturation (SpO2) measurement. The main challenge in ${\mathrm { SvO}}_{2}$ estimation is large inter-individual variability in the cardiopulmonary dynamics, which seriously deteriorates the efficacy of standard EKF. To cope with this challenge, we employ the GSEKF in which individual EKFs are designed using a mathematical model of cardiopulmonary dynamics whose operating points are selected from (i) population-level generative sampling (thus “population-informed”) and (ii) Markov chain Monte Carlo (MCMC) sampling based on a one-time SpO2-SvO2 measurement (thus “personalized”). Using the experimental data collected from 8 hypoxia trials in 4 large animals, we showed the ability of the PI-P-GSEKF to estimate ${\mathrm { SvO}}_{2}$ from ${\mathrm { SpO}}_{2}$ in comparison with its PI-EKF (EKF with population-level generative sampling as the source of process noise) and PI-GSEKF (GSEKF with population-level generative sampling alone) counterparts (average ${\mathrm { SvO}}_{2}$ root-mean-squared error: PI-EKF 4.7%, PI-GSEKF 4.3%, PI-P-GSEKF 3.0%). We also showed that population-level generative sampling and MCMC sampling both had respective roles in improving ${\mathrm { SvO}}_{2}$ estimation accuracy. In sum, the PI-P-GSEKF demonstrated its proof-of-principle to enable non-invasive continuous ${\mathrm { SvO}}_{2}$ estimation.