(299a) Mechanistically Inspired Data-Driven COVID-19 Pandemic Modeling for Multiple Countries

The COVID-19 pandemic has quickly spread over 200 countries and regions in the world after the first wave of the pandemic in China was almost contained in early March 2020. The proportions of deaths in several countries or regions are high, such as Italy, Spain, and the State of New York, while in other countries such as Korea, Germany, and Hong Kong the proportions of deaths remain low. This situation presents a challenging task for the pandemic analysis and predictions.

The epidemic process in a region is composed of two sub-systems. One sub-system is the infection process from susceptible to infected cases, which depends heavily on the contagious characteristics of the coronavirus, personal protection measures, and social distancing. The other sub-system is the curing process, which depends on the region's capacity and effectiveness of the health care responses. The two subsystems are interconnected. If a high number of infections and a high number of severe cases happen simultaneously in a region, it will strain the capacity and reduce the ability of the region to provide adequate medical and societal responses. Therefore, it is important to predict the development of the infection process as well as the curing process of a region.

Traditional epidemic models include the classic susceptible-infectious-recovery (SIR) model, which is a compartmental model analogous to the reaction system in a batch reactor. An extension of the SIR model is the susceptible-infectious-recovery-dead (SIRD) model, which includes the death process that parallels the recovery process.

Data from China since Jan. 20, 2020 contains a complete cycle of the infection and curing dynamics. However, many other countries are still undergoing rapid changes in the pandemic, with relatively few days of data to allow for reliable modeling. The data from China shows that, for the COVID-19 pandemic, the dynamics of the recovery process is significantly slower than that of the death process. This evidence makes the traditional SIRD model inadequate, since the SIRD model leads to identical dynamics for the recovery and death processes as will be demonstrated in this work.

Our proposed approach in this work is to develop data-driven models with their structures inspired by a mechanistic second-order SIR-type of models, which includes the infected and intensely cared critical cases as inventories in the curing process. For the infection sub-process we use multi-layer perceptron neural networks to fit and extrapolate the data. This approach is able to account for the intensively cared patients which not only have a higher death rate, but also are separated from the population to stop further infections. This model is referred to as the susceptible-infectious-recovery-critical-dead (SIRCD) model. We show that the model is suitable for the COVID-19 pandemic data since it yields different dynamics for the recovery and death cases. We discretize the SIRCD model to give rise to a data-driven model and estimate the model relations from real data.

We use data from Mainland China to build models to predict data from other regions over a period of 14 days. We build input-output causal models from daily confirmed cases to predict daily deaths and recovered cases. Although data from Mainland China might arguably not contain all infected cases, we believe that the recovered and death data for the confirmed cases are accurately recorded. The proposed models for the recovery and death processes are analogous to chemical reaction processes, each of which has three to four free parameters and captures the essential dynamics. The recovery and death models seem to have very different dynamics, with the recovery process much slower than the death process. The models from Mainland China indicate that it takes about 12 days for the first thrust of recovered cases and 2 days for death cases to show sizable numbers. It takes about 60 days to complete the recovery process or 30 days to complete the death process from the date of confirmation.

Although data from Mainland China show different characteristics between Wuhan and the rest of China, the overall models predict well for the regions we tested, including Mainland China, U.S., Canada, Germany, and Spain. However, the death model estimated using data from China gives poor predictions for the death data from Italy and Lombardy. We build another death model using data from Italy for the predictions in Italy and Lombardy. We perform 14 days of predictions, which gives enough lead time to estimate the inventories for planning of heath care needs. We observe that, even though the steady state death rates are different among the regions, the 14 days of predictions are satisfactory since it is only a fraction of the time to steady state for the recovery and death processes.