In August, in partnership with the Harvard Global Health Institute, Google launched a set of models — the COVID-19 Public Forecasts — that provide projections of COVID-19 cases, deaths, ICU utilization, ventilator availability, and other metrics for U.S. counties and states. Today, the two organizations released what they claim are significantly improved models — trained on public data from Johns Hopkins University, Descartes Labs, the United States Census Bureau, and elsewhere — that expand beyond the U.S.
The COVID-19 Public Forecasts are intended to serve as a resource for first responders in health care, the public sector, and other affected organizations, Google says. The forecasts allow for targeted testing and public health interventions on a county-by-county basis, in theory enhancing users’ ability to respond to the rapidly evolving pandemic. For example, health care providers could incorporate the forecasted number of cases as a datapoint in resource planning for PPE, staffing, and scheduling. Meanwhile, state and county health departments could use the forecast of infections to inform testing strategies and identify areas at risk of an outbreak.
When initially launched, the COVID-19 Public Forecasts included regional predictions for 14 days into the future. The model, which learns from epidemiological human prior knowledge, as well as data, is now roughly 50% more accurate and includes projections for a 28-day horizon with confidence intervals to account for uncertainty.
Google says it is investigating support for other countries as it introduces the COVID-19 Public Forecasts for Japan. As in the U.S., the forecasts are free and based on public data, such as the COVID-19 Situation Report in Japan. The daily-retrained model predicts confirmed cases, deaths, recoveries, and hospitalizations each day and can look 28 days into the future for every prefecture.
Beyond these improvements, Google says it has made the initial forecasting models customizable to new problems and datasets. The company is also developing an AI-driven “what-if” model to be used for decision-making around COVID-19 and other infectious diseases.
“We partnered with a handful of early testers, including HCA Healthcare, to help us understand how the forecasts should be formatted, what they should forecast, and even test early versions of the forecasts,” Google Cloud AI research head Tomas Pfister wrote in a blog post. “These efforts helped improve the forecasts before they went to the general public. We also exposed the work to significant scientific scrutiny inside Google, having statistical and epidemiological experts vet the work to make sure it was following the highest scientific standards. We designed a responsible daily forecast launch process that first runs over 100 sanity checks looking for any abnormalities, and we required a human to do a qualitative analysis to check for issues. Every day our model training searches over hundreds of hyperparameter options, and the team works to ensure the best models reach our users.”
Pfister says Google also worked with fairness and ethics experts internally to run a fairness analysis, looking at how both relative and absolute errors differ across demographic groups (particularly Black and Latinx populations) and interpreting the results.
Over 100 employees across Google parent company Alphabet contributed to the development of the COVID-19 Public Forecasts.
Best practices for a successful AI Center of Excellence:
A guide for both CoEs and business units Access here