A team of researchers from the Data Mining Lab in the Graduate School of Knowledge Service Engineering have developed “Hi-COVIDNet,” a deep learning AI system that predicts the number of inbound COVID-19 patients entering Korea from overseas. The team, led by Professor Jae-gil Lee and PhD student Minseok Kim, presented their research at the AI for COVID Virtual Conference hosted by the Special Interest Group on Knowledge, Discovery and Data (SIGKDD) of the Association for Computing Machinery on August 24.

While other recent AI studies focused on predicting the spread of COVID in local regions, Hi-COVIDNet, which stands for “a Hierarchical model that estimates the inbound COVID-19 cases using neural Networks,” is the first software that strives to predict the trend of imported cases from overseas.

This prediction model is a neural network composed of two layers: the country-level encoder considers the temporal trend of COVID within a country and the continent-level encoder captures the geographical context among countries within a continent. It accounts for the fact that the infection risk of a country depends on a complex, spatio-temporal relationship.

To test the model, a case study was performed by inputting data from both intra-country (inside Korea) and inter-country (between Korea and other countries) records, which included  confirmed cases and deaths in foreign countries, search keywords related to COVID, amount of international roaming , and number of flights into Korea from the respective countries. After undergoing a training period from March 22 to May 5, Hi-COVIDNet was tested in the subsequent 14 days, resulting in a prediction that closely resembled the actual trend of cases. Compared to other existing prediction algorithms, Hi-COVIDNet received the lowest RMSE (Root Mean Square Error) score and achieved about 35% higher accuracy, highlighting its superior hierarchical architecture.

Since cases from overseas may constitute a major source of COVID transmission within the nation, it is necessary to accurately predict the number of these imported cases. Hi-COVIDNet is expected to facilitate border management  by providing a guide to allocating limited resources to quarantine centers and even possibly to help determine immigration restrictions by identifying which countries pose a high risk to virus containment.  

Currently, the KAIST team is working with the Korean government to effectively utilize Hi-COVIDNet and plans to extend its use beyond Korea and COVID-19.