MCI Management Simulations

• MCI management simulations are computational models using multi-agent systems and AI to optimize response during mass casualty incidents.
• They integrate reinforcement learning, decision-support algorithms, and constraint-based optimization to improve resource allocation and patient transfer efficiency.
• Simulation platforms support human-AI collaboration and ontology-driven reasoning for strategic planning and operational training in emergency scenarios.

Mass Casualty Incident (MCI) Management Simulations represent a multidisciplinary domain where computational models, agent-based architectures, decision-support algorithms, and simulation platforms are deployed to improve the preparedness, efficacy, and real-time response of emergency systems facing mass casualty events. These events, which overwhelm healthcare or emergency services with sudden surges of critically injured or at-risk individuals, demand optimized allocation of limited resources, rapid and accurate triage, adaptive evacuation strategies, and support for dynamic collaboration. The field integrates multi-agent systems, reinforcement learning, constraint-based optimization, ontology-driven knowledge representation, and dynamic scenario testing.

1. Simulation Frameworks and Multi-Agent Architectures

Modern MCI management simulations frequently employ multi-agent system (MAS) architectures, wherein distinct entities such as patients, clinicians, transport vehicles, and decision-makers are represented as autonomous agents possessing individual decision logic. In simulation platforms such as MasTER, each agent encapsulates state vectors—injury severity, required resources, or location—with the overall system evolving through agent interactions and event scheduling (Liu et al., 10 Sep 2025). Crisis management simulations often extend this paradigm by incorporating layers representing human resources, vehicles, hospitals, and rescue points, with mobility and operational constraints encoded in ontology-backed databases (Le et al., 2023).

Key features found across platforms:

• Agent representation with dynamic state: Each agent (e.g., patient, doctor, vehicle) is described by a set of properties—acuity level, specialization needs, current status—with transitions triggered by probabilistic, rule-based, or reinforcement learning mechanisms.
• Layered simulation design: Separation of concerns between individual agent modeling (micro level) and system-wide operational flow (macro level), as in two-layer modeling approaches with discrete event simulation (DES) overlaying agent-based decision logic (Siebers et al., 2010).
• Integration with reinforcement learning: Deep RL agents use observed states to suggest or execute optimal allocation and transfer actions, often learning within high-dimensional state and action spaces using algorithms such as Proximal Policy Optimization (Liu et al., 10 Sep 2025).

2. Optimization of Resource Allocation and Patient Transfer

A fundamental challenge in MCI simulations is allocating scarce resources (ambulances, hospital beds, specialized medical care) in real time. Decision-support AI agents—typically deep RL models—are central in platforms designed for simulated and operational management of MCIs.

• State Space Formalization: For a patient $p_i$, $p_i = (id_i, s_i, r_i)$, with $s_i \in \{0,1,2,3\}$ as severity and $r_i \in \{0,1\}^8$ denoting binary requirements for resources (e.g., ventilators, burn centers, ORs).
• Action Space: $$A_t = \{ a \in \{1, ..., |H|\}^{|P|} : C(a) \leq c \}$$, where assignments $a$ must satisfy hospital capacity $c$ constraints.
• Reward Function: $R = \sum_{i \in P} R_i$, with individual rewards such as

$$R_i = 300 \cdot PQ_i + \begin{cases} 300 \cdot PT_i & \text{if } h = 1 \ 150 \cdot PT_i & \text{if } h = 2 \ 0 & \text{if } h = 3 \end{cases}$$

wherein $PT_i = \max(0, 1 - t_i / T_i)$ penalizes delays and $PQ_i = q_i / Q_i$ captures resource match accuracy (Liu et al., 10 Sep 2025).

Constraint-based recommendation systems utilize optimization matrices $M \in \mathbb{R}^{x \times y}$ assigning driver/vehicle resources to rescue points under capacity and response-time constraints:

• $\sum_{u=1}^x M_{u,v} \leq 1 \quad \forall v$
• $$\sum_{v=1}^y (M_{u,v} \times PA_{R_v}) \geq NB_{P_u} + 2 \times NB_{P_u}^{DIS} \quad \forall u$$ where $PA_{R_v}$ is seating capacity, $NB_{P_u}$ is evacuee count, and $NB_{P_u}^{DIS}$ is evacuees with disabilities (Le et al., 2023).

3. Decision-Support Modalities and Real-Time Collaboration

Simulation platforms now support varying degrees of AI involvement:

• Human-only: All allocation and triage decisions made without AI support. Often results in lower performance from non-experts (Liu et al., 10 Sep 2025).
• Human-AI collaboration: AI agent provides recommendations; human operators may accept or reject. Evidence suggests non-experts, when assisted by AI, can attain expert-level transfer decisions, lower mortality, and faster response (Liu et al., 10 Sep 2025).
• AI-only: Agent runs fully autonomously. Comparative studies show higher decision speed, consistency, and resource match accuracy with increased AI autonomy, particularly under high-complexity scenarios.

Usability results, such as reduced NASA-TLX workload scores and high SUS ratings, demonstrate practical viability for real-world deployment and training of both experts and novices.

4. Ontology, Constraint Satisfaction, and Event-Driven Mechanisms

Ontology-driven data models organize domain entities (vehicles, drivers, shelters) into queryable classes and relationships, facilitating dynamic reasoning and flexible adaptation in scenarios. Real-time operational capacity is achieved via layered architectures:

• Data Layer: Structured entity definitions (e.g., RDF triples for driver/vehicle pairs)
• Intelligent Layer: Constraint-based reasoning over available resources and requirements
• Service Layer: Integration with geospatial services (e.g., OSMNX) to compute time-sensitive metrics such as $T_{R-P}$ for vehicle-to-rescue-point assignments (Le et al., 2023).

Event-driven mechanisms (e.g., reactionary event selection in cyber incident exercises) model stochastic emergence of new incidents correlated with organizational vulnerability (“attack surface”). These mechanisms aid in capturing non-deterministic, cascading effects common in crisis management (Shreeve et al., 2023).

5. Simulation-Based Strategic Decision Support and Scenario Testing

Platforms such as MasTER enable running hypothetical (“what-if”) scenarios for training, preparedness assessment, and operational evaluation. By varying patient counts, hospital resource profiles, and transportation logistics, simulation engines can generate:

• Curvilinear relationships between resource/transfer choices and outcomes (e.g., mortality rates, completion times) (Liu et al., 10 Sep 2025).
• Quantitative metrics for scenario-by-scenario comparison: patient mortality, resource match rate, total completion time.

Such frameworks permit testing policy interventions, evaluating triage and transport protocols, and identifying bottlenecks prior to real-world implementation. Integration of simulation outputs into multi-criteria decision analysis (MCDA)—incorporating both static and dynamic performance measures—enables long-term strategic planning, highlighting that inclusion of dynamic criteria can materially alter optimal decision rankings (Aickelin et al., 2020).

6. Limitations, Future Directions, and System Scalability

Challenges in MCI management simulation include:

• Model assumption sensitivity and data calibration requirements
• Scalability with increasing complexity (patient volume, resource diversity)
• Adaptiveness to unmodeled crisis domains (transport, cyber, clinical)

Research trends focus on enhancing agent autonomy via reinforcement learning, optimizing multi-agent collaboration, and integrating MCDA with dynamic simulation outputs to improve robustness under uncertainty. Ontology-driven knowledge representation and modular architectures support extensibility and rapid system upgrades, while user feedback and controlled studies guide interface optimization and future development.

A plausible implication is that these simulation technologies, refined by empirical validation and scalable architecture, can be integrated with electronic health records and emergency systems to support real-time operations, preparedness training, and resource allocation in mass casualty situations.