A DNN assisted parallel control architecture for automated anesthesia delivery addressing inter-patient variability

The control of depth of hypnosis (DoH) in anesthesia is critical for ensuring patient safety, minimizing clinical intervention, and improving postoperative outcomes. However, the nonlinear and patient-specific nature of anesthetic dynamics poses significant challenges to traditional single-loop PID controllers which often fail to provide robust performance across diverse patient profiles. To address these limitations, this work proposes a robust and adaptive control framework based on a two-degree of-freedom (2-DOF) parallel control scheme, which decouples setpoint tracking from disturbance rejection. To further enhance adaptability, an offline deep neural network (DNN)-based gain scheduling strategy is integrated, enabling controller parameters to be precomputed based on patient-specific characteristics. Genetic algorithms are used to optimize the controller parameters for 13 clinically validated patient categories by minimizing the integral of absolute error (IAE). Simulation results demonstrate substantial improvements in maintaining the Bispectral Index (BIS) within the desired range of 40-60, validating the framework's ability to deliver personalized and precise control of anesthesia. The paper presents the motivation behind the proposed approach, details the control architecture and optimization methodology, and concludes with a discussion of performance evaluation across various patient models.