A hybrid approach to predict COVID-19 cases using neural networks and inverse problem

We derive a novel hybrid approach, a combination of neural networks and inverse problem, in order to forecast COVID-19 cases, and more generally any infectious disease. For this purpose, we extract a second order nonlinear differential equation for the total confirmed cases from a SIR-like model. That differential equation is the key factor of the present study. The neural network and inverse problems are used to compute the trial functions for total cases and the model parameters, respectively. The number of suspected and infected individuals can be found using the trial function of total confirmed cases. We divide the time domain into two parts, training interval (first 365/395 days) and test interval (first 366 to 395/ 396 to 450 days), and train the neural networks on the preassigned training zones. To examine the efficiency and effectiveness, we apply the proposed method to Canada, and use the Canadian publicly available database to estimate the parameters of the trial function involved with total cases. The trial functions of model parameters show that the basic reproduction number was closed to unity over a wide range, the first from 100 to 365 days of the current pandemic in Canada. The proposed prediction models, based on influence of previous time and social economic policy, show excellent agreement with the data. The test results revel that the single path prediction can forecast a period of 30 days, and forecasting using previous social and economical situation can forecast a range of 55 days.