August 21, 2020
When I first heard about a “novel coronavirus” originating out of China, I wasn’t particularly concerned. I kicked back and waited for artificial intelligence (AI) to make COVID-19 a history lesson. Since the 1918 Spanish Flu, which took some 50 million lives, I had seen AI discover planets, unlock quantum theories, and beat the greatest human minds at our most challenging games. So, I naively assumed that AI could also beat a microscopic virus. As it would turn out, I was wrong. Twenty million people have been infected, there is yet to be a proven vaccine, and I find myself in a post-apocalyptic mask-wearing world.
I’ve adjusted my expectations: AI is not the cure for COVID-19. Not today, not tomorrow, and not next year. The reason why is that a pandemic involves too much real-world complexity for current AI systems to handle.
Take contact tracing, for example. Government and pharma’s big idea was that AI could sort out who else a person who contracted COVID-19 may have infected, and then keep the healthy population separate from the sick. But an AI with those capabilities can’t be trained on data from past COVID-19 pandemics because there weren’t any. Since the only available training data is on-the-fly -- most of which is unreliable because up to 80 percent of COVID-19 carriers are asymptomatic -- accurate AI-driven contact tracing is hard to come by.
A similar challenge arises when using AI to develop a COVID-19 vaccine. AI can traverse thousands of research papers and analyze millions of chemical compounds to identify one or many potentially effective vaccines. However, an AI’s outputs are only conjecture. Until they are rigorously tested in the lab and in human patients (a process that usually takes upwards of 10 years), their output is virtually meaningless. AI insights can guide pharmacology, but today, it’s still humans who are building vaccines.
Given the problem-solving power of AI, it makes sense to look there first for solutions to our most pressing issues. But AI is not yet where it needs to be to cure COVID-19 on its own. A pandemic is too dynamic and a virus is too uncertain to be captured in an algorithm’s scope. Still, even if it’s not the be-all-end-all, AI can complement human efforts and bring us closer to containing and eradicating COVID-19. Over the past few months, the AI community has rallied to use AI in focused and effective ways.
On March 16, the White House issued a call to action to AI experts to help scientists answer fundamental questions about COVID-19. To fuel subsequent initiatives, the White House also released the COVID-19 Open Research Dataset (CORD-19), a collection of over 90,000 machine-readable studies on COVID-19. At the time, there was a prevailing information crisis in the research field: COVID-19 literature was being created too quickly for scientists to keep up and avoid making errors or duplicates. After all, 90,000 papers is one hefty summer reading list. Researchers needed some way to filter studies and isolate specific data, and AI was the solution.
Let’s look at a question like “What is known about COVID-19 risk factors?” Data about how climate affects transmission or how socio-economic status impacts infection rates can be hidden in the margins of any research paper. So, you need an AI that can understand, analyze, and classify large chunks of human language, a capability known as natural language processing (NLP). By encoding the linguistic hierarchies that dictate how words relate to each other, NLPs can deduce context and pick out bits of information from even massive datasets like CORD-19. This is the same technology that allows Amazon Alexa to interpret your verbal commands and Google Translate to decipher your Spanish homework. A NLP can learn to recognize text related to COVID-19 risk factors, such as “pregnancy” or “co-infection,” and then return a concise dataset on patient susceptibility and transmission dynamics for use in policy-making or future research.
AI experts at Lawrence Berkeley answered the White House’s call with their NLP-driven search engine COVIDScholar. The online tool allows users to comb through CORD-19 by entering basic keywords and applying filters to narrow their search. The Allen Institute for AI developed a similar product, SPIKE-CORD, with the goal of making the power of NLP accessible to those without programming experience. Their search tool allows not just for the retrieval of CORD-19 papers, but also for extraction of information from them, using a simple query language.
So AI can filter through a dataset and isolate information. That’s helpful, but it still doesn’t leverage AI to its fullest capacity. Rather, AI makes its most far-reaching contributions to the fight against COVID-19 when NLP insights are applied to detecting the virus, limiting its spread, and researching potential vaccines.
At 3:18 AM on December 31, 2019, Epidemic Intelligence from Open Sources (EIOS) picked up an article release citing a unique cluster of pneumonia cases in Wuhan, China. That report would go on to be the first record of COVID-19, obtained by a NLP that gathers and analyzes data from health monitoring programs and medical databases. With millions of pieces of online medical material, it’s a challenge to find what may be the epicenter of a global pandemic. But the EIOS’s AI identified COVID-19 at its outset and gave policy-makers additional time to prepare and respond. Today, EIOS is informing the World Health Organization of new COVID-19 outbreaks, and may help diagnose and neutralize the next global pandemic before it becomes a pandemic in the first place.
Other organizations are examining ways to use AI to limit the spread of COVID-19, particularly among vulnerable populations. AI start-up ClosedLoop developed and open-sourced the C-19 Index, a predictive model that identifies people most at-risk of severe complications from COVID-19 via a short questionnaire. Risk factors were determined using a NLP and then structured in an algorithm that places users on a risk spectrum. The C-19 Index is used by healthcare systems, care management organizations, and insurance companies to recognize high-risk individuals, contact them to share safety resources, and provide them with personal protective equipment.
Although AI can’t spit out a cure for COVID-19, it is helping researchers develop resourceful and cost-effective strategies for making counter-COVID-19 medicine. Synthia, an AI-backed drug synthesis program, is being applied by University of Michigan chemist Timothy Cernak and colleagues to create new recipes for COVID-19 drugs and prevent supply shortages. Just this month, Cernak’s lab identified novel solutions for making 11 out of 12 compounds now being tested as COVID-19 therapies, in one case with cheaper starting materials than those currently in use. As new drugs are proven effective against COVID-19, ensuring that they can be manufactured in abundance will be high-priority.
I will gladly wait for the day that the skies part and the AI-gods descend from the heavens with cures for our mortal problems. But until then, the role of AI in fighting COVID-19 is at least a big step toward accelerating the AI health care revolution.
According to a 2019 study, the global AI health care market is expected to grow at a compound annual growth rate of 41.7 percent, from $1.3 billion in 2018 to $13 billion in 2025. Hospital workflow, wearables, medical imaging and diagnosis, therapy planning, virtual assistants, and drug discovery all promise to be transformed with the introduction of AI. COVID-19 will only expedite those trends.
It’s easy to turn to AI with our questions, but important to remember that it may not always have the answers. Perhaps by that $13 billion valuation in 2025, we will be able to kick back and watch AI tackle the next pandemic for us.