The disease became invisible not because it wasn't there, but because the system was never built to see it.
Across 58 American hospitals, an artificial intelligence tool has done what official public health infrastructure was never designed to do: follow a patient through time, connecting a COVID infection to the chronic conditions that quietly accumulated in its wake. The result — one in six infected people developing long COVID, more than double the government's own count — is less a discovery than an exposure, revealing that over ten million Americans have been living with serious chronic illness inside a surveillance system that was structured, whether by neglect or design, not to see them. The question this raises is older than any algorithm: for whom does public health exist, and who decides which suffering gets counted?
- An AI analysis of nearly 458,000 patient records found long COVID affecting 16.3% of COVID-19 patients — a rate more than twice what federal billing-code surveillance had been capturing.
- Over ten million Americans with long COVID have effectively disappeared from the public health record, their symptoms scattered across cardiology, neurology, and endocrinology clinics with no thread connecting them back to COVID.
- The burden falls hardest on those least able to carry it: essential workers, low-income communities, and economically disadvantaged populations who could not shelter from exposure and cannot absorb the cost of chronic illness.
- Harvard economists estimate long COVID has cost the U.S. economy $3.7 trillion — rivaling the Great Recession — while millions of workers have left the labor force entirely because of the disease.
- Rather than expanding surveillance, recent federal decisions have cancelled RECOVER program grants and shuttered the Office of Long COVID Research, choices the researchers frame not as budget constraints but as political decisions about whose suffering gets made visible.
A research team at Mass General Brigham trained an AI algorithm on the complete medical histories of nearly 458,000 patients across 58 hospitals in four American regions, looking for something the standard public health system had been missing. What they found was a prevalence of long COVID at 16.3 percent — roughly one in six people who contracted COVID-19 — compared to the federal government's own figure of under seven percent, derived from a single billing code that was never built to trace chronic illness across time.
The gap represents more than a statistical discrepancy. It means over ten million Americans with long COVID have vanished from the official record, their suffering distributed across specialist clinics — new cardiac arrhythmias here, metabolic disorders there, unexplained cognitive fog somewhere else — with no mechanism connecting those symptoms back to a COVID infection. The algorithm, called P2RC, did what billing codes could not: it read each patient's longitudinal history, identified new chronic conditions that could not be explained by pre-existing disease, and followed the trajectory of illness across months and years. Prevalence ranged from 13.6 percent in western Pennsylvania to 22.7 percent in southern California, and showed no sign of leveling off through mid-2024.
The conditions were not transient. Roughly 14.5 percent of infected patients developed new chronic illnesses requiring ongoing care — heart disease, diabetes, cognitive decline, fatigue syndromes — findings consistent with research from the VA, the NIH's RECOVER initiative, and international cohort studies. The burden was not evenly shared. Studies of economically disadvantaged populations and essential workers — healthcare staff, educators, those who could not work from home — showed substantially higher rates of persistent illness among those least equipped to absorb its costs.
Harvard economist David Cutler placed the total economic toll at $3.7 trillion: $2.2 trillion in lost quality of life, nearly $1 trillion in reduced earnings, and $528 billion in direct medical spending. Millions of full-time-equivalent workers have exited the labor force entirely. The people the economy most depended on were the same people the disease hit hardest.
What the Brigham study ultimately demonstrated was that the undercounting of long COVID was never a technical failure. The clinical data had always existed. The P2RC tool simply did what the official system was structured to prevent: it looked. The recent cancellation of RECOVER grants and the closure of the Federal Office of Long COVID Research were not administrative footnotes — they were decisions about which suffering would remain countable and which would not. The capacity to see the full scale of this crisis had always been there. The will, the researchers made clear, was the variable that had always been political.
A research team at Mass General Brigham applied artificial intelligence to the medical records of nearly 458,000 patients across 58 hospitals in four American regions and found something the official health surveillance system had been missing: roughly one in six people who caught COVID-19 went on to develop long COVID. That's 16.3 percent. The federal government's own tracking systems, which rely on a single billing code to flag the condition, were catching less than seven percent of cases.
The gap matters because it means more than ten million Americans with long COVID have essentially vanished from the public health record. They show up in cardiology clinics with new heart rhythm problems. They see endocrinologists for metabolic disorders that weren't there before. Neurologists hear complaints of cognitive fog they cannot explain. But because these symptoms arrive without a label connecting them back to a COVID infection, the disease itself remains invisible—not because it isn't there, but because the system designed to count it was never built to see it.
The researchers, publishing their work in JAMA Network Open, used an algorithm called P2RC that did something the standard billing codes could not: it read each patient's complete medical history before and after infection, looking for new chronic conditions that could not be explained by pre-existing disease. What emerged was a portrait of a much larger crisis than official numbers suggested. Prevalence varied by region—13.6 percent in western Pennsylvania, 22.7 percent in southern California—but the pattern was consistent. And there was no sign of plateau. The cumulative prevalence kept climbing through mid-2024, with statistical models pointing toward sustained growth over the next decade if current trends hold.
The conditions themselves were serious. About 14.5 percent of infected patients developed new chronic illnesses requiring ongoing clinical care: heart disease, diabetes, cognitive decline, fatigue syndromes, metabolic and endocrine disorders. This was not a post-viral syndrome that would fade on its own. This was chronic disease, the kind that reshapes a life. The findings align with other major research—work from the Department of Veterans Affairs analyzing millions of patient records, the NIH's RECOVER initiative, international cohort studies—all pointing to long COVID prevalence between ten and twenty-five percent among infected adults, with substantial numbers still severely ill one to three years after acute infection.
But the burden was not evenly distributed. A 2024 cohort study found that economic hardship and structural inequality significantly increased the risk of developing persistent symptoms. A British study of over 200,000 working-age adults showed people in the most economically disadvantaged areas faced substantially higher risk than those in the least disadvantaged. Essential workers—healthcare staff, educators, those who could not work from home and had to stay present to keep the economy running—faced the highest rates. These were the people least able to absorb the cost of chronic illness.
The economic toll was staggering. Harvard economist David Cutler estimated long COVID's total cost to the U.S. economy at 3.7 trillion dollars, a figure rivaling the aggregate damage of the Great Recession. Of that, 2.2 trillion reflected lost quality of life, nearly one trillion came from reduced earnings, and 528 billion represented direct medical spending. Brookings Institution analysis showed millions of full-time-equivalent workers had left the labor force entirely because of long COVID. The people most burdened by the disease were the same people the economy depended on.
The invisibility of long COVID was not a measurement error. It was institutional. Defective billing codes, fragmented care, the absence of systematic patient tracking, and the political decision to declare the pandemic over all played a role. The recent cancellation of RECOVER program grants and the closure of the Federal Office of Long COVID Research and Practice were not mere administrative matters. They were choices about which suffering gets counted and which gets hidden. What the P2RC tool demonstrated was that the data had always been there. The clinical records contained the truth. The question was whether anyone had the will to look.
That raised a harder question: for whom does public health exist? The answer, the Brigham researchers showed, had always been a political choice, not a technical limitation. For years, the bureaucratic foundations of disease surveillance—a single defective billing code, fragmented specialist care, no mechanism to follow a patient over time—functioned as a form of censorship that did not merely fail to count the sick but made them uncountable. P2RC broke that architecture by reading the complete longitudinal record and doing what the official system was structured to prevent: connecting a COVID infection to the chronic conditions accumulating in its wake across months and years, tracking how health deteriorated in ways no billing code was designed to capture. The capacity to do this had always existed. The will had not. And that gap was not technical. It was political.
Citações Notáveis
More than 10 million people with long COVID would remain completely invisible to the billing code on which health systems and policymakers rely to track disease burden. The figures we have discovered are almost certainly an undercount.— Hossein Estiri, Mass General Brigham
The cardiologist observing new heart rhythm problems, the endocrinologist detecting new metabolic disease, the neurologist facing inexplicable cognitive complaints—some of these manifestations are long COVID arriving without the label that would connect them to a COVID-19 infection.— Jiazi Tian, data scientist and lead author
A Conversa do Hearth Outra perspectiva sobre a história
Why does a single billing code matter so much? Couldn't doctors just diagnose long COVID based on symptoms?
A billing code is how the system counts. If a patient has heart problems after COVID, the cardiologist treats the heart. If they develop diabetes, the endocrinologist treats the diabetes. Each specialist sees their piece. But no one is connecting those pieces back to the original infection, and no one is counting that patient as a long COVID case. The code is supposed to do that work. When it fails, the disease becomes invisible to public health.
So the AI tool just looked at the same medical records that already existed?
Exactly. The data was always there. Every symptom, every new diagnosis, every lab result—it was all in the electronic medical records. The AI just read the full story instead of looking for a single label. It connected the dots that the fragmented system had left scattered.
And the numbers are significantly higher than what officials have been reporting?
More than double. The official surveillance caught less than seven percent. This study found 16.3 percent. That's the difference between a manageable problem and a massive public health crisis. Over ten million people unaccounted for.
Why would the government close the RECOVER program if the problem is this large?
That's the question the article is really asking. It wasn't a lack of data or evidence. It was a choice. The pandemic was declared over because the economy needed to reopen, not because the disease had stopped. Long COVID didn't fit that narrative, so the infrastructure to track it was dismantled.
Who bears the heaviest burden?
The people who had to keep working. Essential workers, low-income people, those in economically precarious communities. They couldn't isolate. They caught the virus at higher rates. And now they're the ones living with chronic illness while trying to stay employed, often unable to work at all. The disease is most severe in the people least able to afford it.
What does the 3.7 trillion dollar figure actually represent?
It's the total economic cost—lost wages, lost quality of life, medical expenses. To put it in perspective, that's comparable to what the entire 2008 financial crisis cost. Except this one is ongoing and concentrated among working people.