An Optimal Control Approach to Learning in SIDARTHE Epidemic model

Andrea Zugarini, Enrico Meloni, Alessandro Betti, Andrea Panizza, Marco Corneli, Marco GoriarXiv 2020 · arXiv

Abstract

Standard epidemiological models often assume constant infection and recovery rates, which fail to capture the dynamic nature of real-world conditions such as lockdowns, behavioral shifts, and medical advancements. This paper introduces a machine learning framework based on optimal control to learn smooth, time-varying parameters for the complex SIDARTHE COVID-19 model. By utilizing a novel "temporal momentum" gradient flow algorithm, the authors successfully and accurately forecast epidemic trends in Italy and France.

Introduction

To predict the spread of infectious diseases, researchers rely heavily on compartmental models like SIR or SEIR. The SIDARTHE model (Susceptible, Infected, Diagnosed, Ailing, Recognized, Threatened, Healed, Extinct) improves upon these by explicitly separating undetected cases from detected ones and stratifying by symptom severity.

The SIDARTHE compartmental model structure, illustrating the transitions between different infection and recovery states.

However, most models assume transition rates are constant over time. In reality, these rates change drastically due to government interventions and improved medical treatments. Previous attempts to fix this involved manually setting piece-wise constant rates, which requires perfect knowledge of external events. This paper addresses the challenge of learning these time-variant parameters directly from raw epidemic data.

Methodology

The authors frame parameter estimation as an optimal control problem. Instead of learning static numbers, the model learns a function of time for each of the 18 parameters in the SIDARTHE Ordinary Differential Equation (ODE) system.

Gradient Flow and Temporal Momentum

The model uses PyTorch's automatic differentiation to backpropagate through an ODE solver (Heun's method). A key innovation is the Temporal Momentum update rule. The authors identified that parameters early in the timeline have a compounding effect on the final loss, making training unstable. Temporal momentum introduces a time-dependent learning rate that stabilizes the process by effectively propagating gradients across the timeline.

Derivative Regularization

To prevent parameters from fluctuating unrealistically, the authors add a regularization term that penalizes the first derivative of the parameters. This forces the learned rates to be smooth, plausible curves that reflect gradual shifts in human behavior and viral spread.

Experimental Setup and Results

The framework was trained on official daily COVID-19 reports from Italy and France spanning February to September 2020. The first 188 days were used for training, followed by validation and testing phases.

Key Results

  • Italy: The model achieved highly accurate forecasts with a Mean Absolute Percentage Error (MAPE) under 20% on average. Errors for critical metrics like Healed (H) and Deceased (E) were remarkably low at 4% and 6%, respectively.
  • France: The model successfully forecasted Threatened/ICU (T) and Healed (H) states with MAPEs of 16% and 2%.
  • Ablation Study: Sensitivity analysis demonstrated that without the novel temporal momentum term, the model performs poorly on test data and exhibits massive confidence intervals.
Sensitivity analysis of test loss relative to hyperparameters 'a', 'b', and training size, demonstrating the impact of the proposed learning framework.

Discussion and Insights

A significant finding was that the model naturally learned medical progress. By allowing parameters to float, the model deduced that recovery rates for ICU patients steadily increased over the summer of 2020, reflecting improved hospital treatments without being explicitly told so. This demonstrates the power of smooth priors in capturing underlying real-world trends.

Conclusion

While the model's performance is dependent on data quality—as seen in the higher error rates for noisy early French data—the optimal control approach represents a significant step forward in epidemiological modeling. Future work will focus on solving the full constrained problem using Euler-Lagrange differential equations to further refine these time-varying predictions.

Summary updated

Abstract

The COVID-19 outbreak has stimulated the interest in the proposal of novel epidemiological models to predict the course of the epidemic so as to help planning effective control strategies. In particular, in order to properly interpret the available data, it has become clear that one must go beyond most classic epidemiological models and consider models that, like the recently proposed SIDARTHE, offer a richer description of the stages of infection. The problem of learning the parameters of these models is of crucial importance especially when assuming that they are time-variant, which further enriches their effectiveness. In this paper we propose a general approach for learning time-variant parameters of dynamic compartmental models from epidemic data. We formulate the problem in terms of a functional risk that depends on the learning variables through the solutions of a dynamic system. The resulting variational problem is then solved by using a gradient flow on a suitable, regularized functional. We forecast the epidemic evolution in Italy and France. Results indicate that the model provides reliable and challenging predictions over all available data as well as the fundamental role of the chosen strategy on the time-variant parameters.
BibTeX
@misc{zugarini2020optimal,
  author = {Andrea Zugarini and Enrico Meloni and Alessandro Betti and Andrea Panizza and Marco Corneli and Marco Gori},
  title = {{An Optimal Control Approach to Learning in SIDARTHE Epidemic model}},
  year = {2020},
  eprint = {2010.14878},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2010.14878},
}
Figure