Deep Reinforcement Learning for Optimal Critical Care Pain Management
Date:
I delivered an oral presentation at the Engineering in Medicine and Biology Conference (EMBC 2019) in Berlin, summarizing our work on using deep reinforcement learning to support optimal pain management in the intensive care unit (ICU). The project introduced a sequential decision making framework that learns clinically interpretable morphine dosing strategies personalized to each patient’s evolving physiological and pain state, based on retrospective ICU data from the MIMIC-III database.
The talk presented the full methodological pipeline described in our paper, including the construction of a continuous state space from pain scores, vital signs, and analgesic administration events, and a discretized action space reflecting clinically meaningful morphine bolus doses. I described our dueling double deep Q network architecture, reward function design, and the use of prioritized experience replay to stabilize learning in this high-dimensional, noisy clinical environment.
In the presentation, I discussed the qualitative alignment between model-recommended interventions and physician practice, the model’s tendency to recommend more consistent dosing than clinicians, and how these results relate to existing clinical literature comparing intermittent versus continuous opioid administration. Although exploratory, this work demonstrated the potential of reinforcement learning to aid clinical decision making by identifying personalized dosing strategies using real-time physiological data readily available in critical care settings.