KAIST, AI model Hi-COVIDNet
Forecast of the number of confirmed cases imported from overseas over the next two weeks


As the number of confirmed COVID-19 cases worldwide surpasses 20 million, a domestic research team has developed a technology capable of predicting the number of confirmed cases imported from overseas.

A research team led by Professor Jaegil Lee of the Department of Industrial and Systems Engineering at the Korea Advanced Institute of Science and Technology (KAIST) announced on the 19th that it has developed big data and AI technology that can predict the number of confirmed cases of COVID-19 imported from overseas over the next two weeks.

This study was presented on the 24th under the title 'Hi-COVIDNet: Deep Learning Approach to Predict Inbound COVID-19 Patients and Case Study in South Korea' at the 'AI for COVID-19' session of 'ACM KDD 2020', an international academic conference in the field of data mining.

When calculating the COVID-19 risk in each country, the research team primarily used the reported number of confirmed cases and deaths. However, because these figures are dependent on the number of diagnostic tests, they also used the frequency of COVID-19-related keyword searches as input data to calculate the COVID-19 risk in each country in real time.

In addition, the number of real-time arrivals is confidential information and is not disclosed to the public, so it was inferred from the number of flights arriving in Korea and the number of roaming customers entering the country each day. Data on the number of roaming customers entering the country was received from KT, but the limitation of only including KT customers entering the country was resolved by also considering the number of daily flights.

In addition, to predict the number of confirmed cases coming from overseas, geographical connections between countries must also be considered.

The research team designed an AI model called 'Hi-COVIDNet' to accurately predict the number of confirmed cases imported from each continent based on a geographical hierarchy consisting of countries and continents to learn geographical correlations, ultimately accurately predicting the total number of confirmed cases imported from overseas.

The research team used Hi-COVIDNet, a model created using only a short training period of about a month and a half, to predict the number of confirmed cases from overseas over the next two weeks. The results confirmed that this model achieved up to 35% higher accuracy than existing predictive machine learning or deep learning models based on time-series data.

This research was supported by the KAIST Global Strategy Institute's COVID-19 AI Task Force Team and the roaming dataset provided by KT and the Ministry of Science and ICT's COVID-19 Spread Prediction Research Alliance.