clinicians.build · interactive · september 3, 2026
Two Denominators
A Medicare add-on payment attaches to cases where the technology was used. A positive predictive value describes the alerts it got right. Same product, two denominators — and only one of them appears on the claim. Here are a thousand dots, colored both ways.
Primary source: “Artificial intelligence tools in sepsis prediction: a systematic review and meta-analysis,”
npj Digital Medicine, August 31, 2026 — 53 studies, >7 million admissions, pooled PPV 34.2%
Payment figures: Bayesian Health NTAP approval (PR Newswire, Sep 2, 2026) · CMS FY2027 IPPS final rule
Starting October 1, hospitals can bill up to $61.84 per case for a continuous AI sepsis monitor, across an estimated 739 MS-DRGs. It is the first line in the Medicare inpatient payment system drawn for continuous monitoring software.
The obvious objection is that 82% sensitivity is not good enough to pay for. That is the wrong number to be angry about. Watch what happens when the same thousand dots get counted two different ways.
The other reflex is that $61.84 is a rounding error. It is. Drawn to scale against the average Medicare payment for DRG 871 — septicemia or severe sepsis with MCC, $15,105 — the add-on is the red edge below. That is 0.41%.
But DRG 871 is the single largest MS-DRG in Medicare fee-for-service inpatient — 561,795 discharges across 2,678 hospitals. A sliver of the biggest denominator in the building is $34.7M a year at perfect capture, and it is the first line in the system with continuous monitoring software's name on it.
80/20 lens
If you are building an inpatient tool, the number that decides your revenue is not your AUC and not the per-case amount. It is which MS-DRGs the add-on covers and how much volume those DRGs carry. Sepsis, heart failure, pneumonia and respiratory infection are 27.8% of all Medicare inpatient discharges between them. Twenty-two DRGs are half. The companion map plots all 534.
Where this argument is thin
The two grids are not the same population, and that is the whole point — but it also means you cannot subtract one from the other. Denominator 1 is admissions on an eligible DRG where the monitor ran. Denominator 2 is alerts the monitor fired. This piece deliberately draws them at the same size to make the framing visible; the real alert rate per admission is not published, so any “dollars per alert” figure here is a ratio on a hypothetical thousand, not an observed one.
34.2% is a pooled summary, not a property of any one model. It comes from 53 heterogeneous studies with an accelerated-preview publication and no reported confidence interval. PPV moves with prevalence and with wherever a site set its threshold. A monitor tuned tight in a high-acuity ICU and one tuned loose on a med-surg floor are both inside that pooled number.
Bayesian's own device is not the meta-analysis. Its 510(k) performance is its own; nothing here is a measurement of that product. The pooled figure is the prior you should hold before a vendor shows you site data, not a verdict on the vendor.
False positives are not free, and this does not price them. The cost of a wrong sepsis alert is a nurse's attention, a lactate, sometimes an antibiotic. None of that is in the $61.84 and none of it is in this graphic.
“Up to” and “three years.” $61.84 is 65% of a $95.14 average per-case cost, capped at three years. It is a ramp, not an annuity.
The argument worth having is not whether the model is good enough. It is that Medicare just built a payment whose denominator is usage, for a class of product whose failure mode is over-firing. Those two facts point the same direction, and nobody has to be dishonest for it to go badly.