The story so far
During the first half of 2020, the news was replete with stories about how AI-powered technology would help us solve the COVID-19 crisis. One of the earliest articles about the pandemic and AI was from Wired early in February 2020, just days after the first COVID-19 case in America was reported. The author optimistically describes how AI is diligently tracking cases of the virus by monitoring online activity for mentions of illness, generating visual models of the potential spread, and expanding “surveillance efforts in the US.” We now know that most of these efforts came to nothing, and even though the article maintains a positive tone, it does point out that even the most advanced AI struggles to identify something as amorphous as COVID-19 from social media posts: not all user information is available for analysis, and the virus can present so differently in various people that one person’s COVID-19 symptoms could be another’s seasonal allergies.
Tellingly, this early article points out that efforts to track spread in China have been more successful than those to track spread in North America. Why? In part, because China has far fewer privacy restrictions on user data than we have in the United States. Similarly, an April 2020 article from CNBC described AI-aided efforts to reduce COVID-19 spread in Middle Eastern countries using methods that would probably not be accepted by an American population: surveillance drones, location tracking via smartphones, enforcement robots, and even repurposing speed cameras to reduce people’s movement.
Unlike these other regions, the US has relied heavily, and almost exclusively, on voluntary data for AI-driven pandemic mitigation efforts. Because voluntary data tends to come from people seeking treatment, the data America has to work with has skewed heavily towards 1) people who have contracted rather than avoided COVID-19, 2) people who are more severely ill, and 3) people who have access to or are comfortable seeking healthcare.
Obviously, this data is not representative of the entire population. Compounding the limited data is the fact that in Americans, as well as the citizens of most other developed countries, have not taken up any method of contact tracing en masse. There are many reasons for this, from privacy concerns to underfunding to a cultural distaste for limiting personal freedoms, but the result has been that some of the most effective potential methods of AI-assisted pandemic mitigation have not been leveraged. Whereas South Korea cracked down hard on COVID-19 my utilizing personal information to contact-trace and enforce quarantines, the United States has made little attempt to trace infections or do anything beyond “recommend” quarantine for infected persons.
What we can do next
Given that America is unlikely to adopt the methods used in these other countries, how can we use AI to make a dent in COVID-19 numbers?
One way that AI has been used since quite early in the pandemic is in diagnosing COVID-19. You have probably seen a variety of questionnaires online, intended to predict your risk of severe symptoms, or to predict whether you are currently suffering from COVID-19 infection. These types of questionnaires, often administered before both routine and illness-related medical appointments, can help keep infectious people out of unequipped medical facilities and direct them to the best treatment options. They also serve as effective tools for gathering data that can be used to track the prevalence of COVID-19 in particular geographic areas.
As the pandemic has gone on, AI-powered diagnosis tools have proliferated and become more effective. We now have AI that can help detect asymptomatic cases merely by “listening” to a person’s cough, even if they are not being bothered by an involuntary cough. So far, this model has been trained on data from the sound of thousands of people’s coughs and has proven over 98% effective at detecting infection.
Another technology is one often used for detecting cancerous tumors: AI-read CT scans. Not only can this AI detect COVID-19 infection with 90+% accuracy, it can also distinguish COVID-19 lung disease from generic pneumonia, which may be caused by other conditions especially in elderly patients and those with already compromised lungs. This technique is effective even in early stages of infection, when patients may not even be experiencing breathing difficulties or dips in oxygen saturation.
Vaccine Optimization & Drug Discovery
AI is proving to be a crucial tool in managing COVID-19 vaccine candidates and in predicting which vaccines might prove most effective against proliferating mutations of the virus. Algorithms are being used to analyze virus strains directly from infected patients, enabling vaccine creators to produce vaccines with better affinity to specific strains. Although these articles are from August 2020, the technique described has led to the development of the vaccines we are now distributing.
More recent news has included announcements about how AI is being used to adapt vaccines and therapies for mutating versions of the virus, helping ensure efficacy even as we have more versions of COVID-19. AI-powered computer models are able to predict potential mutations and enable the creation of “updated” vaccines in just minutes.
The bottom line
Even though the US has chosen not to pursue the aggressive mitigation techniques favored by countries like South Korea, we are still leveraging AI in the fight against COVID-19. Once vaccine production is well established, AI-powered technologies should be able to keep them effective for years to come, even at rapid rates of mutation. We may also be able to expand access to diagnostic technologies that could enable infected people to quarantine themselves more quickly and access the right level of healthcare when they need it.