Robust Nonlinear State Space Model Identification for Hemorrhage Resuscitation

Elham Estiri; Hossein Mirinejad

Abstract

Fluid resuscitation is a medical intervention commonly used in hypovolemic scenarios to compensate for the lost blood volume and stabilize critically ill patients. Fluid management is currently ad-hoc and dependent on the physician's style and expertise. Such ad-hoc protocols lack the capability of accurately adjusting the fluid infusion dosages due to their empiric nature, especially in the presence of clinical disturbances, posing significant risk of adverse effects such as under- and over-dosing. Thus, treatment performance is compromised due to the lack of appropriate dosage adjustment tools available. This paper presents a novel modeling framework namely, robust nonlinear state space modeling (RNSSM), for predicting hemodynamic responses in hemorrhage resuscitation. The proposed approach innovatively integrates autoencoder learning and variational Gaussian inference (VGI) in a unified framework to develop nonlinear state space models that are highly amenable to the closed-loop control design from limited, noisy critical care data. The goal is to develop subject-specific models that can reliably predict mean arterial pressure (MAP) in response to fluid infusion in hemorrhage scenarios. The RNSSM approach aim to improve (1) model accuracy by considering subject-specific characteristics and drug attributes and (2) model reliability by accounting for uncertainties that inherently exist in clinical data. Enabling reliable, personalized hemodynamic models amenable to the closed-loop control design can potentially lead to development of efficient model-informed precision dosing strategies, promoting patient safety and outcomes in critical care.

Keywords: Hemodynamic Model; Fluid Resuscitation

Links
[Full text PDF][Bibtex]