KIT: Significant effect of school closures
11.12.2020 09:06
Before enough people are vaccinated, non-pharmaceutical interventions (NPI) - such as reducing social contacts - remain necessary to protect the population as best as possible from Covid-19 infection. Researchers at the Karlsruhe Institute of Technology (KIT) used machine learning methods to investigate the effectiveness of NPI. Their analysis of data from nine European countries and 28 US states shows that early school closures in particular can be the cause of the trend reversal in daily case numbers. The preprint of the study, which will be published in the European Journal of Information Systems, can be read at https://publikationen.bibliothek.kit.edu/1000126905.
The scientists at KIT's Karlsruhe Service Research Institute (KSRI)/Institute of Information Systems (IISM) examined the daily case numbers from Austria, Belgium, collected by the US Johns Hopkins University between January 22nd and May 12th, 2020. Germany, Italy, Norway, Spain, Sweden, Switzerland and Great Britain as well as 28 US states.
Significant influence of school closures on the trend reversal in the number of cases
The researchers looked at four measures: general restrictions on gatherings, reducing personal social contacts, school closings and a lockdown. Country-specific characteristics such as age structure, population density, medical infrastructure and the climate in the respective countries and states were also taken into account. “According to our research approach, a significant effect on the time between the NPI decision and its impact was identified in the data for school closures,” says Dr. Niklas Kühl, head of the Applied AI in Services Lab at KSRI/IISM. The sooner the schools were closed, the clearer the effect of falling case numbers became, said Kühl. “If we had waited a day longer in Germany in the spring before closing the schools, according to our analyses, this would have meant 125,000 additional infections, and the closure seven days later would have even meant 400,000 additional cases,” says the business information scientist.
The researchers found that it takes an average of two weeks for a measure to take effect. However, the result of the effectiveness of school closures does not mean that other NPIs or factors that were not included in the model could not also have a significant influence on containing the pandemic, the scientists said in their publication. The wearing of masks was not analyzed in the study because this measure was mostly introduced late in the countries examined. All research related to measuring the effectiveness of various NPIs involves a high degree of uncertainty, the researchers said. This is particularly due to the fact that the specific implementation varies greatly depending on the country or federal state and the population adheres to the measures in different disciplines.
AI recognizes influence on case numbers
The scientists Lucas Baier, Niklas Kühl, Jakob Schöffer and Gerhard Satzger from KIT used concept drift detection to make fact-based statements about the effectiveness of measures against the corona pandemic. This method from the field of machine learning is used to detect a structural change - a so-called drift - in a data-generating environment. “When the question of when a measure affects the number of cases was publicly discussed in the fight against the corona pandemic, we decided to use procedures from this area to determine this,” says Kühl. Concept drift detection is usually used to make AI robust: The aim is to ensure that the AI, trained with existing data, recognizes when the environment changes in action. This means that it should continue to work accurately even with input data that differs from what has been learned. The scientists used the method to detect a fundamental change in the data and correlate it with the previous introduction of an NPI. “Our analysis shows how important a timely response to the spread of the pandemic is in order to keep active cases at a manageable level,” write the Karlsruhe scientists. With their study they want to contribute to the fact-based basis for national and supranational health policy decisions.
Original publication:
Lucas Baier, Niklas Kühl, Jakob Schöffer, Gerhard Satzger: Utilizing Concept Drift for Measuring the Effectiveness of Policy Interventions: The Case of the COVID-19 Pandemic. Published in: European Journal of Information Systems.
Preprint at: https://publikationen.bibliothek.kit.edu/1000126905