19.2 External Benchmarking: CMI, Severity, Mortality and Length of Stay

External Benchmarking: CMI, Severity, Mortality and Length of Stay

An external comparison is useful only when its population and measurement rules fit the question. A different case mix, grouper version, data source or observation window can change a result without any change in documentation quality. Use the comparison to identify a question for investigation, then examine the records and applicable methodology.

Choose the correct source and release

The CMS MedPAR description identifies it as a stay dataset. It is not itself the annual Table 5 of current DRG weights. The FY2027 IPPS final-rule page separately lists Table 5, correction notices and other files. For example, its published CMI file describes historical FY2025 claims grouped with v42; it should not silently be treated as a v44 re-grouping of the same cases.

Calculate and interpret case mix index

For a defined population using one specified set of MS-DRG weights:

$\text{CMI}=\frac{\sum_g n_gw_g}{\sum_g n_g}$

Here $n_g$ is the number of included discharges in group $g$, and $w_g$ is that group's weight. The denominator counts discharges, not inpatient days. The weight describes relative resource use under the classification; CMI is not a direct measurement of every patient's physiologic severity or the hospital's quality of care.

Hypothetical training weights—not actual CMS DRG values:

CMI is $360/300=1.20$. Do not attach those training weights to named current DRGs. For actual analyses, identify the annual weights and corrections used.

Medicare and all-payer comparisons also need explicit inclusion rules. An all-payer MS-DRG CMI may apply MS-DRG weights to included discharges regardless of how each payer actually pays. It should not mix MS-DRG and APR-DRG relative weights into one unexplained average. Under HFMA MAP FM-5, normal newborns and specified Medicare-exempt units are excluded. A custom report including them measures a different population.

There is no universal 0.30–0.50-point gap between Medicare and all-payer CMI. Obstetric and neonatal services can change a population's mix, but v44 MS-DRG 795 is Normal Newborn, not normal vaginal delivery. Check the actual definition and inclusion rules instead of attributing every difference to nursery days.

Decompose changes before attributing them to CDI

A new high-resource surgical service, referral change, fewer low-weight discharges, different payer mix, coding correction or updated weights can change CMI. Compare service-line composition and within-group results using compatible definitions. A before-and-after rise alone does not isolate a causal CDI effect.

Case-level validated documentation and coding changes can demonstrate specific contributions. Preserve the original and final grouping assumptions and distinguish newly available clinical information from a correction. Do not label all within-service variation “true CDI yield” or assume that a rise is always desirable.

Keep APR-DRG severity and mortality distinct

Solventum's APR-DRG description distinguishes the base APR-DRG from separate severity-of-illness and risk-of-mortality subclasses. The four subclass labels are minor, moderate, major and extreme. SOI and ROM are different dimensions; neither is the same thing as an MS-DRG CC/MCC designation.

The complete case and the applicable APR-DRG version determine its classification. A condition does not universally increase a subclass, and the same subclass does not mean one fixed individual probability of death in every context. Compare distributions only after checking the population, version and method. An academic hospital with a particular proportion in minor/moderate categories is not thereby proven to have “massive under-capture.” There is no universal community-versus-academic percentage table that establishes that conclusion.

Read an observed-to-expected mortality ratio

For an explicitly defined cohort and outcome:

$\text{O/E}=\frac{\text{Observed deaths}}{\text{Expected deaths}}$

In a model that assigns each included patient a predicted death probability, expected deaths are the sum across all included patients, including survivors. The outcome may be inpatient mortality or another defined interval; use the actual measure specification. Claims, clinical records, demographic information or other sources may contribute, depending on the model.

Hypothetical example: Ten observed deaths divided by eight expected deaths gives $10/8=1.25$. If the same ten observed deaths are compared with ten expected deaths, the ratio is 1.00. That demonstrates the denominator's mathematical effect. It does not establish that adding a particular diagnosis would cause that change in a real model.

Missing valid risk information can affect a model when it uses that information in the relevant time window. But not every additional diagnosis, CC or MCC is an eligible predictor, and not every model uses claims alone. Verify the risk factors, timing, exclusions and version. The CMS risk-adjustment overview and procedures explain those model-specific choices.

A ratio below 1 means fewer observed deaths than expected under that model; a ratio above 1 means more. Interpretation also needs sample size, uncertainty, data quality and model adequacy. It does not automatically prove superior or negligent care, establish a public star rating or trigger a payment penalty. The applicable quality program determines its own measures and consequences. Keep quality-score improvement out of patient-specific queries.

Interpret LOS benchmarks without changing diagnoses to fit them

CMS Table 5 publishes geometric and arithmetic mean LOS separately. A geometric summary differs from an arithmetic average. A hospital's actual arithmetic average divided by published GMLOS is a chosen comparison, not a universal efficiency standard requiring a result at or below 1.00.

For illustration, the FY2027 corrected Table 5 gives these values:

These are version-pinned reference values, not instructions to assign a higher tier. A four-day stay divided by 2.1 is approximately 1.90; divided by 3.8 it is approximately 1.05. Neither calculation determines the patient's diagnosis or proves efficient or inefficient care.

In a case appropriately classified in this family, unspecified AKI N17.9 is a CC, not automatically an MCC placing the case in 291. Principal-diagnosis exclusions and the full case also matter; do not assume pulmonary edema automatically adds an effective MCC to heart failure. Check the actual v44 definitions and exclusions.

Payment rules are another question. IPPS cost outliers depend on applicable cost/payment thresholds, not merely exceeding GMLOS. GMLOS also appears in specified transfer-payment rules and certain transfer outlier calculations; saying it never matters to outliers would also be wrong. See CMS outlier guidance and the Claims Processing Manual, Chapter 3, Section 20.1.2.4. Use the particular payment rule rather than treating a dashboard ratio as a claim calculation.

Use PEPPER as an audit signal

As of October 2026, the official PEPPER FAQ identifies Index Analytics LLC, with its partners, as the CMS contractor. STACH reports are quarterly; other facility types have their own schedules. Review the report's target definitions rather than treating PEPPER as a universal physician scorecard or a single hospital-wide CC/MCC capture measure.

Distinguish a target-area rate from its percentile rank. In a hypothetical coding-focused target, 18 target discharges among 60 eligible discharges gives 30%. Its national percentile tells where that rate falls among comparison facilities; it is not necessarily the 30th percentile.

PEPPER flags high outliers at or above the 80th percentile. The low boundary is at or below the 20th percentile for coding-focused target areas. These are inclusive boundaries, not simply “above 80” and “below 20.” Low-outlier logic does not apply indiscriminately to every target.

An outlier is not proof of a payment error, leading queries, underdocumentation or a required recoupment. Confirm the target's current definition, population and comparison group; review representative claims and records to understand the pattern. Preserve valid coding, correct demonstrated errors and monitor the result. Neither a percentile nor a desired CMI authorizes adding unsupported diagnoses or suppressing supported ones.