Safety-Critical Control of Active Interventions for COVID-19 Mitigation

Aaron D. Ames, Tamás G. Molnár, Andrew W. Singletary, Gábor OroszIEEE Access 2020 · DOI

TL;DR

This paper introduces a novel framework for guiding public policy on COVID-19 mitigation by treating epidemiological models as control systems. By applying safety-critical control techniques, the authors demonstrate how to synthesize active intervention policies that formally guarantee bounds on infections, hospitalizations, and deaths, offering a more robust approach than traditional methods.

What Problem Does This Paper Address?

The COVID-19 pandemic highlighted the critical need for effective strategies to mitigate its spread. While epidemiological models have been instrumental in understanding disease dynamics and informing policy, they often treat interventions (like social distancing or quarantines) as static or implicitly updated. This can lead to inaccurate predictions and suboptimal policy decisions, especially as new data emerges and interventions need to be dynamically adjusted. The authors address the challenge of how to dynamically guide these interventions to ensure public safety, specifically by guaranteeing that key metrics like hospitalizations and deaths remain within acceptable limits. This is particularly relevant as regions begin to relax restrictions in a decentralized manner.

Methods

The core of the paper's contribution lies in reframing epidemiological models (like SIR and SIHRD) as control systems. Instead of viewing them as autonomous dynamical systems, the authors treat them as systems where human behavior, through active interventions, acts as a control input.

Key techniques include:

  • Control Barrier Functions (CBFs): This framework provides necessary and sufficient conditions for ensuring that a system remains within a "safe set" (e.g., a set of states where infections are below a certain threshold). CBFs allow for the synthesis of controllers (active intervention policies) that guarantee this safety.
  • Optimization-Based Control: The authors formulate the synthesis of active intervention policies as optimization problems. This allows for finding the "minimal" intervention needed to maintain safety, thereby minimizing the economic and social disruption associated with strict measures.
  • Predictor Feedback: To account for time delays inherent in data collection (incubation periods, testing times), the authors employ predictor feedback. This involves using a model to predict the current state of the system based on delayed measurements, enabling the delay-free control design to function effectively.
  • Parameter Identification: For both the SIR and SIHRD models, the authors use optimization techniques (specifically CMA-ES) to fit model parameters to real-world data, including mobility data, to ensure the models accurately reflect observed trends.

Datasets

The paper utilizes publicly available datasets for COVID-19 cases, hospitalizations, and deaths in the United States. Mobility data from SafeGraph was also used to estimate the effectiveness of interventions. The authors fit their models to data up to specific dates (e.g., May 30, 2020) and then use these fitted models to predict future trends and synthesize policies.

Key Insights and Ideas

  • Epidemiological Models as Control Systems: The fundamental shift in perspective, viewing epidemiological models not just as descriptive tools but as controllable systems, is the paper's most significant conceptual contribution. This opens the door to applying rigorous control theory to public health policy.
  • Formal Guarantees for Safety: Unlike many existing modeling approaches, the safety-critical control framework provides formal mathematical guarantees that key metrics (infections, hospitalizations, deaths) will remain within predefined safe bounds. This is crucial for building public trust and informing policy.
  • Dynamic Intervention Policies: The approach allows for dynamic, time-varying intervention policies that can adapt to changing conditions, rather than relying on static or pre-determined strategies. This is essential for navigating the evolving nature of a pandemic.
  • The Importance of Time Delays: The paper explicitly addresses the impact of time delays in data and disease progression, demonstrating how predictor feedback is essential for maintaining safety in real-world applications.
  • State-Level Policy Differentiation: The framework can be applied at granular levels (e.g., individual states), revealing that different regions may require distinct mitigation strategies due to varying parameters and local conditions.

Key Results

  • Accurate Modeling: The SIHRD model, when fitted with real-world data, accurately captures trends in infections, hospitalizations, and deaths.
  • Superior Intervention Policies: Safety-critical active intervention policies consistently outperform naive linear "opening up" policies. These policies lead to significantly lower peaks in infections and hospitalizations, and ultimately, fewer deaths.
  • Bounded Outcomes: The synthesized policies formally guarantee that hospitalizations and deaths remain below predefined maximums (Hmax and Dmax), a critical outcome for healthcare system capacity.
  • State-Specific Recommendations: The study provides state-level policy recommendations, illustrating how different states might open up at varying paces based on their specific epidemiological data and the synthesized safety-critical control inputs. For instance, some states can open more quickly, while others require sustained mitigation.

Limitations and Future Work

  • Data Dependency: The accuracy of the predictions and policy recommendations is inherently tied to the quality and availability of data. The authors emphasize the need for continuous data assimilation and higher-fidelity data.
  • Model Simplification: While the SIHRD model is more complex than the SIR model, it is still a simplification of reality. The authors acknowledge that these models may not be universally applicable to all infectious diseases or all geographical regions.
  • Parameter Uncertainty: While optimization is used to estimate parameters, there remains inherent uncertainty. The authors suggest that future work could explore robust control strategies that account for parameter uncertainty.
  • Policy Implementation: The paper presents a "proof-of-concept" for translating control inputs into public policy via a gating criterion. Further research is needed to refine these translation mechanisms and integrate them into actual policy-making processes.
  • Generalizability: The authors suggest that the framework can be applied to more complex compartmental models and different geographical regions, indicating a broad scope for future research.

Summary updated

BibTeX
@article{ames2020safety,
  author = {Aaron D. Ames and Tamás G. Molnár and Andrew W. Singletary and Gábor Orosz},
  title = {{Safety-Critical Control of Active Interventions for COVID-19 Mitigation}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.3029558},
}
Figure