At a recent HAI conference, experts suggest new market structures for vaccine development and machine learning tools to improve clinical outcomes.
On June 1, 2020, the Stanford Institute for Human-Centered Artificial Intelligence (HAI) held its second virtual conference on COVID-19, bringing together experts from across disciplines to discuss the way forward, with focus on the intersection of social, economic, political, health care, and technological domains.
“This is the first conference series focused on applications of AI to disease tracking and mitigation,” says co-moderator and HAI associate director Russ Altman, Stanford professor of bioengineering, genetics, medicine, biomedical data science, and computer science. Fellow co-moderator and HAI associate director Rob Reich, Stanford professor of political science, adds, “It’s like we’re dealing with the 1918 pandemic and 1968 election year [with its multiple social protests] at once.”
In this context, several speakers addressed the health care road forward, with the focus on epidemiology, predictive models, and vaccine development.
Here are the speakers’ key insights.
Understanding Infection Rates
With ongoing challenges related to COVID testing, it has been difficult to assess the disease’s incidence and transmission patterns.
“We need to take a broad-based, population-level approach to measurement,” says Yvonne Maldonado, Stanford senior associate dean and professor of pediatrics and health research policy. Her team is conducting multiple research studies to examine the patterns.
The first study — currently in progress — examines household transmission rates. Specifically, a sample of those who test positive for COVID at Stanford Hospital are asked to have family members self-test (using nasal swabs with proven accuracy) for 21 days. Results will help scientists, clinicians, and policymakers understand susceptibility and rate of spread at the household level.
A second study, sponsored by the Chan Zuckerberg Initiative, will examine rates of new incidence of disease and antibody presence in two populations: Bay Area residents at large across six counties and local health care workers previously found to be COVID-negative. Researchers are aiming to enroll 9,500 participants and will continue the study to 2020’s end. “This will help us understand what hot spots are emerging over time and whether antibodies protect against reinfection,” Maldonado says.
Finally, Maldonado is helping to oversee a series of outpatient clinical trials for COVID treatment. “We need a toolkit of drugs for outpatient treatment,” she says. The first trial, for Lambda Interferon, is being conducted on-site at Stanford.
Predicting Clinical and Public Health Outcomes
Along with COVID’s epidemiology, researchers are working to predict clinical and public health patterns and outcomes.
One such expert is Eric Horvitz, Microsoft technical fellow and chief scientific officer. “AI predictive models can be really valuable in clinical decision making,” he says. Horvitz has already published work predicting mortality of ICU patients.
On the clinical care side, Horvitz and Microsoft colleagues are collaborating with a team from the NYU Langone Health to test machine learning models including one predicting whether COVID patients will have an adverse outcome (intubation, transition to hospice, others) over the next 96 hours based on vital signs and other predictors. They plan to test the models further, using randomized controlled trials across hospital systems.
On the public health side, machine learning models can support decisions related to reopening, testing protocols, contact tracing and notification, and others. “The predictions are valuable for public discourse, optimization, and decision analysis,” Horvitz says.
So far, his and other teams have worked to describe risk by region, age group, and other factors, to answer such questions as “Is it safe for me to return to work?” They will continue to harness data and predictive models to support clinical and community decisions.
Creating the Right Incentives for Vaccines
Finding a COVID vaccine is paramount.
Susan Athey, Stanford Graduate School of Business professor and HAI associate director, highlights the dual value of this breakthrough: “COVID is not just an economic problem or a health problem. It’s both.” She points to IMF estimates of a $9 trillion pandemic-related loss over 2020-2021. “So accelerating a vaccine by six months could mean gains of over $2 trillion,” she says. “The cost-benefit is indisputable.”
The problem is that incentives for vaccines have resulted in large-scale market failures. Because most vaccines fail, firms don’t want to invest manufacturing capacity until a given product has gained FDA approval. But scaling takes time, and delays can cost trillions, as the figures above suggest.
The answer, Athey says, lies in proper incentive design: “We have to pay firms to invest in capacity at risk, before approval.”
She helped create a simple economic model highlighting the benefits of early installation of vaccine-production capacity. The research modeled the health and social benefits of scenarios with and without incentive programs, showing that accelerating capacity installation yields dramatically larger benefits.
“The best design uses ‘pull’ incentives to pay firms for positive outcomes and guarantee vaccine purchase prices and ‘push’ funding to reimburse up to 85 percent of capacity installation costs,” Athey says, also suggesting price paid per dose could be lowered over time, to motivate speed.
Athey’s research also used data on vaccine candidates and success probabilities to determine the optimum number of vaccines to invest in: about 18. “We need all shots on goal,” Athey says.