리서치 하이라이트

Public health: Improving COVID-19 border control with reinforcement learning

Nature

2021년9월22일

Eva, a reinforcement learning system that led to efficient and targeted SARS-CoV-2 testing of travellers arriving in Greece in the summer of 2020, is described in Nature. Eva is reported to identify nearly twice as many asymptomatic, infected individuals than would be expected for random testing or entry restrictions based on population-level epidemiological measures. The system also provided early warnings for high-risk regions, guiding government border control policies to reduce the spread of SARS-CoV-2.

Since the first wave of the pandemic, many countries have restricted non-essential travel to reduce the spread of SARS-CoV-2. A range of ad-hoc border controls have been adopted, such as only accepting travellers who tested negative for infection, enforcing quarantine periods or blocking travel from some countries. Many nations based these measures on population-level epidemiological metrics such as case or death rates; however, such metrics may be imperfect owing to underreporting, reporting biases and reporting delays.

Greece took a different approach to control the influx of infected travellers in summer 2020, implementing a reinforcement learning algorithm that predicts which travellers should be tested for infection. The system, nicknamed Eva, uses demographic data (country, region, age and gender) collected from travellers along with results from exploratory testing and previously tested travellers to estimate COVID-19 prevalence within certain demographics (or types of traveller), Kimon Drakopoulos and colleagues explain. Using these estimates, Eva identifies a subset of travellers for PCR testing based on their traveller type; that is, if certain demographics are found to have an increased prevalence of SARS-CoV-2, travellers that fit this profile will be tested. To prevent ‘blind spots’, the system also allocates some tests to traveller types for whom data are limited, a crucial feedback step that improves how the algorithm continues to learn.

This reinforcement learning approach was found to identify 1.85 times as many asymptomatic, infected travellers than expected for randomly allocated tests; completely random surveillance would have identified around 54% of the cases that Eva caught. Eva identified 1.25–1.45 more asymptomatic, infected travellers than expected for testing policies based on epidemiological metrics. Eva’s estimates of COVID-19 prevalence were also used to provide early warnings for high-risk regions, which the Greek government used to adjust travel protocols by grey-listing these nations (requiring travellers to have negative PCR test results before entry). The authors estimate that Eva prevented an additional 6.7% of infected travellers from entering the country through its early grey-listing decisions in the peak season. The results suggest that country-wide bans of incoming travellers based on population-level epidemiological metrics may not be the most efficient approach to opening up safe travel, the authors conclude.

doi: 10.1038/s41586-021-04014-z

영어 리서치 하이라이트