Stefano Giovanni Rizzo, Giovanna Vantini, Mohamad Saad, Sanjay Chawla
We propose an extension of the standard Susceptible Incubation Infection and Recovered (SEIR) model to capture the specifics of COVID-19 pandemic. In particular our model takes into account that (i) transmissions rates before and after lockdown interventions are different (ii) the majority of positive case are asymptomatic or goes unreported , and (iii) there is lag between the time an infected subject is tested and the day this is publicly reported. Besides a new model we use automatic differentiation to estimate the gradients (of the loss function) to learn the parameters of the model. While the trained model predicts and forecasts all the SEIR compartments, only the confirmed cases are required for fitting.
Using auto-differentiation on an extended SEIR model, we optimize per-country epidemic parameters to forecast future outcomes of the COVID-19 pandemic. View 1-page Paper
The following charts show the 6-weeks forecast of future confirmed new cases, given the current observed data. Data is daily updated from the following three sources:
Using older data from areas that have been affected earlier, we can train our model up to an older data and use the following days for validation. Note, training the model here means estimating the parameters of the ODE system in a data-dependent manner. The time series is split into training and test intervals. Data from the training interval is used to estimate the parameters and make a forecast for the test interval. The observed data in the test interval is shown in "red" while the matching forecast is on the "blue" line.
Training data up to February 18. Note: despite the peak on February 12, given by a change in the confirmation policy, the model show robustness to outliers.
Validation using training data up to May 1st.
Despite no lockdown intervention was enforced, intervention time is set as February 19, when the reported outbreak caused residents to spontaneously isolating and practicing social distancing.
Validation using training data up to March 26.