![A screen from the COVI app, on top of a colorful background.](/_next/image?url=https%3A%2F%2Fimages.prismic.io%2Fthedecisionlab%2Fd75a815c-f596-4afd-b669-7e35fc7dd724_Screen%2BShot%2B2023-04-19%2Bat%2B10.24.04%2BAM.png%3Fauto%3Dcompress%2Cformat&w=3840&q=75)
Flattening the Curve of COVID-19 With AI and Behavioral Product Design
![read time - icon](/_next/image?url=%2F_next%2Fstatic%2Fmedia%2Ficon__clock.6914af7a.png&w=48&q=75)
0 min read
In March 2020, the COVID-19 pandemic was only just beginning. Almost overnight, countries around the world went from business as usual to a state of total shutdown. Suddenly, the simplest facets of daily existence — things like going to the store, getting on the subway, meeting up with a friend — were potentially laced with danger.
In this atmosphere of uncertainty and fear, governments were desperately trying to design solutions that would eventually allow people to return to their normal lives, while still mitigating risks related to COVID. But at the time, we still had very little information about how the novel coronavirus affected people and spread through the population.
What’s more, different parts of society faced vastly different levels risks from COVID. Children, older adults, and people who are immunocompromised are all more at risk of severe health outcomes from the coronavirus, including death. This made it much harder for governments and public health organizations to chart a course forward: different people required individually tailored guidelines.
Teaming up with Mila
Mila is a Montreal research institute that focuses primarily on machine learning. Founded by the renowned computer scientist Yoshua Bengio, Mila is globally recognized for its contributions to our understanding of artificial intelligence (AI).
When lockdowns began, the team at Mila envisioned building a state-of-the-art contact tracing app to help ordinary people navigate the “new normal” of pandemic life. Developed at the intersection of epidemiology and machine learning, the COVI app would be able to inform users when they had potentially been exposed to the virus. Not only that, but it would also be able to predict each user’s personal risk level in a variety of everyday situations, based on their demographics, health history, and contacts, to name just a few factors.
TDL partnered with Mila and Libéo, a leading software developer in Quebec, on a joint project to develop COVI. Because of the urgency of the situation, the project needed to be lightning-fast: COVI came together in a matter of weeks, the product of intense collaboration and many long nights. Our work was covered by major news outlets across Canada, including the CBC.
Tailored guidance for every user
As we were working on COVI, there were a slew of other contact-tracing apps in development worldwide, as every developer and their grandmother raced to come up with workable tools for fighting COVID-19.
What made COVI stand apart from its contemporaries was its focus on gradations of risk. A machine learning approach allowed us to develop an app that would integrate rich levels of information into a graded score that would in turn let users understand their own personal risk level in any given situation. Armed with this information, users could then make evidence-based decisions about what actions to take, depending on their personal risk tolerance and needs.
Comprehensive behavioral research
As behavioral leads on the COVI project, TDL’s role was to behaviorally optimize the interface that users would be seeing. We were responsible for figuring out the most pressing decision-making challenges citizens were facing. How could we channel the insights afforded by machine learning to provide them with something truly helpful, especially in a time where emotions and stress were running so high?
To answer those questions, we administered repeated pulse surveys to thousands of Canadians, tracking how their attitudes changed over the first few months of the pandemic. We asked participants not just about COVID itself, but about their beliefs and attitudes towards the government, digital mental health apps, tracking apps, and so on. This data let us identify potential pain points and other pitfalls — as well as nudges and other behavioral interventions that could be built into the interface to help alleviate user concerns.
In addition to our user research TDL also conducted an ecosystem scan and a behavioral literature review, identifying evidence-based best practices to be followed in the app’s UX design. Finally, we took care of the project’s complex stakeholder management, mapping the needs and preferences of the numerous stakeholders involved: citizens, various levels of government, and specific communities who were especially impacted by the COVID-19 pandemic.
Earning users’ trust
When users’ health data is in play, privacy is paramount. As we’ve learned from our other work in the digital health and wellness space, users tend to be wary of any platform that asks them share personal information — and for good reason. That’s why the team was careful to meet the highest standards of data privacy and protection when developing COVI.
On the back end, COVI’s system was designed so that all user information was fully encrypted. Data was anonymized and only analyzed in aggregate, to ensure that no individuals would ever be identifiable. The app also minimized the collection of user data to only the essentials, and deleted user data after 30 days.
Protecting the public
By mid-2020, COVI was under serious consideration by the Government of Canada to become Canada’s official contact-tracing app. Although our proposal was ultimately not retained, the codes and models used to develop COVI are still available in open source on the Mila website.
AI and behavioral science both hold immense potential for public health. It’s our hope that by making COVI’s source code accessible to all, we’ll be helping other researchers and governments to develop similar AI-enabled health apps in the future.