MEDEVAC 작전환경에서 위험인지 ICU 배정을 위한 책임 분리형 LLM-DRO 기반 의사결정 프레임워크

초록

This paper proposes a responsibility-separated framework integrating Large Language Models (LLMs) with Distributionally Robust Optimization (DRO) for allocating limited Intensive Care Unit (ICU) resources in military medical evacuation (MEDEVAC) operations under delay risk and operational constraints. The LLM is restricted to an input-generation layer, while ICU allocation is performed using a Wasserstein-based DRO model that buffers distributional estimation errors. The problem is formulated as a mixed-integer model with Conditional Value-at-Risk (CVaR) to capture the mean performance and tail risk. The results showed that increasing the robustness radius does not monotonically reduce the performance but can induce discontinuous changes in allocation decisions. Even small robustness levels significantly alter selection structures, while tail performance is more sensitive to estimation errors than mean performance. These findings suggest that robustness acts as a structural switching mechanism rather than a simple conservatism adjustment. The framework provides a foundation for risk-aware resource allocation in high-risk military and healthcare environments.

키워드