Toward Standardized ICU Trauma Resuscitation via Markov Decision Processes

Area of Science:

• Critical Care Medicine
• Computational Biology
• Trauma Surgery

Background:

• Post-hemostasis resuscitation guidelines for trauma patients are limited.
• Intensive care unit (ICU) trauma resuscitation involves complex fluid and vasopressor decisions.
• A Markov Decision Process (MDP) framework can model sequential decision-making in critical care.

Purpose of the Study:

• To define an interpretable MDP for ICU trauma resuscitation.
• To evaluate the potential of an MDP-based approach to support clinical decision-making.
• To adapt a sepsis care framework for trauma resuscitation.

Main Methods:

• Retrospective observational study of adult trauma patients (2012-2019) with lactate ≥ 2 mmol/L.
• Framed resuscitation as an MDP focusing on intravenous fluid (IVF) and vasopressor administration.
• Trained a reinforcement learning (RL) agent to recommend treatments and compared its policy to observed clinical practice.

Main Results:

• 4,305 patients included; 70% met resuscitation endpoint.
• Learned RL policy aligned with clinical practice but differed significantly in vasopressor use (less norepinephrine, more vasopressin).
• Offline evaluation suggested the RL policy underperformed observed practice, consistent with sepsis studies.

Conclusions:

• A clinically grounded MDP framework for trauma resuscitation was established.
• Reinforcement learning (RL) can be applied retrospectively in critical care settings.
• The study validates the MDP relevance in trauma, demonstrates RL feasibility, and provides a foundation for decision support.