Abstract
This paper proposes a data-driven Model Predictive Control (MPC) framework to optimize social isolation policies during an epidemic, using SARS-CoV-2 in Mauritius as a case study. By combining a compartmental SIHRD model with Genetic Algorithms for parameter tuning, the researchers demonstrate that dynamic, well-timed isolation levels can reduce fatalities by 40% while keeping hospitalizations within healthcare capacity.
Introduction
During a pandemic, governments face a grueling "balancing act": they must prevent healthcare systems from collapsing and minimize deaths without causing total socio-economic paralysis. Traditional responses often rely on blunt "on-off" lockdowns or static policies that do not adapt to real-time data. This research addresses the challenge of designing a systematic, automated way to decide when and how much to restrict social movement. The goal is to identify a "middle path"—a control strategy that maintains hospitalizations below a critical threshold and minimizes deaths, all while allowing for the maximum possible social and economic activity.
Methodology
The researchers developed a multi-stage control pipeline consisting of epidemiological modeling, system identification, and predictive control.
Epidemiological Modeling
The study utilizes an SIHRD model (Susceptible-Infected-Hospitalised-Recovered-Dead). To make the model "controllable," the authors introduced a state ($\Psi$) representing the population's response to isolation policies and a control input ($u$) representing the "containment index" or isolation level.
System Identification via Genetic Algorithms (GA)
Because epidemiological parameters—such as transmission and recovery rates—vary by region and virus variant, the authors employed Genetic Algorithms. These algorithms "evolve" a set of parameters that best fit real-world data from the third COVID-19 wave in Mauritius, ensuring the model reflects local dynamics.
Model Predictive Control (MPC)
MPC, an advanced control method typically used in aerospace and chemical processing, was applied to the epidemiological model. At each daily time step, the controller:
- Predicts the virus's spread over a 14-day horizon.
- Solves an optimization problem to find the best sequence of isolation levels to minimize a "cost" (a weighted sum of infections, hospitalizations, deaths, and the economic cost of isolation).
- Applies only the first day's decision and repeats the process the next day with updated data.
To ensure policy stability, the controller was restricted to shifting isolation levels only once every seven days, and the control variable was limited to discrete levels (e.g., partial vs. full lockdown).
Data and Experimental Setup
The study utilized real-world epidemiological data from Mauritius covering the period from December 2021 to May 2022, representing the island's most severe wave. Data points included daily new infected cases, daily deaths, and weekly new hospitalizations (interpolated to daily). Sources included the World Health Organization (WHO) Coronavirus Dashboard and the Mauritius Ministry of Health and Wellness.
Key Insights
The Power of Granularity
Increasing the number of available isolation levels (e.g., moving from 3 levels to 5) allows for a significantly smoother and more effective containment strategy. Higher granularity enables the government to "fine-tune" the response rather than relying on the blunt instrument of a full lockdown.
Strategic Timing
The MPC often prioritized removing isolation measures ($u=0$) at the start of a wave, implementing short, sharp increases in isolation only when predicted hospital loads threatened to exceed capacity. This suggests that the timing of interventions is more critical than their total duration.
Weighting Priorities
By adjusting the "weights" in the MPC's cost function, policymakers can explicitly define the trade-off between economic pain and public health. The study found that penalizing infections and the isolation level itself had the most significant impact on the resulting policy outcomes.
Results
Simulation results, validated against historical data from Mauritius, showed significant improvements over the actual historical response:
- Fatality Reduction: The MPC scheme reduced total deaths from 240 to 140, a 40% decrease.
- Hospital Capacity: While the actual wave saw hospitalizations peak at 525 (exceeding the 200-bed limit), the MPC successfully kept the peak at 190.
- Infection Control: Peak infections were reduced from 26,000 to 9,000.
- Socio-economic Balance: These gains were achieved without permanent lockdowns; the controller dynamically toggled isolation levels to allow for periods of normalcy when risks were low.
Limitations and Future Work
Despite the promising results, the authors identify several areas for further research. The current SIHRD model is deterministic; moving toward "grey-box" modeling using Gaussian Processes could better account for the inherent randomness in epidemic data. Additionally, the model assumes perfect population compliance with isolation levels. Future iterations must account for "lockdown fatigue" and the challenges of data sub-notification to ensure the controller remains robust in real-world scenarios.
Summary updated