In-silico evaluation of three control methodologies with model adaptation to minimize risk of overdosing in anesthesia

The ideal conditions for extracting good models for control are not attainable in clinical settings, due to patient safety and further enforced by ethical and regulatory frameworks. From prior observations, the patient model defined by the pharmacokinetic part is piecewise linear and mostly invariant among the patients, while the drug-dose effect relationship exhibits large variability, resulting in significant large gain variations in patient's model. In this paper, we propose a model for the gain adaptation as a twoinput (Propofol and Remifentanil) one output (hypnotic state BIS variable) linear area of the nonlinear surface of the dose-effect for general anesthesia. The new patient model is used for tuning controllers without over-dosing, i.e. no BIS-nadir values below 50 and avoid negative values of median prediction error indicative of over-dosing. A comparison of target controlled infusion (this is manual control with anesthesiologist closing the loop) against two control strategies is performed. A model based predictive control and a PID control scheme with model adaptation and co-administration in ratio control mode are compared before and after the patient model adaptation. The results indicate the adaptation step minimizes risk for over-dosing, as it minimizes modeling errors. Robustness of controllers has been assessed before the identification, encouraging the claim that predictive control closely mimics the human-in-the-loop target controlled infusion profiles. Evaluation criteria from clinical practice further enhance the added value of our solution. Real clinical data evaluation confirms the results from the simulation tests, showing a considerable match between the drug profiles titrated by anesthesiologist and those calculated by the proposed control algorithms. (c) 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).