A pre-bill review tool that compares the coded DRG against the captured encounter is not a documentation product. It is a dial, and the dial already has a position at every hospital in Medicare. Here is where each one sits — the share of cases it codes to the with-MCC side of the three biggest severity-split DRG families.
Three DRG families carry most of the severity money in medical inpatient care, and all three split cleanly on one question: did a condition from the CMS major-complication list make it into the documentation?
Sepsis is the largest inpatient family in Medicare. In 2023 it ran 666,174 discharges across the two codes, and 84.3% of them landed on 871 — the with-MCC side. Urinary tract infection splits near half. Fluid and electrolyte disorders, just over half. The gap between the two codes is $7,763, $2,715 and $4,537 a case respectively.
Every hospital is already somewhere on this dial. The question a pre-bill tool raises is not whether it can move — it is whether anyone would be able to tell that it did.
One dot per hospital per family. Horizontal position is the share of that family’s cases it coded to the with-MCC side. Dot area is its volume. The vertical line is the national pooled rate. Colour asks the only fair question about a rate: is this hospital far enough from the national rate that chance can’t explain it, given how many cases it had?
The honest move with any rate table is to ask how much of the spread is sample size. Drag the minimum-cases slider up and most hospital scorecards collapse: the outliers were small hospitals having a bad year, and once you require real volume the differences fall inside the noise.
That is not what happens here. At sepsis, with every hospital in the file included, 720 of 2,224 — 32% — sit outside the range binomial noise would explain around the national rate, and 297 sit outside a 99.8% range. Push the volume floor up and that share does not fall. It rises: 31% of the 1,861 hospitals with 100+ cases, 40% of the 516 with 400 or more. Severity coding is not mostly noise. It is a real, stable, hospital-level difference in how often the word gets written down.
Which is exactly why it is a product surface. You cannot sell a tool that reliably moves a number that doesn’t move.
Pick your own hospital out of the cloud above and write down the number. If you ship anything that touches documented severity, that number is your baseline, and the only defensible version of “it worked” is a change you can attribute case by case: which human reviewed which machine-surfaced finding, when, and what they changed. Build that artifact before the feature, not after the audit letter.
Also today: the price of one word → — all 43 Medicare DRG pairs where a single documented complication changes the payment, on one chart.