Toward Standardized ICU Trauma Resuscitation via Markov Decision Processes

Guideline-based recommendations for posthemostasis resuscitation in trauma patients remain limited. This study aimed to define an interpretable Markov Decision Process (MDP) to model intensive care unit (ICU) trauma resuscitation and evaluate its potential to support clinical decision-making using a previously established framework from sepsis care.

Retrospective observational study.

High-volume, level I academic trauma center.

Adult trauma patients admitted to the ICU between 2012 and 2019 with lactate ≥2 mmol/L.

Resuscitation was framed as a sequential decision-making problem using an MDP and focused on intravenous fluid (IVF) and vasopressor (norepinephrine and vasopressin) administration. Resuscitation was defined as concluding 1 hour after the last IVF bolus (≥250 mL) or vasopressor administration within the first 48 hours of ICU admission. A reinforcement learning (RL) agent was trained to recommend IVF and vasopressor treatments. Learned policy actions were compared to observed clinical practice using McNemar's test, and model performance was evaluated retrospectively.

A total of 4,305 subjects were included; 3,027 (70%) met the resuscitation endpoint (median resuscitation time: 22 hours). The learned policy aligned with clinical practice but differed significantly in vasopressor use, recommending norepinephrine less often (4% ± 6% vs . 12%, P  < 0.001) and vasopressin more frequently (13% ± 12% vs . 4%, P  < 0.001). IVF ≥250 mL was recommended in 19% ± 12% of decisions versus 13% in observed practice. Offline evaluation estimated the learned policy underperformed the behavior policy, consistent with prior findings in sepsis.

This study presents a clinically grounded MDP framework for trauma resuscitation that enables RL to be applied retrospectively in critical care. The learned policy mirrored trends previously observed in sepsis, aligning with clinician behavior while exposing variation in vasopressor use. These findings validate the relevance of the proposed MDP in trauma care, demonstrate the feasibility of applying RL in this context, and establish a foundation for future modeling, evaluation, and decision support.