Treating COVID-19: How Researchers Are Using AI to Scale Care, Find Cures, and Crowdsource Solutions
The latest technologies are helping researchers and health care practitioners find vaccines and scale care. Photo | Getty Images
In the fight against COVID-19, researchers are using everything from crowdsourcing to rapid gene-sequencing to intelligent health care delivery. Powering much of these efforts are artificial intelligence and analytics.
On April 1, the Stanford Institute for Human-Centered Artificial Intelligence (HAI) held a virtual COVID-19 and AI Conference, featuring speakers discussing how best to understand, treat, and prevent the coronavirus by means of the latest technology-enabled tools.
Crowdsourcing Coronavirus Solutions
Kaggle, the world’s largest machine-learning community, with 4.4 million members, is meeting the COVID challenge with challenges of its own: specifically, open competitions to create machine learning (ML) systems that facilitate scientific literature review, forecasting, and, data set curation related to the coronavirus.
“We’re working with the White House on using the expertise of our community against COVID,” says Kaggle founder and CEO Anthony Goldbloom.
The literature-review challenge, for example, will make the 44,000 research articles already published on COVID more accessible, enabling users to ask and answer specific questions on biochemistry, diagnosis, and other topics. The forecasting challenge aims to unearth the factors most predictive of cases, as well as fatalities overall and by region (using Johns Hopkins University data).
The initiatives are quickly generating takeaways available to the public.
Automating and Scaling Care
Meanwhile, multiple businesses are developing scalable intervention solutions. One is Curai, a startup building a medical platform for doctors and patients to communicate, with AI-enabled capabilities at both patient and provider ends.
“Our platform is chat-based,” says Xavier Amatriain, Curai co-founder and CTO. “It features personalized diagnostic assessment, such as asking health care workers if they are likely to have been exposed to COVID-19.” The solution also offers daily follow-up and inputs into treatment plans, in partnership with health care systems. Curai built the system using machine learning and human-expert input, in part to get people the care they need at home rather than by taking risky trips to health care facilities.
One of the pandemic’s highest-risk populations is the elderly, who may not recognize early signs of illness but may also take unnecessary trips to medical providers, potentially exposing themselves to the virus and burdening the health system.
“Older people are most at risk and are passing away without proper help,” says Fei-Fei Li, Stanford computer science professor and Denning Family Co-Director of HAI. “We’re trying to help them stay home while monitoring them to detect illness.”
She’s part of a team of researchers proposing AI-powered, in-home technology that can collect data through cameras and thermal sensors, then identify clinically relevant patterns to alert users, family members, and caregivers.
Thermal sensors can detect small temperature rises signaling a fever, for example. Other elements of the system could identify falls, immobility, or even dietary changes. Still in design stages, the technology isn’t meant to replace clinicians and caretakers, but to complement their care.
Protecting Frontline Professionals
A large part of the COVID challenge is delivering care to infected patients while protecting frontline health care workers. Among such efforts is a push for more “intelligent” care delivery, including using ML-enabled information and processes.
“Most patients aren’t critically ill when they come in,” says Ron Li, Stanford clinical assistant professor of medicine, hospital medicine, and biomedical informatics. “But they can quickly worsen and end up in the ICU. That constrains resources and puts staff at risk, from labor-intensive processes like intubation due to aerosolization of the virus.”
He is part of a group working to semi-automate identification of at-risk patients before they need critical care, by mapping processes with frontline physicians and feeding these into an ML-based model that will generate a predictive score to complement human judgment.
Higher-risk patients will be subject to specific workflows, for their benefit and that of the staff. The Stanford system is already performing well, with ongoing validation.
Developing Drugs and Vaccines
Multiple ongoing efforts are aimed at developing COVID treatments and vaccines.
In this context, sequencing the genome of the virus behind COVID-19 is critical, because this contains molecular targets for diagnostic tests, vaccines, and antiviral drugs.
For example, IBM’s Functional Genomics Platform studies microbial life at scale, as described by Kristen Beck, IBM lead bioinformatician in the AI and cognitive software group: “We have a database that covers more than 300 million bacterial sequences, connecting genotypes [genetic material] with phenotype [how that material is expressed, including as disease].”
As COVID has spread, IBM has used the platform to map the underlying genomic profile of SARS-CoV-2 (the specific coronavirus), including how that compares with those of SARS and MERS, to derive treatment implications. IBM offers cloud-based genomic data and tools to enable researchers to use the information in AI-based and other approaches.
It’s important to recognize that novel coronavirus may not require a novel treatment agent. Instead, existing drugs may be repurposed to fight COVID-19. But that involves quickly understanding the right therapeutic target — or what protein to bind or block with an existing compound.
Stefano Rensi, a Stanford bioengineering research engineer, is part of a team carrying out this mission. The group first used a natural language processing tool to map relationships between chemicals, genes, and diseases in the literature, to home in on the mechanism of action to pursue.
From the research, the TMPRSS2 protein emerged as facilitating viral invasion of cells. So the team conducted docking simulations to predict binding to that protein for 18 molecules. The top-scoring compounds were then compared with existing drugs, ideally those already approved for specific indications or cleared for testing in humans. One of the compounds identified is already undergoing clinical testing for COVID-19 in Japan.
Most agree that the “holiest grail” in the fight against COVID-19 will be an effective vaccine. While some regions have tried a tuberculosis vaccine, there is no evidence of efficacy.
Finding an effective vaccine requires identifying the immunogenic components of coronavirus, to understand how the virus enters human cells — and machine learning can help. “B-cells and T-cells can destroy infected cells to stop the virus,” says Stanford genetics MD/PhD student Binbin Chen, who’s part of a research team analyzing fragments of the SARS-CoV-2 virus. “We’ve used a neural network [ML-based approach] to predict what sites on the virus for T-cells to bind to. Some of our best vaccine candidates look like SARS antibodies.”
The group continues to seek the best vaccine candidates, developing a growing list.
Missed the conference? Watch the conference recording, and check the HAI blog for additional in-depth coverage. Also, sign up for HAI’s email newsletter to learn about upcoming events.