Medical Error Is Probably Not the Third Leading Cause of Death. In a Thousand Reviewed Hospital Deaths, 5.2 Per Cent Were Preventable.

In 2016 the BMJ published an analysis putting deaths from medical error in the United States at 251,454 a year, which would make it the third leading cause of death after heart disease and cancer. The figure went everywhere. It is still quoted in journal introductions, medical school curricula and news coverage almost a decade later.

Two of the field’s most careful methodologists replied that the estimate is, in their words, very likely to be wrong. And in the meantime somebody had done the obvious thing: taken a thousand hospital deaths, chosen at random, and had physicians read every record to decide whether better care would have prevented each one. The answer was 5.2 per cent.

This is not an article arguing that hospitals are safe. The best current measurement finds an adverse event in nearly a quarter of admissions. It is an article about the difference between a harm rate, which is real and large, and a death toll, which is an extrapolation — and about why the second keeps coming out so much bigger than anyone can find by counting.

Where the number came from

The chain starts with the Harvard Medical Practice Study, which reviewed 30,121 records from 51 New York hospitals in 1984. It found adverse events in 3.7 per cent of hospitalisations, judged 27.6 per cent of them due to negligence, and found that 13.6 per cent of adverse events led to death. Those percentages, applied to national admissions, produced the Institute of Medicine’s 1999 estimate of 44,000 to 98,000 deaths a year.

An audit published the following year found the foundations thinner than the citation implied. The two studies did support the claim that adverse events occur in 2.9 to 3.7 per cent of admissions. Support for the assertion that about half were preventable was, in the reviewers’ assessment, less clear: the original studies did not define preventable adverse events, and the reliability of subjective judgements about preventability was never formally assessed. The methods behind the 98,000 upper bound were described as highly subjective, with unknown reliability.

The estimates then climbed. A 2013 review using the Global Trigger Tool put a lower limit at 210,000 and a true figure above 400,000. Makary and Daniel’s 251,454 followed in 2016. Each step rested on applying rates from a few chart reviews to the whole national admission count.

The estimates went up. The counting did not.

Year	Claim	How it was produced
1991	Adverse events in 3.7% of admissions; 13.6% of them led to death	Review of 30,121 records from 51 New York hospitals, 1984 data
1999	44,000–98,000 deaths a year	Those rates applied to national admissions
2013	210,000 lower limit, “more than 400,000”	Weighted average of four Global Trigger Tool studies
2016	251,454 deaths a year, third leading cause	Rates from a small set of studies applied to 35 million admissions
2012	5.2% of hospital deaths preventable	1,000 randomly selected deaths, each record read by a physician

The last row is out of chronological order deliberately. It is the only one in the table that counted deaths rather than inferring them.

Brennan, T. et al., NEJM, 1991; Institute of Medicine, To Err Is Human, 1999; James, J.T., Journal of Patient Safety, 2013; Makary, M. and Daniel, M., BMJ, 2016; Hogan, H. et al., BMJ Quality & Safety, 2012.

What happens when you review the deaths themselves

Helen Hogan and colleagues took ten English acute hospital trusts, stratified by region, size and teaching status, and drew 100 consecutive deaths from each for the year 2009. Paediatric, obstetric and psychiatric admissions were excluded, as were patients admitted explicitly for palliative care. Seventeen experienced physician reviewers read the whole record for each case — nursing notes, drug charts, test results — and judged first whether a problem in care had contributed to the death, and then whether the death itself was preventable.

They used “problem in care” rather than “adverse event”, because it captures failures to act as well as things actively done wrong. Preventability was scored on a six-point scale, and a death counted as preventable only if the reviewer thought there was a better than even chance that better care would have prevented it.

Fifty-two of the thousand deaths met that bar: 5.2 per cent, with a confidence interval of 3.8 to 6.6. A problem in care contributing to death was found in 131 cases, so about two in five of those did not translate into a preventable death. Scaled to English hospital deaths, that gives roughly 11,859 preventable deaths a year, against the 60,000 to 255,000 that had been the working figure.

The answer moves with where you put the line

Share of 1,000 English hospital deaths judged preventable, by the threshold used on a six-point scale.

Including “possibly preventable but not very likely”8.5%

Better than an even chance — the headline figure5.2%

If two reviewers had been required to agree2.8%

Strong evidence or definite only2.3%

The authors argue their known biases — single-reviewer judgement and hindsight bias — would push the estimate up rather than down, so 5.2% is more likely too high than too low.

Hogan, H., Healey, F., Neale, G., Thomson, R., Vincent, C. and Black, N., BMJ Quality & Safety 21:737–745, 2012.

The part nobody quotes

The comparison everybody reaches for is the jumbo jet: a plane full of people falling out of the sky every day. Hogan’s study contains the finding that breaks that image, and it is not in the abstract of anything you will have read.

The median estimated remaining life expectancy of the patients whose deaths were judged preventable was six months, with an interquartile range of four months to two years. Sixty per cent of them had less than a year. These were, in the main, people who were already dying.

An American study made the same point more sharply. Among 111 in-hospital deaths at seven Veterans Affairs centres, reviewers rated 6.0 per cent as probably or definitely preventable — close to Hogan’s figure. Then the reviewers were asked a second question: in the absence of any problem in care, would this patient have lived at least three more months in good cognitive health? They said yes for 0.5 per cent of the deaths, with a confidence interval of 0.3 to 0.7 per cent — roughly one patient per 10,000 admissions.

Two questions about the same deaths

6.0%Of 111 US hospital deaths, the share physicians rated probably or definitely preventable. In line with the English figure of 5.2%.

0.5%The share who, absent any problem in care, would have lived at least three more months in good cognitive health. About one patient per 10,000 admissions.

6 monthsMedian remaining life expectancy of the English patients whose deaths were judged preventable. 60% had less than a year.

This does not make those deaths acceptable, and the commentary that assembled these figures says so explicitly. It makes the aviation comparison wrong, and it makes hospital mortality a poor instrument for finding unsafe care.

Hayward, R. and Hofer, T., JAMA, 2001, as reported in Shojania, K.G., “Deaths due to medical error: jumbo jets or just small propeller planes?”, BMJ Quality & Safety, 2012; Hogan et al., 2012.

Now the number that is real

None of the above says hospital care is safe. The best modern measurement of harm, as opposed to death, is a 2023 study in the New England Journal of Medicine that took a random sample of 2,809 admissions to eleven Massachusetts hospitals in 2018. Trained nurses screened the records using trigger methods and physicians adjudicated what they found.

At least one adverse event was identified in 23.6 per cent of admissions. Among the 978 adverse events, 222 — 22.7 per cent — were judged preventable, and 316 — 32.3 per cent — were serious. That is roughly one admission in four involving some harm from care, more than three decades after the Harvard study found 3.7 per cent, which mostly reflects how much harder people are now looking.

Harm is common. Death from harm is rare.

	Rate	Source
Admissions with at least one adverse event	23.6%	2,809 admissions, 11 Massachusetts hospitals, 2018
Adverse events judged preventable	22.7% of 978 events	Same study
Adverse events judged serious	32.3% of 978 events	Same study
Hospital deaths judged preventable	5.2% (95% CI 3.8–6.6)	1,000 deaths, 10 English trusts, 2009
Deaths where 3 more good months were lost	0.5% (95% CI 0.3–0.7)	111 deaths, 7 US Veterans Affairs centres

Bates, D.W. et al., “The Safety of Inpatient Health Care”, New England Journal of Medicine, 2023; Hogan et al., 2012; Hayward and Hofer, 2001.

Why “preventable” is the weak joint

Every one of these estimates rests on a judgement that a trained person makes by reading a record, and one of those judgements is much shakier than the other. A Dutch team had a committee re-review the same records twice and measured the agreement. Detecting whether an adverse event occurred was reproducible. Deciding whether it was preventable was not — only fair reliability, which the authors attributed to the absence of any agreed definition of what preventability means.

That is not a small technicality. It is the hinge on which every headline number turns. Hogan’s estimate moves from 2.3 per cent to 8.5 per cent depending purely on where the threshold sits on the same six-point scale, applied to the same thousand records by the same reviewers. And a French comparison of three different detection methods on one patient sample found that they identify different numbers of cases and different preventable proportions, with no gold standard to arbitrate between them.

Two judgements, two different qualities of evidence

ReproducibleDid an adverse event occur? Re-reviewing the same records gives good agreement. This part of the measurement holds.

Not reproducibleWas it preventable? Only fair agreement, which the researchers attribute to there being no agreed definition of preventability at all.

2.3% – 8.5%The range the English preventable-death figure covers, from the same thousand records and the same reviewers, purely by moving the threshold.

Klein, D.O. et al., PLOS ONE, 2018; Hogan et al., 2012.

The argument has moved to diagnosis

The live version of this dispute is now about diagnostic error, and it has exactly the same shape. A 2023 analysis combining disease-specific error rates with national incidence data estimated that nearly 800,000 Americans a year die or are permanently disabled by diagnostic error. That figure is now widely quoted, including in Science.

Against it: a systematic review that pooled 22 studies covering 80,026 hospitalised adults, in which physicians had read case series of admissions and identified harmful diagnostic errors directly. The pooled rate was 0.7 per cent, with a confidence interval of 0.5 to 1.1 per cent.

The two are not measuring quite the same thing: the first covers all settings and includes permanent disability, the second covers hospitalised adults and counts what reviewers could find in the notes. But the pattern repeats: modelled estimates come out far above counted ones, and the modelled ones are what travel.

Questions people ask

So is medical error the third leading cause of death or not?

Almost certainly not, on the evidence available. The claim rests on extrapolation from small chart-review studies; the studies that reviewed deaths directly find around 5 per cent of hospital deaths preventable, and the patients concerned were mostly within months of death anyway. Two leading patient-safety methodologists concluded in print that the 251,454 estimate is very likely to be wrong.

Does that mean hospital safety is not a problem?

No, and the same researchers making the argument above say so first. Roughly a quarter of admissions involve an adverse event; a third of those events are serious. Most patient safety problems do not kill anyone — they cause pain, anxiety, temporary or permanent disability, longer stays and more treatment. Using deaths as the measure understates the problem and points at the wrong things.

Should this change whether I go to hospital?

No. Nothing in this literature compares being admitted with the counterfactual of staying at home with the condition that put you there, and the studies exclude people admitted for palliative care precisely because the comparison would be meaningless. This is research about how to make hospital care safer, not about whether to accept it.

What actually goes wrong most often?

In the English preventable deaths, the three commonest problems were clinical monitoring — failing to act on a test result or a change in condition — at 31.3 per cent, diagnosis at 29.7 per cent, and drug or fluid management at 21.1 per cent. In 73.1 per cent of preventable deaths, more than one problem was identified. Failures compound; they rarely arrive alone.

Is there anything a patient or family can usefully do?

The evidence here is much thinner than the advice industry around it suggests, and we are not going to invent a checklist. What the data does say is that the common failure is nobody acting on information that was already in the record, and that most preventable deaths involved more than one problem. Asking what a result means and what happens next is not a proven intervention, but it is at least aimed at the right place.

The short version

• The claim that medical error is the third leading cause of death in the US rests on 251,454 deaths a year, extrapolated from a small number of chart reviews. Two senior patient-safety researchers published that it is very likely to be wrong.
• When 1,000 randomly selected English hospital deaths were reviewed individually, 5.2% (95% CI 3.8–6.6) were judged preventable. A US study of 111 deaths found 6.0%.
• The median remaining life expectancy of those patients was six months; 60% had less than a year.
• Asked whether the patient would have lived even three more months in good cognitive health, US reviewers said yes for 0.5% of deaths — about one per 10,000 admissions.
• Hogan’s figure moves between 2.3% and 8.5% depending only on where the threshold sits on the same six-point scale.
• Detecting an adverse event by record review is reproducible. Judging it preventable is not, and there is no agreed definition of preventability.
• The harm rate is real and large: at least one adverse event in 23.6% of a random sample of 2,809 US admissions in 2018; 22.7% of the events preventable, 32.3% serious.
• The same dispute is now running over diagnostic error: a modelled estimate of nearly 800,000 deaths or permanent disabilities a year, against a pooled chart-review rate of 0.7% of hospitalised adults.

This article summarises published research on patient safety measurement. It is not medical advice, it is not guidance about any individual’s care, and nothing in it should be used as a reason to delay, decline or change treatment. If you have a concern about care you or a family member received, the place to raise it is with the clinical team or the hospital’s own complaints or patient safety process.

Further reading: Hogan and colleagues’ 2012 study and Shojania’s accompanying commentary are both open access and are the two documents that changed how this question should be asked. The 2023 NEJM paper is the current benchmark for how much harm there is.

Sources

• Hogan, H., Healey, F., Neale, G., Thomson, R., Vincent, C. and Black, N., “Preventable deaths due to problems in care in English acute hospitals: a retrospective case record review study”, BMJ Quality & Safety 21:737–745, 2012. (1,000 deaths, 100 each from 10 English acute trusts, year 2009; paediatric, obstetric, psychiatric and palliative admissions excluded. 52 deaths, 5.2%, 95% CI 3.8–6.6, judged preventable at better than an even chance; 39.7% of the 131 cases with a problem in care contributing to death. Stricter threshold 2.3%, looser 8.5%, requiring two reviewers to agree 2.8%. Commonest problems: clinical monitoring 31.3%, diagnosis 29.7%, drugs or fluids 21.1%; more than one problem in 73.1% of preventable deaths. Median life expectancy of those patients six months, IQR four months to two years, 60% under one year. National estimate 11,859 against the Chief Medical Officer’s 60,000–255,000. The authors state their biases would lead to overestimating, not underestimating.)
• Shojania, K.G., “Deaths due to medical error: jumbo jets or just small propeller planes?”, BMJ Quality & Safety, 2012. (Commentary reporting that among 111 in-hospital deaths at seven US Veterans Affairs medical centres, reviewers rated 6.0% as probably or definitely preventable, and that when asked whether the patient would have lived at least three months in good cognitive health absent any problem in care, clinicians estimated 0.5%, 95% CI 0.3–0.7%, roughly one patient per 10,000 admissions. Argues that the small proportion of preventable deaths makes hospital mortality a poor performance measure, while stating that this does not undermine the importance of patient safety.)
• Shojania, K.G. and Dixon-Woods, M., “Estimating deaths due to medical error: the ongoing controversy and why it matters”, BMJ Quality & Safety 26:423, 2017. (Response to Makary and Daniel’s attribution of 251,454 deaths per year in US hospitals to medical error. The authors state that the new estimate is very likely to be wrong and address the role of mortality as a patient safety indicator.)
• Bates, D.W., Levine, D.M., Salmasian, H., Syrowatka, A., Shahian, D.M. and Lipsitz, S.R., “The Safety of Inpatient Health Care”, New England Journal of Medicine, 12 January 2023. (Retrospective cohort study, random sample of 2,809 admissions to 11 Massachusetts hospitals during 2018. At least one adverse event identified in 23.6% of admissions. Of 978 adverse events, 222, 22.7%, judged preventable and 316, 32.3%, of serious severity. Trigger method screening by trained nurses with physician adjudication.)
• Brennan, T.A., Leape, L.L., Laird, N.M., Hebert, L., Localio, A.R. and Lawthers, A.G., “Incidence of adverse events and negligence in hospitalized patients: results of the Harvard Medical Practice Study I”, New England Journal of Medicine, 1991. (30,121 records from 51 randomly selected New York hospitals, 1984. Adverse events in 3.7% of hospitalisations, 95% CI 3.2–4.2; 27.6% due to negligence, 95% CI 22.5–32.6; 70.5% caused disability lasting under six months, 2.6% permanent disability, 13.6% led to death.)
• Sox, H.C. and Woloshin, S., “How many deaths are due to medical error? Getting the number right”, Effective Clinical Practice, 2000. (Audit of the studies the Institute of Medicine cited. They substantiate adverse events in 2.9–3.7% of admissions; support for the claim that about half were preventable is less clear, the original studies did not define preventable adverse events, and the reliability of subjective preventability judgements was never formally assessed. The methods behind the 98,000 upper bound are described as highly subjective with unknown reliability.)
• Klein, D.O., Rennenberg, R., Koopmans, R.P. and Prins, M.H., “Adverse event detection by medical record review is reproducible, but the assessment of their preventability is not”, PLOS ONE, 2018. (Committee re-review of the same records. Good reliability for the presence of an adverse event, only fair reliability for preventability; the authors attribute this to the absence of a definition of preventability and call for international consensus on what constitutes it.)
• Michel, P., Quenon, J.L., de Sarasqueta, A.M. and Scemama, O., “Comparison of three methods for estimating rates of adverse events and rates of preventable adverse events in acute care hospitals”, BMJ, 2004. (778 patients across 37 wards in seven French hospitals. Three detection methods applied to the same sample identify different numbers of cases and different preventable proportions; the authors note there is no gold standard.)
• Newman-Toker, D.E., Nassery, N., Schaffer, A.C., Yu-Moe, C.W., Clemens, G. and Wang, Z., “Burden of serious harms from diagnostic error in the USA”, BMJ Quality & Safety, 2023. (Combines disease-specific diagnostic error and harm rates with national disease incidence from 21.5 million sampled US hospital discharges and cancer registries, with Monte Carlo uncertainty estimates. The widely quoted result is that nearly 800,000 Americans die or are permanently disabled by diagnostic error each year.)
• Gunderson, C.G., Bilan, V.P., Holleck, J.L., Nickerson, P.A., Cherry, B.M. and Chui, P.W., “Prevalence of harmful diagnostic errors in hospitalised adults: a systematic review and meta-analysis”, BMJ Quality & Safety, 2020. (22 studies, 80,026 patients, 760 harmful diagnostic errors identified by physician review of consecutive or randomly selected admissions. Pooled harmful diagnostic error rate 0.7%, 95% CI 0.5–1.1%.)
• James, J.T., “A New, Evidence-based Estimate of Patient Harms Associated with Hospital Care”, Journal of Patient Safety, 2013. (Weighted average of four Global Trigger Tool studies giving a lower limit of 210,000 deaths a year associated with preventable harm, and the author’s own estimate of more than 400,000 given the tool’s limitations.)