Particle filter algorithms for identification of minimally parametrized Wiener models of drug administration effect

Abstract

Closed-loop drug administration is performed by a feedback controller keeping the patient response around a pre-set level in the face of process disturbance and uncertainty. To guarantee the safe operating envelope of the closed-loop drug delivery system, controller design for performance and robustness is typically based on a mathematical model of the plant. Minimally parametrized Wiener models comprising a linear dynamic block and a smooth output nonlinearity have been shown to accurately capture the patient-specific pharmacokinetics and pharmacodynamics in closed-loop drug administration. Due to the nonlinear model dynamics, identification techniques based on linear or linearized models yield biased parameter estimates and thus introduce superfluous uncertainty. The present paper compares the performance of three nonlinear system identification algorithms: an extended Kalman filter, a conventional particle filter (PF), and a PF making use of a orthonormal basis to estimate the probability density function from the particle set. A database of patient models estimated under PID-controlled neuromuscular blockade during general anesthesia is utilized, along with the clinical measurements. The bias and variance of the estimated models are related to the computational complexity of the identification algorithms, highlighting the superiority of the PFs in this safety-critical application.