Will 10 Million People Die a Year due to Antimicrobial Resistance by 2050?

Marlieke E.A. de Kraker, Andrew J. Stewardson, Stéphan Harbarth•View original

OverviewBalancedalloy voice

You've probably heard the line by now: by 2050, antimicrobial resistance could kill ten million people every year. It's a chilling sentence. It lands in your chest. It was popularized by the twenty fourteen AMR Review led by Jim O'Neill, and it did exactly what a headline like that is supposed to do — it moved politics. But here's the uncomfortable part. When you dig into how that number was built, as de Kraker, Stewardson, and Harbarth did, the ground gets wobbly. Not because resistance isn't a serious threat — it is — but because turning a messy, patchy global reality into a single death toll requires a stack of assumptions. If those assumptions are shaky, the tower sways. The headline worked. In twenty sixteen, the United Nations held a General Assembly session on antimicrobial resistance. Countries reaffirmed surveillance and research plans under the World Health Organization's Global Action Plan from twenty fifteen. Advocacy groups like CARA pushed governments to keep their promises. That momentum matters. It opened doors for funding, stewardship, and surveillance. De Kraker and colleagues don't dismiss any of that. They just argue that political urgency shouldn't come at the cost of statistical clarity. If we're going to steer the ship for decades, we need a navigation chart we can trust — and we need it to show the fog as well as the land. So where do the numbers come from? In Europe, the backbone is the European Antimicrobial Resistance Surveillance Network, or EARS-Net. It's a hospital-based system. It records invasive infections, like bloodstream infections, that are diagnosed in hospitals. Participation varies by country, and tertiary hospitals — the big, complicated referral centers — are more likely to be in the network than small community facilities. That's already a tilt. Then comes the leap: incidence measured inside these hospitals is extrapolated to entire national populations. Catchment overlaps between hospitals are ignored. For countries outside Europe, the method goes even rougher: take average infection rates from EARS-Net, apply them per one hundred thousand people, multiply by population size. You can see the problem. A hospital-based numerator gets stretched over a population denominator that was never truly sampled. And that numerator is narrow. EARS-Net focuses on bloodstream infections because they're well defined and clinically important. But bloodstream infections are just one slice of the resistant infection pie. To fill the rest — lower respiratory tract infections, urinary tract infections, surgical site infections — the AMR Review and the European Centre for Disease Prevention and Control leaned on ratios. How many resistant pneumonias per resistant bloodstream infection? How many surgical sites? Those ratios came from very limited sources. One was a multicenter study in Brooklyn that, for the key organism in question, had just twelve extended-spectrum beta-lactamase positive Klebsiella pneumoniae bloodstream infections. Another was a single-center study in Spain that, for methicillin-resistant Staphylococcus aureus, counted four bloodstream infections. Those data points were from nineteen ninety-nine and two thousand two. Useful as clues, yes. Sturdy enough to scale nationally or globally, no. Even before you multiply anything, one technical detail can hijack the whole enterprise: culture rates. To detect resistance, you need cultures, and how often hospitals take them varies a lot. If a hospital cultures almost everyone with a fever, the proportion of isolates that look resistant can actually skew low, because you're sampling widely and picking up many susceptible infections. If a hospital cultures rarely, often in sicker patients who failed first-line therapy, the proportion of resistant isolates can look high — sometimes very high. Incidence rates — the count of resistant infections per patient-days — are less sensitive to that bias, but you need comprehensive data to calculate them. In many settings, especially in low- and middle-income countries, they're just not available. So we're stuck with proportions that over- or underestimate depending on how often blood is drawn, not how often bacteria are spreading. Now, zoom out and look at how the burden model is stitched together. Step one: estimate how many resistant infections there are. In Europe, that means multiplying a national bloodstream infection count by a resistance proportion from EARS-Net. Outside Europe, it means using World Health Organization information or EARS-Net averages as stand-ins. The pathogens of focus are familiar: Escherichia coli, Klebsiella pneumoniae, and Staphylococcus aureus, with resistance defined the way clinicians think about it — third-generation cephalosporin resistant Escherichia coli and Klebsiella pneumoniae, and methicillin-resistant Staphylococcus aureus. Step two: extend beyond the bloodstream by applying those fixed ratios to get estimates for pneumonias, urinary infections, and surgical sites. We've already seen how fragile those ratios are. But suppose we accept them for a minute. Step three is where the stakes spike: attach a mortality burden to each infection and add it up. This is where de Kraker and colleagues get especially firm. The literature often treats per-infection mortality as a simple multiplier. Take the number of resistant infections and multiply by the chance that a person with that infection dies, compared to if the bug were susceptible. But the mortality data going into that multiplier are all over the place. Some studies report crude death proportions. Others report adjusted odds ratios comparing resistant to susceptible infections. Odds ratios don't translate cleanly into attributable deaths in a population. External validity — the idea that an effect measured in one hospital, with one mix of patients and care, applies in another — is weak. The European Centre for Disease Prevention and Control compiled a range of attributable mortality estimates and it runs from zero point two percent up to thirty percent depending on pathogen and context. Even if you pick a number from the middle of that range, what does it mean when you multiply it by a national estimate that started with a biased culture proportion and a hospital denominator stretched across a country? Add a few quieter but very real statistical landmines. Time-dependent bias is one. Patients have to survive long enough in the hospital to acquire an infection and be classified as "exposed." If you ignore that so-called immortal time, you underestimate the impact of infection. Competing risks are another. Discharge alive is a competing event for in-hospital death; if you don't model that properly, mortality can look inflated. De Kraker and colleagues point out that neither of these challenges is consistently handled in the burden estimates that feed the big headline. Then come the future scenarios, the part that grabbed the world's attention. The AMR Review laid out four trajectories: resistance rising by about forty percentage points or going all the way to one hundred percent, each paired with infection rates staying the same or doubling. The models assume that the mortality risk per infection holds steady to twenty fifty. That's a big assumption. In-hospital mortality from sepsis and bloodstream infections has been trending down in many places, thanks to earlier recognition, better supportive care, and more consistent protocols. Some data even suggest faster declines in middle-income settings as systems catch up. The most quoted scenario — a sharp initial rise of resistance by around forty percentage points, stability through twenty fifty, and a doubling of infections — doesn't rest on empirical evidence. And remember the culture-rate issue. If hospitals change how often they culture over time, the apparent resistance proportion can drift for reasons that have nothing to do with biology. Imagine projecting that forward thirty-five years. One more layer: transparency and scrutiny. The AMR Review drew on outside groups like RAND and KPMG to assemble inputs, but the final burden numbers were not published in peer-reviewed journals. They were described by their own authors as "broad brush estimates," with more detailed academic work to follow. Confidence intervals? Sensitivity analyses showing how much the totals swing if you tweak culture rates, ratios, or mortality assumptions? Not reported. For numbers that are supposed to guide global priorities, that's a problem. As de Kraker, Stewardson, and Harbarth argue, estimates of this magnitude should carry explicit uncertainty at each step and be reviewed by independent experts before they're rolled out in press releases or high-level meetings. If that feels like a pile-on, take a breath. The intent here isn't to dismiss the threat of antimicrobial resistance or the need for action. It's to put a spotlight on the foundations so we can build them stronger. Right now, the foundations lean heavily on hospital-based surveillance in high-income settings, focused on invasive infections, with uneven participation, uncertain catchment denominators, and extrapolations that push beyond what the data can actually support. Low- and middle-income countries, where the burden is likely substantial, are underrepresented. Community-acquired infections are often off the radar. And the practice of turning proportions into populations, then multiplying by a mortality factor borrowed from a small, context-specific study, magnifies uncertainty at each step. So what does a sturdier path look like? First, invest in population-based surveillance that captures not just hospitals but communities, not just Europe but low- and middle-income countries where data gaps are widest. Track incidence — resistant infections per person-time — because it's less hostage to how often a clinician orders a culture. Link microbiology to outcomes so that mortality attribution isn't guesswork. Report age and gender-specific burdens; resistance doesn't hit everyone the same way. And at every stage, show the uncertainty. When a pathogen's attributable mortality could plausibly be zero point two percent or thirty percent depending on the scenario, say so, and model what that means for national totals. When ratios tying bloodstream infections to pneumonias come from a Brooklyn cohort with twelve cases or a Spanish hospital with four, explain how fragile that makes the extrapolation. Second, demand peer review before headline numbers become policy levers. That doesn't mean moving slowly. It means moving carefully. We've seen how the ten million figure catalyzed action at the United Nations in twenty sixteen and under the World Health Organization plan in twenty fifteen. Keep the urgency. But pair it with transparent data pipelines and independent checks that pressure-test each assumption. Don't just show the number. Show how it breathes when you push on culture rates, when you swap proportion for incidence, when you adjust for time-dependent bias or competing risks. Show the version for a hospital that cultures everyone and for one that cultures only the sickest. And finally, be honest about what long-range projections can and can't do. When they're built on reliable, population-based data with clear uncertainty, they're guides. When they're layered on estimates that were never meant to carry that weight, they're advocacy tools. Useful in the short run. Risky in the long run. De Kraker, Stewardson, and Harbarth aren't saying don't project. They're saying ground your forecasts in the best surveillance you can build, and carry your uncertainty out into the open. If we do that, we can keep the political momentum — from the United Nations, from the World Health Organization, from civil society — and aim it with precision. Not at a slogan, but at the specific places where better data and better care will actually save lives.

You've probably heard the line by now: by 2050, antimicrobial resistance could kill ten million people every year. It's a chilling sentence. It lands in your chest.

It was popularized by the twenty fourteen AMR Review led by Jim O'Neill, and it did exactly what a headline like that is supposed to do — it moved politics. But here's the uncomfortable part. When you dig into how that number was built, as de Kraker, Stewardson, and Harbarth did, the ground gets wobbly.

Not because resistance isn't a serious threat — it is — but because turning a messy, patchy global reality into a single death toll requires a stack of assumptions. If those assumptions are shaky, the tower sways.

The headline worked. In twenty sixteen, the United Nations held a General Assembly session on antimicrobial resistance. Countries reaffirmed surveillance and research plans under the World Health Organization's Global Action Plan from twenty fifteen.

Advocacy groups like CARA pushed governments to keep their promises. That momentum matters. It opened doors for funding, stewardship, and surveillance.

De Kraker and colleagues don't dismiss any of that. They just argue that political urgency shouldn't come at the cost of statistical clarity. If we're going to steer the ship for decades, we need a navigation chart we can trust — and we need it to show the fog as well as the land.

So where do the numbers come from? In Europe, the backbone is the European Antimicrobial Resistance Surveillance Network, or EARS-Net. It's a hospital-based system.

It records invasive infections, like bloodstream infections, that are diagnosed in hospitals. Participation varies by country, and tertiary hospitals — the big, complicated referral centers — are more likely to be in the network than small community facilities. That's already a tilt.

Then comes the leap: incidence measured inside these hospitals is extrapolated to entire national populations. Catchment overlaps between hospitals are ignored. For countries outside Europe, the method goes even rougher: take average infection rates from EARS-Net, apply them per one hundred thousand people, multiply by population size.

You can see the problem. A hospital-based numerator gets stretched over a population denominator that was never truly sampled.

And that numerator is narrow. EARS-Net focuses on bloodstream infections because they're well defined and clinically important. But bloodstream infections are just one slice of the resistant infection pie.

To fill the rest — lower respiratory tract infections, urinary tract infections, surgical site infections — the AMR Review and the European Centre for Disease Prevention and Control leaned on ratios. How many resistant pneumonias per resistant bloodstream infection? How many surgical sites?

Those ratios came from very limited sources. One was a multicenter study in Brooklyn that, for the key organism in question, had just twelve extended-spectrum beta-lactamase positive Klebsiella pneumoniae bloodstream infections. Another was a single-center study in Spain that, for methicillin-resistant Staphylococcus aureus, counted four bloodstream infections.

Those data points were from nineteen ninety-nine and two thousand two. Useful as clues, yes. Sturdy enough to scale nationally or globally, no.

Even before you multiply anything, one technical detail can hijack the whole enterprise: culture rates. To detect resistance, you need cultures, and how often hospitals take them varies a lot. If a hospital cultures almost everyone with a fever, the proportion of isolates that look resistant can actually skew low, because you're sampling widely and picking up many susceptible infections.

If a hospital cultures rarely, often in sicker patients who failed first-line therapy, the proportion of resistant isolates can look high — sometimes very high. Incidence rates — the count of resistant infections per patient-days — are less sensitive to that bias, but you need comprehensive data to calculate them. In many settings, especially in low- and middle-income countries, they're just not available.

So we're stuck with proportions that over- or underestimate depending on how often blood is drawn, not how often bacteria are spreading.

Now, zoom out and look at how the burden model is stitched together. Step one: estimate how many resistant infections there are. In Europe, that means multiplying a national bloodstream infection count by a resistance proportion from EARS-Net.

Outside Europe, it means using World Health Organization information or EARS-Net averages as stand-ins. The pathogens of focus are familiar: Escherichia coli, Klebsiella pneumoniae, and Staphylococcus aureus, with resistance defined the way clinicians think about it — third-generation cephalosporin resistant Escherichia coli and Klebsiella pneumoniae, and methicillin-resistant Staphylococcus aureus. Step two: extend beyond the bloodstream by applying those fixed ratios to get estimates for pneumonias, urinary infections, and surgical sites.

We've already seen how fragile those ratios are. But suppose we accept them for a minute. Step three is where the stakes spike: attach a mortality burden to each infection and add it up.

This is where de Kraker and colleagues get especially firm. The literature often treats per-infection mortality as a simple multiplier. Take the number of resistant infections and multiply by the chance that a person with that infection dies, compared to if the bug were susceptible.

But the mortality data going into that multiplier are all over the place. Some studies report crude death proportions. Others report adjusted odds ratios comparing resistant to susceptible infections.

Odds ratios don't translate cleanly into attributable deaths in a population. External validity — the idea that an effect measured in one hospital, with one mix of patients and care, applies in another — is weak. The European Centre for Disease Prevention and Control compiled a range of attributable mortality estimates and it runs from zero point two percent up to thirty percent depending on pathogen and context.

Even if you pick a number from the middle of that range, what does it mean when you multiply it by a national estimate that started with a biased culture proportion and a hospital denominator stretched across a country?

Add a few quieter but very real statistical landmines. Time-dependent bias is one. Patients have to survive long enough in the hospital to acquire an infection and be classified as "exposed." If you ignore that so-called immortal time, you underestimate the impact of infection.

Competing risks are another. Discharge alive is a competing event for in-hospital death; if you don't model that properly, mortality can look inflated. De Kraker and colleagues point out that neither of these challenges is consistently handled in the burden estimates that feed the big headline.

Then come the future scenarios, the part that grabbed the world's attention. The AMR Review laid out four trajectories: resistance rising by about forty percentage points or going all the way to one hundred percent, each paired with infection rates staying the same or doubling. The models assume that the mortality risk per infection holds steady to twenty fifty.

That's a big assumption. In-hospital mortality from sepsis and bloodstream infections has been trending down in many places, thanks to earlier recognition, better supportive care, and more consistent protocols. Some data even suggest faster declines in middle-income settings as systems catch up.

The most quoted scenario — a sharp initial rise of resistance by around forty percentage points, stability through twenty fifty, and a doubling of infections — doesn't rest on empirical evidence. And remember the culture-rate issue. If hospitals change how often they culture over time, the apparent resistance proportion can drift for reasons that have nothing to do with biology. Imagine projecting that forward thirty-five years.

One more layer: transparency and scrutiny. The AMR Review drew on outside groups like RAND and KPMG to assemble inputs, but the final burden numbers were not published in peer-reviewed journals. They were described by their own authors as "broad brush estimates," with more detailed academic work to follow.

Confidence intervals? Sensitivity analyses showing how much the totals swing if you tweak culture rates, ratios, or mortality assumptions? Not reported.

For numbers that are supposed to guide global priorities, that's a problem. As de Kraker, Stewardson, and Harbarth argue, estimates of this magnitude should carry explicit uncertainty at each step and be reviewed by independent experts before they're rolled out in press releases or high-level meetings.

If that feels like a pile-on, take a breath. The intent here isn't to dismiss the threat of antimicrobial resistance or the need for action. It's to put a spotlight on the foundations so we can build them stronger.

Right now, the foundations lean heavily on hospital-based surveillance in high-income settings, focused on invasive infections, with uneven participation, uncertain catchment denominators, and extrapolations that push beyond what the data can actually support. Low- and middle-income countries, where the burden is likely substantial, are underrepresented. Community-acquired infections are often off the radar.

And the practice of turning proportions into populations, then multiplying by a mortality factor borrowed from a small, context-specific study, magnifies uncertainty at each step.

So what does a sturdier path look like? First, invest in population-based surveillance that captures not just hospitals but communities, not just Europe but low- and middle-income countries where data gaps are widest. Track incidence — resistant infections per person-time — because it's less hostage to how often a clinician orders a culture.

Link microbiology to outcomes so that mortality attribution isn't guesswork. Report age and gender-specific burdens; resistance doesn't hit everyone the same way. And at every stage, show the uncertainty.

When a pathogen's attributable mortality could plausibly be zero point two percent or thirty percent depending on the scenario, say so, and model what that means for national totals. When ratios tying bloodstream infections to pneumonias come from a Brooklyn cohort with twelve cases or a Spanish hospital with four, explain how fragile that makes the extrapolation.

Second, demand peer review before headline numbers become policy levers. That doesn't mean moving slowly. It means moving carefully.

We've seen how the ten million figure catalyzed action at the United Nations in twenty sixteen and under the World Health Organization plan in twenty fifteen. Keep the urgency. But pair it with transparent data pipelines and independent checks that pressure-test each assumption.

Don't just show the number. Show how it breathes when you push on culture rates, when you swap proportion for incidence, when you adjust for time-dependent bias or competing risks. Show the version for a hospital that cultures everyone and for one that cultures only the sickest.

And finally, be honest about what long-range projections can and can't do. When they're built on reliable, population-based data with clear uncertainty, they're guides. When they're layered on estimates that were never meant to carry that weight, they're advocacy tools.

Useful in the short run. Risky in the long run. De Kraker, Stewardson, and Harbarth aren't saying don't project.

They're saying ground your forecasts in the best surveillance you can build, and carry your uncertainty out into the open. If we do that, we can keep the political momentum — from the United Nations, from the World Health Organization, from civil society — and aim it with precision. Not at a slogan, but at the specific places where better data and better care will actually save lives.

Related lectures