When the next new infectious disease begins to race around the world, Ryan Tibshirani hopes to have a completely different way to track and forecast its spread. During the COVID-19 pandemic, public health officials were responsible both for managing the disease on the ground and reporting key indicators — cases, hospitalizations, and deaths — and that led to inconsistent data and errors.
“This system is too slow and error prone. And it’s a huge burden on public health itself,” said the Amazon Scholar and professor of statistics at the University of California, Berkeley. “You don’t want the people who are primarily responsible for dealing with the pandemic to also be responsible for reporting the data that decisions are based on.”
Tibshirani, an Amazon Scholar with Amazon Web Services (AWS) AI Research and Education organization, is also principal investigator of the Delphi Group, a research team based out of Carnegie Mellon University in Pittsburgh that is developing an epidemiological tracking and forecasting system that ingests data streams that operate outside of public health reporting.
For example, the team has agreements to access de-identified medical insurance claims that hospitals file with insurers to get paid for services performed. That data pipeline already exists and reflects disease activity, noted Tibshirani.
“Data streams that exist in the medical records sphere are sustainable and they can be very localized – you can see something happening in a particular spot and time. That can be very informative,” he explained.
That work, along with his body of research, led to Tibshirani being awarded the Committee of Presidents of Statistical Societies (COPSS) Presidents’ Award at the Joint Statistical Meetings in Toronto in August. The award — which goes to a member of the statistical community under the age of 41 and is considered one of the highest honors in the field of statistics — recognized his academic research, including contributions to theoretical statistics, development of new methodology, and contributions at the interface of statistics and optimization.
“It’s a huge honor,” Tibshirani said of receiving the award. “It’s not something that I would have ever thought that I would win or ever dreamed to win. There are some highly distinguished people who have won this award in the past, including my dad.”
Tibshirani’s father, Robert Tibshirani, a professor of statistics at Stanford University, received the COPSS award in 1996. The father-son duo are frequent collaborators today, including on the Delphi Group’s research.
Foundational research
The COPSS Presidents’ Award recognizes Tibshirani’s contributions to the foundations of statistics: The award citation notes his deep contributions to nonparametric estimation, high-dimensional inference, and spline theory.
Nonparametric estimation refers to a class of statistical models that are used to estimate underlying trends in the data without specifying the shape of the pattern or behavior they’re looking for, Tibshirani explained. Neural networks, for example, are nonparametric. High-dimensional inference is when the number of parameters in a statistical model is large and may exceed the number of observations.
The award citation also notes Tibshirani’s contributions to distribution-free inference, which refers to a class of approaches that quantify uncertainty without making assumptions about the model at hand or the underlying data generating process. This is particularly relevant for quantifying the uncertainty of machine learning models, noted Tibshirani.
An Amazon Scholar since March 2020, Tibshirani has worked on methods for distribution-free inference that are being incorporated into AutoGluon, an AutoML toolkit that AWS open-sourced in 2019. His work on ensembling, a machine learning technique where multiple models designed to predict the same thing are combined, has also been incorporated into AutoGluon.
Going forward, Tibshirani said, the epidemic tracking and forecasting work with the Delphi Group will remain an important focus, balanced with his more traditional academic research. The Delphi Group recently received funding from the U.S. Centers for Disease Control for its Outbreak Analytics and Disease Modeling Network, the first national network for this type of research.
Ultimately, Tibshirani said, he wants epidemic tracking and forecasting to be “as trusted and as used as weather forecasting is today. Right now, I think it is very far from that — but I don’t think it needs to be.”
