A recent study reveals that the early death toll of the Covid-19 pandemic in the United States was significantly underreported compared to official records, highlighting stark differences in the number of uncounted deaths.

Official records show that approximately 840,000 deaths were attributed to Covid-19 in 2020 and 2021. However, researchers employing artificial intelligence techniques estimate there could be as many as 155,000 additional deaths that went unrecorded, primarily occurring outside hospital settings. This suggests that around 16% of Covid-19 fatalities may not have been officially documented during those years.

The findings, published on Wednesday in the journal Science Advances, align closely with previous estimates regarding pandemic-related deaths from that period. However, this new research delves deeper into identifying which specific deaths were more likely to be absent from the official counts.

The study indicates that the unreported deaths were predominantly among Hispanic and other communities of color, particularly in the early months of the pandemic. These individuals were more frequently located in southern and southwestern states such as Alabama, Oklahoma, and South Carolina.

Even six years after the pandemic first struck the United States, significant challenges persist for these communities, according to Steven Woolf, a researcher from Virginia Commonwealth University who was not involved in the study.

“Individuals on the fringes of society continue to face higher mortality rates because they lack access to adequate healthcare,” Woolf stated in an email.

While hospital patients were routinely tested for Covid-19, many who grew sick and died outside of hospitals were not tested – often because at-home testing was not readily available early in the pandemic, said one of the study’s authors, the University of Minnesota’s Elizabeth Wrigley-Field.

In some parts of the country, death investigations are handled by elected coroners who do not necessarily have the specialized training that medical examiners do. Some research has suggested partisan opinions could affect whether a sick person or their family members sought Covid-19 testing, and whether coroners pursued postmortem coronavirus testing. Indeed, some coroners said families had pressed them not to list Covid-19 as a cause of death.

“Our antiquated death investigation system is one key reason why we fell short of accurate counts, particularly outside of big metropolitan areas,” said Andrew Stokes of Boston University, the senior author on the paper.

The Centers for Disease Control and Prevention data count more than 1.2m Covid-19 deaths since the pandemic erupted in early 2020. More than two-thirds of those reported deaths occurred in 2020 and 2021.

The count has long been debated, as false claims on social media said the number of Covid-19 deaths was inflated. Adding to the rancor was Donald Trump, who in August 2020 retweeted a post claiming only 6% of reported deaths were actually from Covid-19 – a post Twitter later removed.

To be sure, there were other kinds of pandemic deaths. For example, uninfected people died from other medical conditions because they could not get care at hospitals overloaded with Covid-19 patients. People with drug addictions died of overdoses as a result of social isolation and losing access to treatment. Other studies that have estimated the actual number of pandemic deaths have taken those deaths into account.

But Stokes and his collaborators wanted to focus on the deaths of people infected by the coronavirus. They used machine learning to sift through the death certificates of infected patients who died in hospitals and then used patterns observed in those records to evaluate death certificates of people who died outside hospitals and whose deaths were attributed to things like pneumonia or diabetes.

Scientists’ understanding of the strengths and weaknesses of machine learning-reliant research is still evolving, but Woolf called this team’s use of it “intriguing”.