A reinforcement learning algorithm to optimize resource utilization in combat casualty care

Fluid resuscitation immediately following a hemorrhagic injury improves clinical outcome. However, future military conflicts are expected to result in mass-casualty incidents with limited availability of fluid-resuscitation resources. Here, we developed and assessed the performance of a reinforcement learning AI method that optimized clinical outcome and fluid allocation under constrained resources. We generated a large cohort of synthetic trauma casualties using a validated cardio-respiratory computational model, simulating vital-sign time-series data for realistic battlefield scenarios involving hemorrhage, tourniquet application, and fluid resuscitation. For each casualty, we simulated three intervention options—infusion or no infusion of one fluid unit every 30 min over 90 min—and assessed whether the interventions restored the vital signs to “healthy” levels. Using these data, we trained the reinforcement learning model to predict the optimal sequence of interventions that maximized the number of casualties restored while minimizing fluid utilization. Using independent simulated data, we found that the AI model was twice as efficient and restored more than twice as many casualties as the current standard of care across varying numbers of casualties and resource limitations. These results highlight the model’s potential to enable personalized interventions, enhance treatment efficiency, and support automated medical decision-making in resource-constrained environments.