The University of Gothenburg is making a bold and pioneering move to revolutionize current epidemic testing and containment strategies using machine learning
“When a large outbreak has begun, it’s important to identify infectious individuals quickly and efficiently. In random testing, there’s a significant risk of failing to achieve this,” states Laura Natali, Doctoral student at Sweden’s University of Gothenburg Physics Department and the lead author of a recent study examining the use of machine learning to find effective screening methods and better control future public health crises.
“We need a more goal-oriented testing strategy,” she adds. “One that can find more infected individuals even when only relatively limited data is available. We believe AI has a huge potential to gain the necessary information helping to decrease the spread of infection.” There’s still some distance before we approach the light at the end of the COVID-19 dark tunnel, with the Delta variant now spreading its dark tentacles into at least 98 countries.
According to the CDC’s estimates, the strain is responsible for about one in four infections in the US and virtually all new cases in the UK. The Delta variant has also infected fully vaccinated individuals and led Israel, Australia and other governments to reintroduce certain social restrictions. While AI-driven solutions have played an active role in mitigating this global health crisis, there aren’t many studies looking into how the technology could be used to contain new waves of infections such as the ones caused by the Delta variant.
“Containment of outbreaks entails great societal and economic costs,” Natali adds. “We want to make the first step in gaining better control over future major epidemics or even pandemics, to reduce the need for widescale lockdowns. However, cost-effective containment strategies rely on diligent contact tracing and identification of infected individuals, but information tends to be limited at the start of an outbreak. Making the best possible use of the available testing resources and being able to predict which individuals should be offered testing become critically important.”
The new strategy proposed by Natali and her research team leverages machine learning to adapt to the specific characteristics of sick individuals automatically and dynamically. The model tells whether screening should be prioritized on an elderly population or a particular geographical area, including neighborhoods and public venues.
“The main assumption here is we can have access to the exact data about a part of the population,” Natali explains. “We simulated an outbreak using the archetypal susceptible-infectious-recovered (SIR) model with data about the first confirmed cases, including who they have been in close contact with, where and for how long, as well as other information that was used to train a neural network that learns to make predictions about the rest of the population.
“We believe such predictions can be used to execute a containment strategy. Rather than hindering the outbreak at the starting point, which is a small group of infectious agents that grow quickly in the absence of implemented measures, the neural network allows targeted detection of infectious agents that help contain the outbreak. We also demonstrated how this method can be used when there’s a possibility of reinfection to effectively eradicate an endemic disease.”
Natali and her research team observed, under a simulated environment, how an outbreak can quickly be brought under control when they deployed their model. Specifically, the epidemic evolution in time is significantly improved and the total fraction of the population infected is reduced by half. Whereas random testing leads to uncontrolled spread with many more infected individuals. Nevertheless, the research team did caution that the present study requires improvement.
“Under real-world conditions, more information such as demographic data, age, and health-related conditions can be added to improve the method’s effectiveness,” Natali adds. “But we do feel that our method is the first step in being able to implement more targeted initiatives to reduce the spread of infection. It’s possible to use relatively simple and limited data to make predictions of who would be most beneficial to test. This allows better utilization of testing resources.”