clinicians.build · interactive · September 25, 2026 · built on BCBSA, Sep 24 2026

Who Moved the Code

BCBSA’s $942M is a total across hospitals. Audits don’t go after totals. They go after particular hospitals. Here is every Medicare hospital with usable data (…), showing how far its “with MCC” coding moved from 2022 to 2024 after adjusting for its condition mix, plotted against the band that chance alone would produce at its volume.

Primary source: Blue Cross Blue Shield Association, “BCBSA analysis: AI coding tools affect healthcare costs,” September 24, 2026 (coverage: Fierce Healthcare).
Data: CMS Medicare Inpatient Hospitals by Provider & Service, CY2022 and CY2024, covering 16 condition families (a DRG split into “with MCC” and lower tiers). CY2022 was cross-checked against MIMI Labs. CY2024 comes from data.cms.gov. Medicare fee-for-service, not the commercial claims BCBSA studied.

Read the BCBSA analysis → Or read today’s newsletter →

Measure

Condition family

Hide hospitals with fewer discharges than —

State

(single-family view only)

Find a hospital

Hospitals shown–

Moved >10 pts up–

Above the 99.8% band–

Below the band–

above band: more MCC coding than chance explainsbelow bandwithin noise

Stress-test it: the big jumps are small hospitals With every hospital shown, … hospitals moved more than 10 points toward MCC coding over two years. That’s the list that ends up in a slide deck. Only … of them sit above the 99.8% band. Drag the filter to 500 discharges and … hospitals are left with a 10-point jump, while … are still above the band. At big hospitals a 3-point shift is a signal. At a 40-discharge hospital a 20-point swing can be a handful of patients.

The middle didn’t move The median hospital’s adjusted change from 2022 to 2024 was …. In Medicare fee-for-service, the AI-coding years don’t show up as a shift across hospitals in general. If they show up at all, it’s in the tail: … hospitals above the band against … below it. Pure chance would put about 0.1% of hospitals, roughly …, on each side.

Turn on the suppressed tiers Pick one family, say GI hemorrhage, and click “include hospitals with suppressed tiers.” A new set of dots appears at impossible extremes. CMS hides any hospital–DRG cell with 10 or fewer discharges, so a hospital whose lower-tier cell vanished looks like it coded every case “with MCC.” Without that toggle, only a few dozen hospitals report every GI-bleed tier in both years. Any upcoding list built from the public file without handling this will name the wrong hospitals.

Where this is thin — read before quoting

A dot above the band is a question, not a finding. Real changes in acuity (a new trauma or transplant program, a shift of simple cases to observation, a merger that brings in sicker patients) look exactly like coding intensity in this file. No hospital here is shown to be upcoding.
No treatment data. BCBSA’s argument is that the code moved and the treatment didn’t. This file has codes and payments and nothing clinical, so it can locate movement but can’t test concordance.
Mix adjustment. “All 16 families” compares each hospital’s MCC count with what national shares predict for its condition mix, in each year. It uses only families where every tier was reported in both years. The bands use a binomial approximation at the national pooled rate and ignore overdispersion, so real bands are somewhat wider and the count above the band is an upper bound.
Medicare FFS only. BCBSA studied commercial claims from 2023 to 2025. A hospital can code the same way for every payer, but audit pressure differs by payer.
Two snapshots. Some hospitals may already have been using AI-assisted CDI in 2022, which would shrink the change measured here. Heart failure and respiratory infections are excluded because their DRG splits shift for definitional and COVID reasons.