Teams turn to AI to tackle the spread of Covid

Since its launch 25 years ago, the XPrize has become synonymous with ‘moonshot’ contests. Contests designed to crowdsource technological solutions to Earth’s most complex and huge-scale problems – deforestation, illiteracy, climate change. All with multi-million dollar incentives.

Yet behind the headlines dominated by $100m (£72m) prize pots, put up by Elon Musk, the XPrize is now using its crowdsourcing approach to make a more immediate difference to everyday lives. Through smaller prizes, aimed at tackling problems that need more pressing solutions, the XPrize has proved it is able to mobilise the world’s greatest minds more quickly. Those minds were recently called upon to tackle the ongoing pandemic.

When Covid took hold, think tanks, businesses, scientists, and experts began gathering data: data on symptoms, data on deaths, data on the effectiveness of masks. It wasn’t known which data would prove useful in the long term, or which sets were best to model algorithms and predictive models on.

In an attempt to make sense and use of this crucial data, a team of evolutionary AI specialists at New Jersey-based Cognizant built a baseline predictions and prescriptors platform. Based on data captured by Oxford University – one of the earliest sets to include policy decisions alongside numeric facts and figures – a Cognizant team, led by Babak Hodjat, created a blueprint. It used AI to extract insights from the data, before using these insights to predict various future scenarios.

The team could then see how prescriptors, such as hypothetical policy decisions made in response to these different scenarios, would potentially play out. When it was built, Hodjat’s team published the platform. The aim was to grab the attention of anyone working in the field, to crowdsource knowledge, but it ended up grabbing the attention of the team at XPrize.  

“XPrize had already been thinking of launching a prize to apply AI for the good of humanity,” Hodjat tells E&T. “They believed our platform presented a blueprint to do this: to use AI to address issues surrounding the coronavirus. Yet with the virus evolving rapidly and in real time, building such a system would be like building the engine of the plane as you’re flying. We’d need to crowdsource the ideas and collaborate on them if it was to work.”

With this in mind, on 30 October 2020 registrations for the $500k (£360k) Pandemic Response Challenge went live. With it, the challenge became the first competition under XPrize’s AI and Data for Good Alliance arm.

Using Cognizant’s blueprint, with Oxford Covid-19 Government Response Tracker data, the challenge asked teams to build AI models that could predict local Covid-19 transmission rates. The teams would need to make recommendations based on these predictions, and the winning team would be the one whose predictions and prescriptors were the most accurate.

By the start of Phase 1, on 17 November, more than 100 teams from 17 countries had submitted entries using a range of approaches. Some opted for simulation-based epidemiological modelling mixed with classical machine learning, others combined probabilistic mixture models with decision-tree algorithms. Another team used machine learning to support human expertise.

The beauty of this phase was that the advisory board and judges, which included epidemiologists, biostaticians, government ministers from Europe, and US policy makers, could test the accuracy of the models by seeing predictions play out in real time. Each team would make regular predictions about how well different measures would reduce infection rates and impact economies, and these predictions would be tracked using real-world data. The 48 teams that most accurately predicted what later played out in real-life were advanced to Phase 2.

In Phase 2, each team’s models were run through a reality simulator to test their predictive powers further. “Following the first phase we’d developed a pretty reliable predictor,” continues Hodjat. “We knew what the error bounds were for a huge array of scenarios, and this allowed us to create a simulation that was effectively a substitute, or surrogate, for reality. We used this simulation to input multiple hypothetical predictions, and run them simultaneously.”

This allowed judges to test millions of ‘what-ifs’ from the top all the way down to a granular level. What if workplaces closed, schools opened, and mask wearing was relaxed? What if workplaces opened, schools closed, and people were told to wear gloves and masks? What if the Rule of Six became the Rule of Seven, or if pubs only opened on Tuesdays, or if people wore two masks instead of one? And so on.

This not only allowed the judges to pit one team’s predictions against another, but it meant the simulations could explore myriad combinations of different teams’ measures. It could effectively pull individual parts from various models together to create new, more accurate models, making the final results greater than the sum of their parts.

The challenge ended on March 9 2021 with two teams sharing the prize: first place went to the Valencia IA4Covid19 team (pictured above), coordinated by Dr Nuria Oliver, commissioner of the presidency of the Generalitat Valenciana. In second place was JSI vs Covid from Slovenia, coordinated by Mitja Lustrek, head of the Ambient Intelligence Group at the Jozef Stefan Institute

The former has been collaborating with the Valencian government on its Covid policies since March 2020. Its model uses deep neural networks that account for both the number of cases and the different interventions used by governments across all 230 regions of the challenge. It then used machine learning to make its recommendations.

JSI vs COVID predicted infections using a traditional SEIR epidemiological model – in which populations are given labels S, E, I, or R, (Susceptible, Exposed, Infectious, or Recovered) – with machine learning.

The winning models will now be made available to policymakers, business leaders and health officials.

“It’s been amazing to see the levels of motivation of the team to address a common challenge, giving the best of each of us,” Dr Oliver tells E&T. “The biggest challenge was building both predictors and prescriptors for 230+ countries/regions in such a short time period. We are focusing now on expanding our work to take into consideration the vaccination campaigns. We hope our work will inspire governments and public administrations to shift towards more evidence-driven public policy-making.”

Amir Banifatemi, chief innovation and growth officer of XPrize adds: “It’s been a year since the beginning of the Covid-19 outbreak, and we’ve sought out various ways in which XPrize can help identify and accelerate solutions. Our hope now is to implement the technologies into emergency response plans for local governments worldwide, utilising AI to better protect everyone against future pandemics.”