The best published AI model matched the exact MS-DRG on 52 percent of charts. Learn where AI DRG assignment breaks and what a coder still has to check.
Sometimes, and not without a coder. A Medicare Severity Diagnosis Related Group (MS-DRG) comes out of fixed CMS grouper logic, so grouping is never the weak step. Errors start upstream: AI has to pick the right principal diagnosis, capture every supported complication or comorbidity (CC) and major CC (MCC), and code the ICD-10-PCS procedures correctly.
In the most cited peer-reviewed study, the best model matched the exact MS-DRG on 52 percent of discharge summaries. That gap is why AI DRG assignment works as AI-assisted coding: the model proposes a working DRG with its evidence, and a Certified Coding Specialist (CCS) confirms the final one.
Why the grouper isn't the hard part
For discharges on or after October 1, 2026, CMS runs MS-DRG Grouper V44.0, which sorts every inpatient claim into one of 768 MS-DRGs. Feed it the same codes twice and it returns the same DRG twice. Any AI DRG error is really a coding error on one of the grouper's inputs.
| Grouper input | What AI has to get right | Typical AI failure | What coder review catches |
|---|---|---|---|
| Principal diagnosis | The condition chiefly responsible for the admission, after study | Picks the most severe or most mentioned condition | Sequencing rules, "after study" logic |
| Secondary diagnoses | Every CC and MCC the record supports | Codes a CC/MCC from a single mention with no clinical support | Clinical validation, CC exclusion against the principal diagnosis |
| ICD-10-PCS procedures | Correct root operation, body part, approach | Wrong root operation moves a medical DRG to a surgical one | Operative note review |
| Present on admission (POA) | Y, N, U or W on each diagnosis | Defaults to Y | A hospital-acquired condition coded N stops counting as a CC/MCC |
| Discharge status | Expired, home, transfer, post-acute | Reads status from a template, not the final note | DRG pairs split by status, such as acute MI alive (280 to 282) vs expired (283 to 285) |
What published research shows
Two peer-reviewed studies set the realistic baseline.
DRG-LLaMA (npj Digital Medicine, 2024) fine-tuned a language model on 236,192 discharge summaries from the MIMIC-IV dataset. Top-1 MS-DRG accuracy was 52.0 percent. When the task was split into base DRG and CC/MCC level, each reached roughly 68 percent. The model saw only the discharge summary, capped at 512 tokens, not the full record a coder works from.
A 2024 NEJM AI benchmark from Mount Sinai tested general-purpose models on code lookup. GPT-4, the best performer, matched the exact ICD-10-CM code 33.9 percent of the time. That study measured lookup from code descriptions, not chart coding, but it shows a general model without coding tools and code tables can't be trusted to produce codes on its own.
Neither result describes a commercial product. Production systems pair models with code tables, edits and rules, and their accuracy has to be proven on your own case mix. That is the question to put to any vendor of inpatient DRG coding software.
Where the errors carry real risk
Missed CCs and MCCs leave payment on the table. Unsupported ones create audit exposure, and that side is already under federal scrutiny.
The HHS Office of Inspector General found in 2021 that Medicare stays billed at the highest severity level rose almost 20 percent between FY 2014 and FY 2019, and that over half of those stays had only one diagnosis qualifying them for that level. A model tuned to find MCCs scales exactly that pattern.
Worked example. A patient is admitted with pneumonia, and sepsis is documented as present on admission. Under ICD-10-CM guideline I.C.1.d.4, sepsis is sequenced first when both conditions drive the admission, which moves the case from the simple pneumonia family (MS-DRG 193 to 195) to septicemia (MS-DRG 871 or 872). If the payer applies Sepsis-3 criteria and the record doesn't meet them, that DRG is denied on clinical validation. The AI coded the guideline correctly. The documentation didn't hold.
How to test a vendor's DRG accuracy claim
A single "DRG accuracy" percentage tells you almost nothing. Ask for these before you trust it:
- The denominator. How many charts, from which DRG families, over what dates.
- Working or final DRG. Was accuracy measured before or after coder review?
- The reference standard. Coder of record, a blinded CCS panel, or post-audit outcome.
- CC/MCC precision and recall, separately. One shows overcoding risk, the other shows missed revenue.
- Payment-weighted error. Missing one MCC on a high-weight surgical DRG costs more than ten misses on low-weight medical DRGs.
- Denials after go-live. Clinical validation and DRG downgrade rates in the six months after deployment.
A DRG match can also hide wrong codes. Two different code sets can land in the same MS-DRG, so a vendor reporting DRG match alone may be masking code-level errors that surface later in a payer audit.
What AI-assisted DRG coding looks like in practice
The workable model keeps a credentialed coder on the final DRG. The AI reads the full record, proposes a principal diagnosis, secondary codes and a working MS-DRG, and links each CC or MCC to the documentation that supports it. The coder confirms or corrects, and anything the record can't support goes to clinical documentation integrity (CDI) as a query instead of onto the claim. Every decision is logged, so the audit trail exists before a payer asks for it.
Related questions
Does the MS-DRG grouper itself use AI?
No. The grouper is deterministic software that CMS publishes each fiscal year with a definitions manual. The same codes always produce the same MS-DRG. AI only enters the process when it proposes the codes that go into the grouper.
Is AI DRG accuracy the same as coding accuracy?
No. A model can reach the correct MS-DRG with incorrect codes, because different code combinations can land in the same group. A sound evaluation measures code-level accuracy and DRG-level accuracy separately.
Why do inpatient programs keep a coder on the final DRG?
Because the costly errors are judgment calls: principal diagnosis sequencing, clinical support for a CC or MCC, and whether the record meets the criteria a payer will apply. Those decisions carry compliance exposure, so a CCS signs off on the final DRG and the decision is documented.
This page is general education on inpatient coding. It is not a coding determination for any specific encounter.
