GuardianTwin: MEDEVAC Triage Support

• GuardianTwin is a decision support system that integrates logic programming and MILP optimization to deliver rapid, explainable triage decisions in MEDEVAC operations.
• It employs a layered architecture combining user-interface, logical reasoning, and optimization to enable modular, transparent, and effective resource allocation.
• Empirical evaluations show a 35.75% casualty reduction compared to baselines, highlighting its operational efficiency and strategic impact.

GuardianTwin is a decision support system designed for high-stakes Forward Medical Evacuation (MEDEVAC) operations. Integrating a logic programming framework with scalable mixed-integer optimization, it orchestrates multiple triage strategies and delivers explainable, rapid resource allocation recommendations targeting the principle of “right patient, right platform, right escort, right time, right destination.” GuardianTwin maintains a consistent digital twin of the operational theater, enabling modular, interpretable, and operationally effective solutions to the medical triage assignment problem. The system has achieved an average simulated casualty reduction of 35.75% compared to the strongest baseline under realistic mission scenarios (Patil et al., 14 Jul 2025).

1. Layered Architecture

GuardianTwin is architected in three principal layers:

• User‐Interface Layer: Provides two primary screens (Strategy Builder and Mission Details). The Strategy Builder allows users (field medics, planners) to select triage strategy variants—Urgency‐based, Reverse Triage, Situational—using high‐level controls, abstracting all direct interaction with underlying logic code or solver parameters. After computation, users receive assignment outputs specifying the “right patient, right platform, right time, right destination.”
• Logic‐Programming Layer: Implements a datalog‐style first‐order logic language supporting truth values True/False/Uncertain. This layer:
  • Encodes patient scoring rules (e.g., NISS, RTS, LIFE).
  • Specifies which optimization criteria and constraints apply.
  • Processes post‐solver symbolic reasoning for further decision support.
  • Utilizes the PyReason-based fixpoint deduction engine to compute the set of all logical consequences (Γ*(Π)), from which it can generate multiple integer programming (IP) problem variants.
  • Translates IP solution outputs back into new logical facts for downstream reasoning.
• Optimization Layer: Constructs and solves a collection of mixed-integer linear programming (MILP) models using python-mip. Each variant employs binary decision variables, linear constraints, and objective functions tailored to a specific triage scenario, with typical solve times of approximately 6 ms.

This separation ensures domain experts interact at a strategic level while modularly supporting new triage paradigms as logical and optimization fragments.

2. Multi-Variant Optimization Formulation

GuardianTwin frames the MEDEVAC allocation as a modular family of binary assignment MILPs parameterized by logical criteria:

• Decision Variables:
  • $x_{p,a} \in \{0,1\}$: casualty $p$ assigned to asset $a$.
  • $y_{p,f} \in \{0,1\}$: casualty $p$ assigned to facility $f$.
  • Each evacuated casualty is assigned exactly one asset and one facility: $\sum_a x_{p,a} = \sum_f y_{p,f} \in \{0,1\}$.
• Patient-Centric Functions (logic-derived):
  • $scr(p) \in (0,1)$: normalized triage severity.
  • $rtd(p) \in \mathbb{R}^+$: predicted return-to-duty time.
  • $lsi(p) \in \mathbb{R}^+$: time until life-saving intervention required.
  • $p$0/$p$1: travel times from asset to casualty, casualty to facility; $p$2.
• Constraints (activated via symbolic facts $p$3):
  • $p$4: $p$5 for all assigned $p$6 (life-saving deadline).
  • $p$7: unassigned $p$8 must have $p$9 (minimum severity for being skipped).
  • $a$0: unassigned $a$1 must have $a$2 (bounded for return-to-duty prioritization).
  • $a$3: impose travel-time bounds on legs for all assigned $a$4.
• Objective Functions:
  • Urgency-based triage ($a$5): $a$6.
  • Reverse triage ($a$7): $a$8.
  • Situational triage: uses above but imposes $a$9.

Symbolic logic orchestrates the construction and sequencing of multiple such formulations. The system can solve these in parallel or series, using rule-based coordination to compare solutions or iteratively relax infeasible constraints.

3. Symbolic Reasoning and Solution Translation

MILP outputs (assignments $y_{p,f} \in \{0,1\}$0, $y_{p,f} \in \{0,1\}$1) are translated into new logical facts such as $y_{p,f} \in \{0,1\}$2, $y_{p,f} \in \{0,1\}$3, $y_{p,f} \in \{0,1\}$4. If multiple optimization runs are staged, these are versioned as $y_{p,f} \in \{0,1\}$5 to denote solution provenance.

The expanded logical fact base (Π ∪ solution-facts) undergoes a secondary fixpoint deduction, driving the following:

• Selection and justification among alternative solutions (e.g., choosing the solution where reverse triage achieves at least 90% of $y_{p,f} \in \{0,1\}$6 while significantly improving $y_{p,f} \in \{0,1\}$7).
• Automated constraint monitoring, with rule-based relaxation (e.g., incrementing $y_{p,f} \in \{0,1\}$8 in $y_{p,f} \in \{0,1\}$9 until feasible).
• Proof trace generation for every decision, enabling explicit explanations (“why was patient $p$0 not evacuated?” yields deductions referencing violation or satisfaction of specific thresholds and functions).

This strictly monotonic, symbolic approach supports traceable, explainable inference from data ingestion to resource assignment.

4. Operational Integration and Workflow

GuardianTwin sustains a digital twin of the forward MEDEVAC domain, encompassing:

• Casualty states: geolocation, vital status, triage scores, intervention requirements.
• Evacuation assets: helicopters/UAVs parameterized by range, speed, crew-hours.
• Facilities: locations, capacities.

A domain ontology, encoded in the logic as predicates (casualty/1, asset/1, facility/1, score/2, use/1), standardizes data semantics and units.

The operational workflow proceeds as:

1. Intake of field data (vitals, locations, asset statuses) as initial facts.
2. Operator selection of desired strategies and constraint thresholds (via high-level UI controls) generates $p$1, $p$2, $p$3, etc.
3. Logic deduction generates MILP(s), which are solved and results reintegrated via logical facts.
4. Mission Details view displays assignments and constraint satisfaction (“P12 skipped because $p$4 threshold”).
5. In live use, the process iterates as asset statuses are updated, supported by timeline visualizations (cf. Gantt chart in Figure 1 in (Patil et al., 14 Jul 2025)).

This cyclic ingestion, deduction, optimization, and explanation enables rapid, evidence-backed decision cycles in mission-critical scenarios.

5. Empirical Evaluation and Impact

GuardianTwin’s evaluation utilized synthetic, exercise-grade datasets representing Arizona/Utah/Colorado operational areas. Key experimental features include:

Scenario Element	Description	Values
Number of runs	Randomized scenarios	250
Casualties	Data per run	n=25
Facilities	Destinations per run	m=10
Assets	Evac platforms per run	$p$5

Baseline Methods

• B1 Random: Random assignment, subject to $p$6.
• B2 Priority by $p$7: Assign in descending $p$8, then random, subject to $p$9.
• B3 Priority by $f$0: Assign in ascending $f$1, then random, subject to $f$2.

Performance Metrics

• Number of casualties evacuated.
• Sum of $f$3 for evacuated.
• Computation time: solver 6 ms per MILP, logic 3–5 s per run.

Findings

• Urgency and Reverse triage strategies evacuated approximately twice as many casualties as B1/B2, and 61.3% more than B3 for moderate asset counts.
• Situational triage is sensitive to strict air-time constraints.
• In multi-problem runs, logic-driven reasoning quantified trade-offs (e.g., reverse triage achieved 66% of $f$4 while boosting $f$5 4× over urgency-based triage).
• Constraint relaxation was frequently necessary for tight parameter settings.
• Across all runs and strategies, mean simulated casualty reduction was 35.75% relative to the strongest baseline (Patil et al., 14 Jul 2025).

A plausible implication is that logic-based formulation and comparative reasoning enable systematic, explainable identification (and operational validation) of non-intuitive but doctrinally compliant triage solutions.

6. Explainability and User Interaction

GuardianTwin’s design foregrounds explainability at both operational and cognitive levels:

• Strategy Builder: Users select among triage objectives (“Urgency,” “Return to Duty,” “Air-Time Limit,” “Life-Saving Intervention”) via checkboxes, each mapping to logical activation of corresponding optimization criteria and constraints.
• Mission Details View: Displays evacuation assignments in a table or map, with interactive explanatory panes linked to every decision. Example explanations include:
  • “Patient P23 not evacuated because $f$6 threshold in $f$7.”
  • “Constraint $f$8 satisfied: $f$9 min $\sum_a x_{p,a} = \sum_f y_{p,f} \in \{0,1\}$0 min.”
• Usability Observations: Preliminary feedback from PRCC/PECC officers indicated that proof-style explanations (ground-inference chains) foster greater trust compared to black-box solvers, and 90% of users could construct strategies and interpret results within two minutes of training.

Explainability is directly traceable to the monotonic, symbolic logic layer, making the full inference chain transparent for each assignment and constraint-check in the system.

7. Significance and Prospective Directions

By modularizing both the optimization and the reasoning processes, GuardianTwin demonstrates the viability of logic programming foundations for high-assurance optimization under uncertainty and operational constraints. The system’s empirical reduction in simulated casualties and its capacity to deliver explicit, proof-backed explanations support adoption in mission-critical settings where traditional black-box methods are unacceptable.

This approach suggests broader applicability in domains where symbolic reasoning over optimization variants is essential for trust, auditability, or doctrinal compliance. Future directions may include real-world deployment, expanded ontology for asset and intervention modeling, and integration with predictive models for dynamic patient status forecasting.

For comprehensive technical details and experimental results, see “Reasoning about Medical Triage Optimization with Logic Programming” (Patil et al., 14 Jul 2025).