Koopman Representations for Early Outbreak Warning and Minimal Counterfactual Intervention in Multi-Agent Epidemic Simulations

Technical Summary: Koopman Representations for Early Outbreak Warning and Minimal Counterfactual Intervention

Problem Statement
The paper addresses the challenge of predicting and controlling epidemic outcomes in complex, multi-agent systems near critical tipping points. In regimes where transmissibility is neither low enough to guarantee extinction nor high enough to guarantee a widespread outbreak, the final outcome (contained spread vs. major outbreak) is highly sensitive to small perturbations in initial conditions, timing, or local interactions. Traditional aggregate compartmental models (e.g., SIR) often obscure individual heterogeneity and spatial contact structures, while agent-based models (ABMs) generate high-dimensional, nonlinear dynamics that are difficult to summarize or predict early. The core problem is twofold: (1) Can early, aggregate observations of a developing epidemic trajectory reliably predict whether the system will cross a major outbreak threshold? and (2) If a trajectory is predicted to be dangerous, can minimal, localized counterfactual interventions (e.g., restricting a single agent's mobility for one day) alter the system's long-term trajectory to prevent the outbreak?

Methodology
The authors propose a three-component computational framework:

1. Multi-Agent Epidemic Simulator:

   • Environment: A spatial grid ($50 \times 50$) with 200 off-grid home locations and 500 agents.
   • Dynamics: Agents follow fixed daily routines with phase shifts, creating heterogeneous contact patterns. Transmission occurs via co-location in grid cells.
   • Disease Progression: Agents have heterogeneous susceptibility and immunity categories (strong, medium, low, compromised), which dictate individual viral load curves. Viral load determines infectiousness and behavioral restrictions (e.g., agents with high viral loads become homebound).
   • Transmission: A deterministic threshold rule based on the product of the source's viral load and the target's susceptibility.
   • Observables: The system records daily aggregate counts (susceptible, infected, recovered, dead, new cases, mobile vs. homebound infected) rather than full microscopic states for the prediction task.

2. Koopman Representation and Early Warning:

   • Koopman Learning: The authors learn a low-dimensional latent representation ($z_d \in \mathbb{R}^6$) of early epidemic trajectories (5-day windows) using a neural encoder. The dynamics in this latent space are approximated by a linear transition matrix ($A$), allowing for short-horizon forecasting ($h=5$ days) via $z_{d+1} \approx A z_d$.
   • Training Objective: The model minimizes reconstruction loss, latent linearity loss, and multi-step prediction loss.
   • Classification: A Random Forest classifier uses features derived from the early observation window (direct statistics) and the Koopman latent space (latent coordinates and forecasts) to predict the probability of a major outbreak (defined as a final attack rate $\rho \ge 0.3$).

3. Counterfactual Intervention:

   • Protocol: For trajectories predicted to be high-risk, the system performs counterfactual simulations. A single agent is quarantined (forced to stay home) for a single selected day, while all other parameters (random seeds, initial conditions, routines) remain identical to the baseline.
   • Metrics: The intervention's success is measured by the reduction in final attack rate ($\Delta \rho$) and peak infected count. An intervention is considered "outbreak-preventing" if it shifts the final attack rate from $\ge 0.3$ to $< 0.3$.

Key Contributions

• Koopman-Based Early Warning: The paper demonstrates that Koopman operator learning can effectively compress high-dimensional, nonlinear multi-agent epidemic trajectories into a low-dimensional latent space that preserves information relevant to long-term outcomes.
• Feature Integration: It shows that Koopman-derived features (latent states and forecasts) are as informative as direct epidemiological summaries (e.g., active infection counts) for classifying outbreak risk, achieving high accuracy with short observation windows.
• Minimal Intervention Analysis: The study provides a systematic method for testing the modifiability of near-critical trajectories. It establishes that minimal perturbations (one agent, one day) can qualitatively alter the system's regime, shifting it from a major outbreak to extinction, provided the intervention occurs before the transmission chain becomes self-sustaining.
• Tipping Point Characterization: The work characterizes the "near-critical" regime where similar parameter settings yield divergent outcomes, highlighting the sensitivity of the system to specific contact sequences rather than just aggregate transmission rates.

Experimental Results

• Dataset: 10,200 simulation runs were generated with susceptibility parameters swept narrowly around a transition point (1.3015–1.3025), resulting in a nearly balanced dataset of contained (4,825) and major outbreak (5,375) runs.
• Prediction Performance:
  • The Koopman model alone achieved a ROC AUC of 0.9980 at the run level.
  • The Random Forest classifier, utilizing both direct and Koopman-derived features, achieved a test accuracy of 0.9958 and a ROC AUC of 0.9999 on held-out 5-day windows.
  • Feature importance analysis confirmed that Koopman-derived features contributed significantly (22.49%) to the classification, comparable to direct infected series features (23.57%).
• Intervention Outcomes:
  • In illustrative counterfactual cases, a single-day quarantine of a selected agent successfully prevented major outbreaks in several instances, reducing the final attack rate from ~0.69 to near zero.
  • In other cases, the same type of intervention only delayed the epidemic peak or reduced the maximum active burden without preventing the final outbreak, demonstrating that the efficacy of minimal interventions depends heavily on the specific dynamical position of the trajectory.

Significance and Claims
The paper claims to provide a computational framework for studying phase-transition-like behavior in multi-agent systems. Its primary significance lies in the joint use of early prediction and counterfactual simulation:

1. Identifying Modifiable Trajectories: The framework identifies trajectories where the outcome is not yet predetermined and remains sensitive to small perturbations.
2. Minimal Intervention Viability: It demonstrates that in near-critical regimes, extremely localized and temporary interventions can be sufficient to redirect a system away from a catastrophic outcome, challenging the notion that only broad, population-wide measures are effective.
3. Methodological Bridge: The work bridges nonlinear dynamics, spectral analysis (via Koopman operators), and data-driven decision support, showing that learned latent representations can serve as both forecasting tools and feature sources for intervention planning.

The authors explicitly state that this is a controlled computational study using abstracted agents and simplified transmission rules. They do not claim the results constitute a calibrated epidemiological decision system for real-world application but rather a proof-of-concept for analyzing tipping points and intervention sensitivity in complex adaptive systems. Future work is suggested to include richer assumptions (vaccination, detailed networks) and cost-constrained policies.