On adaptive control and particle filtering in the automatic administration of medicinal drugs

Abstract

Automatic feedback methodologies for the administration of medicinal drugs offer undisputed potential benefits in terms of cost reduction and improved clinical outcomes. However, despite several decades of research, the ultimate safety of many--it would be fair to say most--closed-loop drug delivery approaches remains under question and manual methods based on clinicians' expertise are still dominant in clinical practice. Key challenges to the design of control systems for these applications include uncertainty in pharmacological models, as well as intra- and interpatient variability in the response to drug administration. Pharmacological systems may feature nonlinearities, time delays, time-varying parameters and non-Gaussian stochastic processes. This dissertation investigates a novel multi-controller adaptive control strategy capable of delivering safe control for closed-loop drug delivery applications without impairing clinicians' ability to make an expert assessment of a clinical situation. Our new feedback control approach, which we have named Robust Adaptive Control with Particle Filtering (RAC-PF), estimates a patient's individual response characteristic in real-time through particle filtering and uses the Bayesian inference result to select the most suitable controller for closed-loop operation from a bank of candidate controllers designed using the robust methodology of mu-synthesis. The work is presented as four distinct pieces of research. We first apply the existing approach of Robust Multiple-Model Adaptive Control (RMMAC), which features robust controllers and Kalman filter estimators, to the case-study of administration of the vasodepressor drug sodium nitroprusside and examine benefits and drawbacks. We then consider particle filtering as an alternative to Kalman filter-based methods for the real-time estimation of pharmacological dose-response, and apply this to the nonlinear pharmacokinetic-pharmacodynamic model of the anaesthetic drug propofol. We ultimately combine particle filters and robust controllers to create RAC-PF, and test our novel approach first in a proof-of-concept design and finally in the case of sodium nitroprusside. The results presented in the dissertation are based on computational studies, including extensive Monte-Carlo simulation campaigns. Our findings of improved parameter estimates from noisy observations support the use of particle filtering as a viable tool for real-time Bayesian inference in pharmacological system identification. The potential of the RAC-PF approach as an extension of RMMAC for closed-loop control of a broader class of systems is also clearly highlighted, with the proposed new approach delivering safe control of acute hypertension through sodium nitroprusside infusion when applied to a very general population response model. All approaches presented are generalisable and may be readily adapted to other drug delivery instances. Show more