Simulation of Covid-19 epidemic evolution: are compartmental models really predictive?
This work addresses the problem of improving epidemic forecasting for public health policy, but it is incremental as it builds on existing SIR models with minor enhancements.
The study investigated whether an enhanced SIR epidemiological model could reliably predict COVID-19 evolution in Lombardy, Italy, using particle swarm optimization for parameter identification, and found that predictions were sensitive to training data size, requiring more data for convergence.
Computational models for the simulation of the severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) epidemic evolution would be extremely useful to support authorities in designing healthcare policies and lockdown measures to contain its impact on public health and economy. In Italy, the devised forecasts have been mostly based on a pure data-driven approach, by fitting and extrapolating open data on the epidemic evolution collected by the Italian Civil Protection Center. In this respect, SIR epidemiological models, which start from the description of the nonlinear interactions between population compartments, would be a much more desirable approach to understand and predict the collective emergent response. The present contribution addresses the fundamental question whether a SIR epidemiological model, suitably enriched with asymptomatic and dead individual compartments, could be able to provide reliable predictions on the epidemic evolution. To this aim, a machine learning approach based on particle swarm optimization (PSO) is proposed to automatically identify the model parameters based on a training set of data of progressive increasing size, considering Lombardy in Italy as a case study. The analysis of the scatter in the forecasts shows that model predictions are quite sensitive to the size of the dataset used for training, and that further data are still required to achieve convergent -- and therefore reliable -- predictions.