By Austin VornhagenSomewhere in a hospital tonight, a test result says something is wrong.
What happens next mostly isn’t a medical question. The scan has been read. The finding has been written down in plain English. The question that decides whether this ends well is administrative, almost boring: whose job is it now?
In 2025, federal investigators went back to a sample of hospitalized Medicare patients they knew had been harmed by their care and asked the hospitals a simple thing: did your own safety systems catch this? For about half of the harm events, the answer was no. Not caught late. Not caught at all.
Medicine’s hardest killers today aren’t a slip of the hand. They’re dangers someone noticed, wrote down, and nobody owned.
That sounds like a modern management problem. It is actually the latest chapter of a 180-year detective story, in which the culprit keeps changing address. Every time medicine caught it, the fix had the same shape. Watch for it.
The killer keeps moving outward
Each ring below is a bigger zone of care, from one pair of hands to an entire health system. Scroll, and follow the red marker as medicine discovers where the danger is hiding.
Fix: make cleaning the hands someone’s job.
The doctors were carrying it
Physicians and students went from autopsies straight to examining women in labor. In 1847 Ignaz Semmelweis made them wash in a chlorinated lime solution. In his First Clinic, the yearly death rate among mothers went from about 11 in 100 in 1846 to about 1 in 100 by 1848. The people trying to save mothers had been carrying the disease that killed them.
Then it lived in the operating field
Joseph Lister published his work on antiseptic surgery: treat the wound, the dressings, and the instruments to control contamination. Deaths that surgeons had accepted as the price of cutting turned out to be preventable by changing routine.
Then it hid in what nobody checked
Surgeon Ernest Amory Codman pushed his “End Result Idea”: follow every patient long enough to learn whether the treatment worked, and study the failures. Colleagues resisted. His idea helped shape the hospital standardization movement that grew into modern accreditation.
Then the hospital stay became the hazard
Elihu Schimmel followed more than 1,000 patients on a hospital medical service. About 20 percent had a complication caused by their care: 240 episodes, 16 of them fatal. Not every episode was an error. But somebody was finally looking for injuries caused by care instead of waiting for lawsuits.
Then one specialty engineered it down
Anesthesia can kill in minutes if oxygen or breathing fails. The Anesthesia Patient Safety Foundation formed in 1985; national monitoring standards followed in 1986. Pulse oximeters and exhaled carbon dioxide monitors watched continuously, and equipment was redesigned so certain mistakes became harder to make.
Then the whole system got a body count
The Harvard Medical Practice Study reviewed about 30,000 New York hospital records from 1984: adverse events in 3.7 percent of stays, and 13.6 percent of those led to death. In 1999 the Institute of Medicine’s To Err Is Human estimated 44,000 to 98,000 Americans died each year in hospitals from medical errors, and argued the fix was systems built for human fallibility.
Now it lives in the gaps
A result flagged in one building and read in another. A diagnosis that was a reasonable guess on Tuesday and never revisited. A follow-up that every party assumes someone else scheduled. The danger has moved into the space no single team owns.
Every win had the same shape. Someone took a danger that lived in the gaps of attention and gave it an owner.
Hands became a washing rule. Wounds became antisepsis. Outcomes became a review. Oxygen became a machine that beeps. And the best modern example is almost embarrassingly simple: in 2006, a project in Michigan intensive care units made five already-known infection-prevention steps for central lines reliably someone’s job, backed by teamwork and measurement. Bloodstream infection rates fell by up to 66 percent, and stayed down over the study period.
So we must be winning, right? How many people still die?
That is where the story gets strange. Nobody can tell you.
Four famous numbers, four different questions
You have probably heard that medical error is the third leading cause of death in America. You may also have heard that’s wildly overblown. Both camps cite real research. Here is the thing almost every headline leaves out: the famous estimates aren’t a low and a high guess at one quantity. They are answers to different questions. Tap each one.
How many Americans die in hospitals each year from medical errors?
- Where it looked
- Hospital stays only.
- How it got there
- Extrapolated nationally from samples of hospital records reviewed in earlier studies, not counted directly.
- The catch
- The famous number that started the modern patient-safety movement. It is a range because the underlying studies disagreed.
Of the people who die in U.S. hospitals, how many deaths did physician reviewers judge preventable?
- Where it looked
- Hospital deaths only. Pooled from eight studies covering 12,503 deaths.
- How it got there
- About 3.1 percent of reviewed deaths were judged preventable, then scaled to the nation.
- The catch
- A strict question: would this person likely have lived? It misses deaths after discharge and harm outside the hospital.
How many U.S. deaths are associated with medical error?
- Where it looked
- Hospital-based studies, extrapolated to the country.
- How it got there
- Combined earlier studies’ death rates and applied them to national hospital admissions.
- The catch
- The source of the “third leading cause of death” line. Critics argue it counts deaths where an error occurred without showing the error caused the death. It is not a settled statistic.
How many U.S. deaths follow a missed or wrong diagnosis, anywhere in care?
- Where it looked
- Clinics, emergency departments, and hospitals.
- How it got there
- Modeled from disease-specific error and harm rates applied to national disease counts, with a wide uncertainty range.
- The catch
- A different kind of failure (diagnosis), in every setting, estimated by modeling rather than chart review. Not every one of those deaths was necessarily preventable.
Bars to scale, per year, United States. For the 1999 range the bar shows the top. None of these is a direct count of deaths in 2026.
Notice that the smallest number and the biggest number aren’t even looking at the same failures. One asks about preventable deaths inside hospitals. The other asks about missed diagnoses everywhere. Lining them up as a range is like averaging a car’s top speed with its fuel economy.
But the deepest disagreement isn’t about scope. It’s about one brutal question, and the next chart shows how much it changes the answer.
Chart 03: the hundred-death sieve
Picture 100 people who died in a hospital
Most of them died of the illness that brought them in. That is what hospitals are for: the sickest people in the country go there. Now ask three questions that sound identical and are not.
“Did anything go wrong?”
In a 2024 study of 2,428 hospitalized adults who died or were moved to intensive care at 29 academic medical centers, reviewers found a diagnostic error in about 23 percent. That is just one kind of error, in a high-risk group. Errors are common in the records of very sick people.
“Did an error contribute to the death?”
In that same study, among the 1,863 patients who died, reviewers judged that a diagnostic error played a role in 6.6 percent. The pile shrinks fast once you ask about cause instead of presence.
“Would they have lived without it?”
A 2020 review pooled eight studies where physicians judged whether hospital deaths were preventable: about 3.1 in 100. That counterfactual is the hardest judgment in medicine, because a very sick person can die after an error and would have died anyway.
Same 100 people. Answers of 23, 7, or 3.
Multiply each answer by a whole country and you get headlines that differ by many times over, all from honest research. A bigger published number can mean a broader question, better detection, or a different setting, not more deaths.
The counting machine leaks too
Researchers can argue about counterfactuals forever. Hospitals have a more basic job: notice when their care hurts someone, so they can stop it from happening twice. That’s what incident reporting and surveillance systems are for.
In July 2025, the HHS Office of Inspector General tested those systems. It took harm events its own reviewers had found among Medicare patients discharged in October 2018, then asked 154 hospitals, covering 266 of those events, whether their systems had captured them. Watch where the 100 go.
Per 100 harm events, rounded. HHS OIG, 2025.
Why did they miss so much? Not mainly because people hid things. For 46 percent of the missed events, staff said they didn’t consider what happened to be harm. For 18 percent, recording that kind of event just wasn’t standard practice. The definition of “harm” was narrower than the harm.
And catching an event was only the start. Of the events hospitals did capture, few were investigated, and fewer led to a change in how care was delivered. A companion report looked at the captured events that hospitals were required to report to outside regulators: 15 of them. Hospitals had reported 5.
That doesn’t mean half of medical deaths go unreported. It means the systems meant to learn from harm are blind to a huge share of it, and you can’t fix a pattern you never see.
So where are we right now?
Honestly: improving where the danger has an owner, stuck where it doesn’t.
The clearest good news is infections, the oldest enemy on our map. CDC’s latest national report compares 2024 with 2023 in U.S. acute care hospitals, using risk-adjusted infection ratios:
- Central-line bloodstream infections−9%
- Catheter-associated urinary infections−10%
- Hospital-onset MRSA bloodstream infections−7%
- Hospital-onset C. difficile infections−11%
- Surgical-site infections, colon surgery−4%
- Surgical-site infections, abdominal hysterectomy+8%
Five of six falling, one rising. Infection prevention is the textbook case of the pattern: known steps, measured relentlessly, owned by named teams.
Now look at the dangers that live in the gaps. Diagnosis is the big one. The 2023 modeling study above estimated that missed and wrong diagnoses cause about 795,000 deaths or permanent disabilities a year in the U.S., concentrated in conditions like stroke, sepsis, pneumonia, blood clots, and lung cancer. A diagnosis is a moving target spread across visits, buildings, and weeks, so no single team’s checklist catches the miss.
The patient-safety organization ECRI, with the Institute for Safe Medication Practices, ranked over-reliance on AI in diagnosis as the top patient-safety concern for 2026, alongside emergency department boarding, under-recognized harm events, and rural access. That ranking is a judgment about rising risk, not evidence that AI is killing the most people. Globally, the WHO estimates unsafe care contributes to more than 3 million deaths a year.
The fixable dangers got owners. The ones left are the ones that fall between owners.
Follow one result through the cracks
The red line is responsibility. Watch it travel from station to station, and watch what happens to it in today’s system and in the one I’d build.
A fall, a scan, and a sentence nobody acts on
Here is a composite scenario patient-safety researchers would recognize. A woman goes to the emergency department after a fall. The CT for her ribs is fine, but the radiologist notices a small spot on her lung and writes: “Recommend follow-up CT in three months.” The ER team’s job is the fall. They treat it well.
Responsibility becomes a piece of paper
She goes home with discharge instructions about her ribs. The lung sentence lives in the radiology report. Ownership of the spot has quietly passed from a team of people to a document.
Three reasonable assumptions
Her primary care office gets the records, but the visit was about a fall. The ER assumes primary care will see the report. Primary care assumes the ER told her. She assumes no news is good news. Nobody has made a mistake you could point to on a chart.
The deadline passes in silence
No alarm goes off, because no system was ever told this result had a deadline. If the spot was early cancer, the cheapest, safest window to catch it is closing. The red line on the floor simply ends.
Now the sentence creates an obligation
In the system this essay proposes, a high-stakes recommendation can’t just be written down. It opens a follow-through ticket: what is unresolved, which team owns it, by when, what counts as done, and what happens if it stalls. Handing her to primary care transfers the ticket explicitly, and primary care has to accept it.
If it stalls, it escalates instead of evaporating
Say she moves and changes insurance. The ticket doesn’t care. On the deadline, an open ticket goes to a funded regional team that can call her, book the scan, and arrange a ride. It closes only when someone has assessed the result and she knows the plan, or she makes an informed choice to decline.
The world I’d build: responsibility follows the patient
Here is the design principle in one sentence: responsibility follows the patient until an identified danger has been dealt with, and comes with the staff, money, and authority to act. Better diagnostic tools help, but only inside a system that reliably connects noticing a problem to doing something about it.
Imagine leaving an appointment knowing what they think is wrong, what’s still uncertain, when you should feel better, and who will contact you if you don’t. You can decline more care and the system respects it. You can say “I’m getting worse” without fighting a phone tree. Six changes make that experience real. Some pieces exist today; the rest are my proposals, labeled as such.
Every high-stakes plan comes with a follow-through ticket
Five fields: what is unresolved, who owns it (with backup), the deadline, what counts as done, and where it goes if it fails. Sending a message doesn’t count as done. Neither does booking an appointment the patient can’t realistically reach.
Exists in pieces today as closed-loop follow-up. The extension: make it portable across hospitals, insurers, and staff turnover.
A diagnosis ships with a forecast
“We think it’s this. You should improve by about here. If you don’t, this team reconsiders.” The system watches for a mismatch between the expected course and the real one, and keeps contradictory evidence visible instead of letting a guess harden every time it’s copied into a new note.
Proposal. Needs testing so it doesn’t drown patients in extra tests and worry.
On track. The forecast quietly closes itself.
An independent second look for high-risk moments
A separate service hunts for consequential gaps: a treatment that conflicts with a verified allergy, a result still waiting for action, a patient getting worse under the current explanation, an owner who vanished during a transfer. Software screens; clinicians and pharmacists judge. It needs the original results and the patient’s own account, because five AI tools reading the same incomplete chart can share the same mistake.
Proposal. Judged by harm prevented and harm introduced, not alert counts.
Some treatments become safe by design
For tightly defined tasks, delivery systems that confirm the patient and the drug, watch the relevant vital signs, and catch a developing problem before the whole dose is in. That’s the anesthesia lesson, extended.
With restraint: studies of rigid electronic “hard stops” have documented unintended consequences, including delayed treatment. Emergency overrides are part of the design.
Failures travel as tests, not just lessons
When one hospital finds that an abnormal result goes invisible during a transfer, it publishes a privacy-protected test case: Can your system detect this result, identify the owner, and get it acted on? Other hospitals run it against their real software. Vendors run it before releasing updates.
Proposal. Rotated and independently audited, so nobody just studies for the exam.
Safety gets its own capacity
Regional teams that can do urgent reassessments, pharmacy reviews, home visits, rides, and specialist bookings, with some reserve capacity funded the way communities fund fire departments: for being ready. Rural areas need local staff and transport, not just a video link.
The expensive part, and the part software can’t replace. A system that notices danger needs somewhere to send the patient.
And the incentives have to reward finding harm
Independent reviewers would sample records for harm and publish risk-adjusted results, uncertainty included. Hospitals would earn credit for finding problems and proving their fixes worked, while concealment and reckless conduct stayed accountable. People harmed by care would get a prompt explanation, practical help, and a fair compensation process. Remember the leaking counter: a reporting system can look spotless precisely because it’s blind.
Start with one narrow, measurable failure
I wouldn’t launch all six at once. I’d pilot the follow-up of clinically important abnormal test results across one region’s hospitals and clinics, with the ticket, a staffed follow-up service, independent audits, and real capacity to book the care. Then measure serious diagnostic delays and harm, and also unnecessary procedures, patient burden, clinician workload, cost, and differences between patient groups. Expand only if the whole service makes care better. A pretty dashboard full of acknowledged alerts proves nothing.
Every leap in safety was the same move. We have one left to make.
Go back down the corridor. Semmelweis didn’t discover a drug; he gave unwashed hands an owner. Lister gave one to the wound, Codman to the outcome, the anesthesiologists to every minute of oxygen, the Michigan ICUs to five steps everyone already knew. Each time, a death that had looked like fate turned out to be a task nobody was holding.
The deaths that remain hardest to count and hardest to prevent live in the last unowned space: between buildings, between shifts, between today and three months from now. That’s why we can’t agree on the body count, and why half the harm slips past the very systems built to see it. The danger isn’t hiding. It’s homeless.
The future will still have brutal diagnoses, unavoidable complications, and patients who die despite superb care. Medicine can’t promise otherwise. What it can promise is narrower and completely achievable.
An identified danger should never quietly become nobody’s responsibility. Paint that line all the way to the end of the hall.
This essay began as a conversation with an AI about the history and current state of medical-error deaths, then was checked against sources: the HHS OIG reports on uncaptured harm and on external reporting (2025); CDC’s 2024 HAI Progress Report; Rodwin et al. (2020); Newman-Toker et al. (2023); Auerbach et al. (2024); Makary and Daniel (2016); the Institute of Medicine’s To Err Is Human; Schimmel (1964); the Harvard Medical Practice Study (1991); and Pronovost et al. (NEJM, 2006). The six-part future system is the author’s proposal, built on existing closed-loop follow-up research; it has not been shown to reduce deaths. The hero film and diagrams are illustrations.
If “give the danger an owner” stuck with you
Six Invisible Walls Stand Between You and a Plane Crash
The industry medicine keeps borrowing from: how aviation stacks owned, tested layers so one slip doesn’t become a crash.
Read the essay →Essay · InfectionThe Superbug Escape Trap
The oldest enemy on our map, still evolving. Why bacteria outran penicillin, and a trap that makes their escape backfire.
Read the essay →Essay · The people in the hallNobody Can Save a Nursing Shift
The humans holding the red line every night, and why heroics can’t substitute for a system with enough capacity.
Read the essay →Your business has open loops too.
A contact form nobody answers. A booking request that sits in an inbox. A payment that fails and nobody follows up. Small versions of the same problem: something important was noticed, and then it wasn’t anyone’s job. At Content Pilots I build fast websites where scheduling, payments, and follow-up are wired together, so every inquiry has an owner and a next step. And if you work in patient safety and think I got something wrong, email me. I’d genuinely like the correction.