It’s an understatement to say there can be a lot of noise online. Disinformation and misinformation flood the social media channels we’ve become increasingly dependant on in a socially distant world.
But amid the cacophony, sometimes the founding collaborative spirit of the Internet reemerges.
David Yu — team lead of hockey analytics at Sportlogiq, a Montreal-based, AI-driven advanced stats company that works with most NHL teams — is the author of one of those recent precious moments.
He isn’t an epidemiologist. Nor is he an expert in infectious diseases.
But the soon-to-be 33-year-old hockey analyst has created a powerful tool, the COVID Projections Tracker, which has provided those in the field a way to find flaws in the model widely used by U.S. health care providers, media outlets and government bodies — including the White House — to make crucial decisions related to the novel coronavirus pandemic.
The website tracks the daily changes in projections of cases and deaths, among other metrics, produced by the aforementioned University of Washington’s Institute for Health Metrics and Evaluation (IHME) model, as well as that of the Los Alamos National Lab.
Yu — who moved from China to Winnipeg when he was six and whose partner’s family hails from Wuhan — was inspired to pitch in when cases of the virus started to spike in Italy.
He initially used his abilities to create the online platform VolunteerAtlas, which aimed to connect Canadian volunteers with those in need — such as the elderly or immunocompromised — so they can get supplies or groceries.
But when he realized there were others with more software development expertise working on similar offerings, he shifted his efforts.
“I’m really trying to just make it easier for the people that are on the frontlines of — whether it’s epidemiology or volunteer co-ordination — for them to do their jobs,” said Yu, who completed five years of a PhD in biology before changing focus. “I could have maybe tried to build models and things like that, but I think (experts are) actually inundated with people that are helping, but not really helping. And so my goal was always to, rather than to try and build something that was better than theirs, build something that would help them do their jobs.”
And that’s exactly what he’s done with the COVID Projections Tracker.
After doing some initial graphing of the data, which was publicly released in late March, Yu noticed that some of the hardest-hit U.S. states had seen their projected death counts drop suddenly. He then reached out via Twitter to prominent University of Washington biologist Carl Bergstrom — who is not affiliated with the IHME model.
The sudden shift in projected death counts was something Bergstrom hadn’t seen before, and he encouraged Yu to keep looking into it. So the long-time Winnipeg Jets fan — who is used to analyzing aspects of hockey such as pace of play, faceoffs and passing — leveraged his data-science and data-visualization talents to develop a platform that could help Bergstrom and other experts make definitive assessments.
It's been all COVID all the time for a while, but excited to present our comprehensive analysis of pass difficulty, value and tendencies in hockey at #ISOLHAC this Saturday
Team effort with @pauly_p14 @ConnorJungle & @SamForstner
Kudos to @alyssastweeting for organizing https://t.co/BwcT4wbvi0
— David Yu @ (@yuorme) May 5, 2020
Bergstrom compared Yu’s tool to forensic analysis. Only, in this case, it is being performed on a model rather than a crime scene.
“It’s been a very useful resource to … get a better sense of what [the IHME model is] going to be good at predicting and what it’s not going to be good at predicting,” said Bergstrom, who studies infectious diseases and misinformation. “It turned out [the model was doing] a reasonably good job of — better than I was expecting — predicting the peak of the curve in most U.S. states, for example, but it did a dismal job of predicting what happened after that.”
In particular, the COVID Projections Tracker made it plain that the IHME model projected deaths to decrease as quickly as they rose when the virus spread initially — even though that hasn’t been and isn’t likely to be the case. That’s a troubling fact given how prevalent the model has been to date.
“Just based on how much the White House and (U.S. President Donald Trump’s) administration cites this model, clearly, they go to it as a source of truth for their outlook on the future,” said Yu.
Now, Yu’s tool has helped people see the need to stop over-relying on the IHME model — which researchers have recently updated to better account for its faults — and start looking to others.
“It just helped us see clearly that the model was getting things wrong in a systematic, understandable way,” he added. “And so I think it gave me a lot more confidence to be critical of that model’s ability to predict what happens after the curve.”
It’s a feat of remarkable data wrangling and coding from an unexpected source in a moment when experts need all the help they can get.
“I think it’s a really great example … like this is a situation where there’s so much talent out there and we need all hands on deck, and we’re learning in real-time how if we get all hands on deck, we can collaborate and not talk over each other and make forward progress,” said Bergstrom. “And this was one of those cases where it really worked beautifully.”
[relatedlinks]
