clinicians.build · interactive · august 17, 2026

Before You Buy the Score

Sentara embedded an MRSA risk model in Epic, redesigned what nurses do when it fires, and cut hospital‑onset MRSA bacteremia about 45%. CMS publishes that exact measure for every hospital in the country. Here is what the measure can actually see.

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), measure HAI_5, reporting period 1 Jul 2024 – 30 Jun 2025

Sentara Health put a predictive MRSA risk score into Epic and then did the part nobody writes up: they changed the flowsheet. Decolonization and line care started happening earlier because somebody redesigned what a nurse does when the number appears. Hospital-onset MRSA bacteremia fell roughly 45%, with about $1.18 million in infection-related costs avoided so far.

The score is the part you can buy. Before you do, it is worth asking what the national version of that number is capable of telling you — because CMS already computes it, for everybody, and publishes it.

Every US hospital reporting MRSA bacteremia to NHSN, Jul 2024 – Jun 2025
Hospitals reporting
4,035
Given a public SIR
1,728
Suppressed — no score at all
2,307
Pooled national SIR
0.69
CMS publishes a standardized infection ratio only when a hospital’s predicted case count reaches 1.0 for the year. The cutoff is exact and visible in the file: the smallest published denominator is 1.002 predicted cases, the largest suppressed one is 0.998. The 2,307 hospitals below that line still had 351 real MRSA bloodstream infections across 14.8 million patient days. None of it shows up in any league table.

1,728 hospitals, plotted against their own denominator

Each circle is one hospital with a published MRSA bacteremia SIR. Horizontal position is the predicted number of cases — the denominator, the thing that decides how much the ratio is allowed to wobble. Vertical position is the SIR itself. The shaded funnel is the exact Poisson range you would expect if every hospital were performing identically at the benchmark and the only thing moving the dots was chance.

1
Hospitals shown
Pooled SIR
Outside the funnel
Highest SIR shown
above the funnel — real signal, worse below the funnel — real signal, better inside — indistinguishable from the benchmark
 

Nine and a half percent. That is the share of American hospitals whose MRSA bacteremia SIR is statistically distinguishable from 1.00 at all. Everything else on the list is a hospital that reported honestly and landed inside the noise.

Drag the filter and watch the league table dissolve

Set the minimum denominator to 1 and the chart looks like wild variation in quality: hospitals at zero, hospitals at four times expected, a spread of more than four points. Push it to 10 and the field flattens into a band. Nothing improved. You just stopped looking at hospitals whose entire annual expectation is two cases.

What this dataset cannot tell you

Read this before you use any of the numbers above. The SIR is risk-adjusted for facility type, bed size, medical school affiliation and inpatient community-onset prevalence — it is not adjusted for how hard a hospital looks. Blood culture rates vary enormously between institutions, and a hospital that draws fewer cultures finds fewer bacteremias. That is a surveillance artifact that looks exactly like performance. NHSN reporting is also self-audited; CMS validates a sample, not the file. And this measure counts only laboratory-identified MRSA bloodstream events — not MRSA pneumonia, not surgical site infection, not MSSA. A hospital could cut this number and move nothing that matters to a patient.

There is a second, quieter problem. The dots here are one year of one measure. Sentara’s 45% is a multi-year change across a multi-hospital system — a big enough denominator that the drop means something. At the US median hospital, with 3.6 predicted cases a year, a 45% drop is roughly what you would expect to see by chance about a third of the time with no intervention at all. The companion piece simulates exactly that.

The actual builder’s question

The Sentara result is not evidence that a risk score works. It is evidence that a risk score plus a rewritten nursing workflow plus enough volume to measure it works. Two of those three are free. The expensive one is the one nobody sells you.

So before the score: pick one alert already live at your site, write down the exact action it is supposed to trigger, who performs it, and where that action gets documented. If you cannot fill all three boxes, you do not have an intervention — you have a notification. And check your own denominator first. If it is under three, you will not be able to tell whether anything you build worked.