Multivariable modeling and control of the response to anesthesia

The long term goal of this research is the development of an automated or partially automated drug delivery system to assist anesthesiologists in surgery, allowing them to concentrate on critical safety issues. Simultaneously, this anesthesia delivery system should optimize drug consumption on an individual patient basis during surgery.; Although there is a long history of research in feedback control of anesthesia, there have been few true applications of feedback control in clinical practice. Conventionally, clinical research studies focusing on anesthesia dosing and delivery have been based on pharmacokinetic (PK) - pharmacodynamic (PD) models. PK models are used to reflect the diffusion, the distribution and the metabolism of anesthetics in the human body. PD models describe how varying anesthetic doses affect the body's vital signs, such as heart rate and blood pressure, as well as brain activity (EEG). Unfortunately, these models fail to capture the response to disturbance inputs, and fail to describe the interrelation among effects in the human body.; To address these shortcomings, this dissertation proposes the use of piecewise-linear multivariable models of the effects of anesthetic and stimuli inputs on patient vital signs. These models are developed by utilizing data-based system identification methods. Different model structures such as linear time invariant (LTI), piecewise linear switching and parameter-varying, are investigated and verified. Robust gain scheduled controllers are constructed and simulated using a linear parameter varying (LPV) framework. The use of multivariable state-space models for patient modeling is novel in this research area, as is the proposed use of piecewise-linear models. The modeling results indicate that the proposed piecewise-linear models yield improved simulation responses over traditional PK-PD models, in comparisons made with respect to single-output effects. Further, individual models may provide reasonable central models in the sense that simulated output responses obtained by applying the input data set for one volunteer to the piecewise-linear models for other volunteers produces an acceptable fit to the output data for the first volunteer. Similar results are found for our constructed individual controllers when applied as a generic controller for other volunteer data.