The pandemic has forced a lot of change in a short period of time. At Embee, we are endeavoring to see if we can pitch in. We’re working with UC Berkeley to build an early warning system that can examine the data we collect from our panel using Berkeley’s artificial intelligence and machine learning capability, combined with its world-leading epidemiology expertise. The goal is to be able to spot outbreaks before they start based on the behavioral cues, app usage, and mobility.
Let’s go through the rationale for this undertaking.
If you think back to the early days of COVID-19, most states were on some kind of lockdown, and movement was very limited. Highways during rush hour looked like a scene from The Walking Dead. Pretty much everyone was at home, and mobile phones weren’t very mobile.
Inevitably, whether we were ready or not, society started to open back up. There was considerable discussion about the risks people would take as they venture out into the world. As an area opens up, how can you mitigate risk?
For example, if you merely drive past a restaurant in your car (assuming you’re by yourself or accompanied by a family member), there’s no risk at all. If you go and sit outside the restaurant, there’s a slightly elevated risk from passersby and fellow diners. Then again, if you go inside the restaurant, there’s an even higher risk because transmission of the virus is more likely indoors. It turns out, knowing the degree of movement, alone, isn’t enough. Instead, it’s about where you go and what you do.
So, while researchers worked to develop therapeutics and vaccines, those of us in technology wondered how we could help.
Could we redirect the tools we use every day to help slow the spread?
Contact tracing, which went from a term no one knew to a buzzword at dizzying speed, became the focus of a number of tech firms. But, useful as it could be in some ways, the idea behind contact tracing wasn’t to prevent outbreaks. Rather, it was to trace an outbreak after it had started. Plus, given the degree to which outbreaks can spread among asymptomatic people, the reality is, by the time it becomes apparent there’s an outbreak, it may already be too late for contact tracing to be effective. As one of the epidemiologists from Berkeley put it to me, once sick people start showing up at hospitals, the “networks already overlap” too much.
So, we asked a question: Could we, effectively, find sick people before they know they are sick? In other words, could we use information about what people are doing to spot patterns and find out whether there’s a heightened risk of an outbreak in a specific area? If so, then we could alert authorities so they can deploy precious testing and medical resources to a pinpointed area to find sick people before they know they are sick, when contact tracing and other mitigation methods can be most effective.
This is what we’re endeavoring to do along with UC Berkeley. We’re working on something of an early warning system—like a buoy out on the ocean that indicates a tidal wave is coming. The work is part of a study (see the announcement here), titled “Social distancing and sheltering in place: Using a nationwide smartphone panel with location data to understand population heterogeneity and inform intervention methods,” led by UC Berkeley researchers Daniel Chatman, Joan Walker, and Daniel Rodriguez.
That kind of timely data is critical so that a community can take action before an outbreak becomes a crisis—or even before it becomes an outbreak.
Embee and Berkeley are working to build a predictive model that combines Embee’s survey and behavioral data (things like indicators of mobility, app usage, and behavioral cues) with Berkeley’s expertise in AI, ML, and epidemiology to predict where outbreaks are likely to happen with the goal of finding sick people before they get sick.
If we suddenly see an uptick in people searching for bars, movie theaters, and restaurants, there’s a high likelihood they’ll end up going to one of those establishments. Looking at people’s intended behavior (in this case, searching) is a great way to anticipate what actual behavior will be. We can also look at other factors, including search patterns, and app downloads and usage.
Looked at in the aggregate, these are leading indicators: Are people going to bars, restaurants, and movie theaters? Are they also searching for COVID symptoms? Are they downloading medical apps? Are they scheduling doctor’s appointments? Embee is working with the Berkeley machine learning and artificial intelligence group to develop AI-based models built on this very data to predict outbreaks and alert local health officials with advice on where to increase testing.
We are early on with this initiative. But we’re hopeful that the results will live up to the ambition we’ve laid out here.
So why are we doing this?
It’s pretty simple. We want to try to make life better for us all.
I don’t know about you. But I’m longing for a return to normalcy—not the new normal we all hear about, but the real normal we all had in February and the months and years before.
One way to do that is to make the existing response systems we have more effective, efficient and surgically precise. By combining UC Berkeley’s expertise and Embee’s data and tools, we can help.
Stay with us as we update you on our progress.