Start Date

5-2020 12:00 AM

Description

High-dose propofol is being investigated for its potential antidepressant effect. Propofol is titrated to induce burst suppression, a specific EEG pattern. However, propofol is difficult to dose due to uncertainty in each patient’s pharmacokinetics (PK) and pharmacodynamics (PD), and the lack of a commercially available monitor of propofol concentration. Clinicians currently infer the proper drug dose after observing the EEG response to the given dose. In this report we share our development of an automated controller to optimally administer propofol-induced burst suppression. We designed a deep deterministic policy gradient (DDPG) algorithm, which includes two deep neural networks and relates a 2-dimensional action space with a 3-dimensional state space. Our DDPG prototype did not satisfy our minimum training criteria. However, we share our diagnosis of current limitations in training a DDPG-based RL agent to administer propofol to PK-PD-simulated in silico patients. We also discuss potential solutions to improve RL agent training and performance.

Comments

Due to COVID-19, the Symposium was not able to be held this year. However, papers and posters were still submitted.

Available for download on Saturday, May 01, 2021

Share

COinS
 
May 1st, 12:00 AM

A Reinforcement Learning Based Control Approach for Propofol-Induced Burst Suppression

High-dose propofol is being investigated for its potential antidepressant effect. Propofol is titrated to induce burst suppression, a specific EEG pattern. However, propofol is difficult to dose due to uncertainty in each patient’s pharmacokinetics (PK) and pharmacodynamics (PD), and the lack of a commercially available monitor of propofol concentration. Clinicians currently infer the proper drug dose after observing the EEG response to the given dose. In this report we share our development of an automated controller to optimally administer propofol-induced burst suppression. We designed a deep deterministic policy gradient (DDPG) algorithm, which includes two deep neural networks and relates a 2-dimensional action space with a 3-dimensional state space. Our DDPG prototype did not satisfy our minimum training criteria. However, we share our diagnosis of current limitations in training a DDPG-based RL agent to administer propofol to PK-PD-simulated in silico patients. We also discuss potential solutions to improve RL agent training and performance.