clinicians.build · interactive · august 17, 2026

Forty-Five Percent, By Accident

Sentara cut hospital‑onset MRSA bacteremia about 45%. Ten thousand simulated hospitals where nothing changed at all — and the size of drop each one posted anyway.

Primary source: Andrea Fox, “Sentara’s Epic EHR model targets MRSA infections early,” Healthcare IT News, 14 Aug 2026
Grounding data: MIMI Labs · CMS Care Compare, Healthcare Associated Infections – Hospital (77hc‑ibv8), 1,728 hospitals with a published MRSA bacteremia SIR
Read the Sentara story → Get the CMS HAI file → Or read today’s newsletter →

Sentara’s result is real, and the interesting part of it is not the model. They embedded an MRSA risk score in Epic and then rewrote what a nurse does when it fires — decolonization and line care moved earlier because somebody changed the flowsheet. That is the expensive half, and it is the half nobody publishes.

It is also a large multi-hospital system, which turns out to matter more than anything else on this page. Here is the same 45% drop, generated by nothing.

Ten thousand hospitals. Two consecutive years. No intervention.
2.5
45%
Post a drop this big anyway
—
Post a rise this big anyway
—
US hospitals at or below this volume
—
Median absolute swing
—
drop at least as big as the headline rise at least as big everything else — the same hospital, unchanged the vertical stripes are real — at low volume only a handful of ratios are arithmetically possible
 

Every dot is a hospital that did nothing. The red ones would have written a press release.

Where your hospital actually sits

The slider is not arbitrary. This is the real distribution of predicted MRSA bacteremias across the 1,728 US hospitals CMS gave a public score to — the denominator that decides whether your quality improvement work is measurable at all. The median is 3.6 predicted cases for the entire year.

The 80/20

Find your denominator before you buy the model. Pull last year’s NHSN MRSA bacteremia report for your facility and read one number: predicted cases. If it is under three, no infection-prevention intervention you deploy this year will produce a result you can defend — not because it will not work, but because a single case in either direction will swamp it. That is an argument for pooling across your system before you measure, not for skipping the work.

And if a vendor shows you a percentage drop with no denominator attached, ask for the raw counts. “45%” from 20 to 11 is a program. “45%” from 2 to 1 is a Tuesday.

What this simulation is not

This is a null model, not a verdict on Sentara. Every dot assumes the true underlying rate did not change at all and that year-to-year counts are Poisson around it. That is deliberately the most favourable possible case for the skeptic. Real hospital counts are usually over-dispersed — outbreaks cluster, units close, case mix drifts — which makes the noise wider than what you see here, not narrower. Sentara reported its result across a multi-hospital system, and you can check the scale in this same CMS file: their 11 hospitals with a published SIR sum to 48.1 predicted cases and 37 observed — a system-level ratio of 0.77. Press the Sentara, all 11 button and watch the red almost vanish. At that denominator a 45% drop is not something chance hands you. The point is not that their number is wrong. The point is that your version of their number probably cannot be evaluated the same way.

The simulation also cannot tell you anything about whether the intervention was worth doing. A hospital with two predicted cases a year still has patients who get MRSA bacteremia, and decolonization and line care still work on them. It only tells you that the measurement will not know.

The companion explorer plots all 1,728 hospitals against their own denominators with exact Poisson control limits — and shows that only 9.5% of American hospitals are statistically distinguishable from the benchmark at all.