Why AI fell short in slowing the spread of COVID-19

Though hopes were high, experts say a lack of high-quality data meant models were both underwhelming and "anti-constructive."

By Kat Jercich

July 23, 2020

04:35 PM

This spring, much of the healthcare industry hoped that artificial intelligence could be a key tool in stemming the spread of the COVID-19 pandemic across the world.

But the results weren't just underwhelming. In some cases, they were "anti-constructive," said Dr. Isaac Kohane, chair of the Department of Biomedical Informatics at Harvard Medical School, during a FutureMed presentation on Thursday.

"We in healthcare were shooting for the moon, but we hadn't gotten out of our own backyard," said Kohane.

In the United States there were several attempts to use aggregate data from electronic health records. Kohane used Epic as an example, pointing to its system to predict severity of disease based on admissions.

"It didn't perform very well at all," said Kohane.

According to Kohane, a lack of high-quality data contributed to the shortfall.

"Most of the data that was being shared for the first three months was literally just case counts and death counts," he said. "To the extent that there was sharing of clinical courses, it was from single institutions," rather than interstate efforts.

"We did not have a real collective intelligence," he said.

But hope isn't completely lost for AI's role in addressing the pandemic. Kohane noted that companies are using it to develop vaccines – specifically, using large databases of protein interactions and docking simulations to figure out the best protein domain to block.

In December or sooner, he said, we'll see "the results of Phase 2 trials from purely machine-learned trials."

U.S. Food and Drug Administration Principal Deputy Commissioner Amy Abernethy said during the presentation that AI might be used to help sort through the available drugs and to help get data sets cleaned up "to better understand how drugs are performing."

Meanwhile, Eran Segal, a professor in the computer science department at the Weizmann Institute of Science, pointed to the use of AI in conjunction with surveys to help predict, based on reported symptoms, which individuals should be tested.

Ultimately, said Dr. Karen DeSalvo, chief health officer at Google Health and former National Coordinator for Health IT, those building AI tools to confront the pandemic must not replicate existing biases in medicine, a possibility that is a continued concern among many developers.

"There's a really important challenge to look at fairness: to make sure that whatever we are building is not going to exacerbate inequities in health outcomes," said DeSalvo.