20.2 Mortality Reviews and Observed-to-Expected Ratios

20.2 Mortality Reviews and Observed-to-Expected Ratios

Quick Answer: An observed-to-expected (O/E) mortality ratio below 1 generally means fewer deaths than expected; a ratio above 1 generally means more deaths than expected. Clinical documentation integrity (CDI) can improve a high ratio by raising expected mortality through accurate severity capture. It does not improve the ratio by changing who died.

Section 20.1 established that coded SOI/ROM (and similar risk engines) feed expected deaths. This section is the operational half: how to read the ratio, how a mortality review uses it, and where CDI may and may not intervene. Domain VIII weights analysis heavily here. A specialist who can define ROM but cannot explain why ten deaths with a thin coded profile look worse than ten deaths with a complete profile will miss the item.

Observed, expected, and the ratio

Observed deaths (O) count decedents who meet the measure’s inclusion rules. Typical inpatient profiling uses in-hospital death (discharge status expired). CMS 30-day mortality measures count death within 30 days of admission, including deaths after discharge, using a claims-based model (Chapter 21). If you mix those cohorts, you will “investigate” the wrong list.

Expected deaths (E) are not a guess typed by quality staff. They are the sum of predicted probabilities the model assigns to each stay. A ROM 1 pneumonia discharge might contribute a small probability; a ROM 4 septic-shock stay contributes a larger one. Add the probabilities across the cohort and you have E. When documentation is thin, each stay’s probability is too small, E is too small, and O/E rises.

O/E = O ÷ E (usual teaching form)

Vendors sometimes scale the ratio (index = O/E, or index = O/E × 100). Always read the axis. CCDS items still expect the directional reading above for the unscaled O/E. CMS does not publish a single national “pass” cut score for a hospital’s APR-DRG mortality index. Do not invent 0.80 or 1.20 as an official threshold. Use those figures only as arithmetic examples.

Same deaths, different expected: why CDI can move O/E

Worked teaching case (illustrative counts, not a published benchmark):

• Cohort: 400 medical discharges, 10 observed deaths.
• Month A, thin coding: expected deaths 7.5 → O/E = 10 / 7.5 ≈ 1.33.
• Month B, same 10 deaths, complete supported secondaries: expected deaths 12.5 → O/E = 10 / 12.5 = 0.80.

Nothing about who died changed. Expected changed because the risk inputs changed. That is the entire CDI mechanism on this metric. Programs that celebrate a falling O/E should be able to show what newly reportable, clinically valid diagnoses appeared—not a campaign to keep dying patients out of the file.

Compliant CDI cannot lower O by:

• Recoding discharge status from expired to home, hospice, or transfer to hide the death.
• Dropping death encounters from the quality extract.
• Asking providers to delay pronouncement until after discharge.
• Querying “document this so the death looks expected.”

Those actions would be compliance failures, not documentation integrity. The 2026 ACDIS/AHIMA query standard still bars quality-outcome language in the query itself. Mortality-review education can teach physicians why shock and acute respiratory failure matter; the individual query still has to stand on clinical indicators and independent provider judgment.

What a mortality review actually does

A mortality review is a structured look at deaths (or at deaths that look statistically unexpected) to answer: Did the coded risk match the clinical course, and was the death a documentation problem, a coding problem, a true quality-of-care problem, or a mixture?

Typical concurrent or retrospective steps:

1. Pull the list using the same cohort the dashboard uses (in-hospital versus 30-day; all payers versus Medicare only; exclusions for transfers or hospice if the model uses them).
2. Read the record for the death: presenting illness, trajectory, procedures, code status, palliative decisions, and whether shock, respiratory failure, organ failure, metastatic disease, and similar high-ROM conditions were named, linked, and reportable.
3. Compare coded SOI/ROM (or the vendor risk score) with that clinical story. A death with ROM 1 and a chart full of vasopressors and mechanical ventilation is a documentation/coding signal until proven otherwise.
4. Separate issues. Missing acute respiratory failure is a CDI/coding pathway. A missed perforation after elective surgery with complete coding may be a quality pathway. Both can exist on one stay.
5. Query or recode only what the record supports. Clinical validation applies to decedents the same as to survivors. Do not “complete” ROM with unsupported terms.
6. Feed education and systems. If one service line repeatedly under-names shock, that is a physician-education and template problem, not a reason to write leading queries on every death.

Palliative care, comfort measures, and do-not-resuscitate orders do not by themselves remove a death from observed counts in most in-hospital models. Some 30-day measures have measure-specific exclusions (for example, certain hospice or transfer rules). Those exclusions are specified by the measure steward. CDI should not invent an exclusion by changing disposition. Documenting the palliative diagnosis and the conditions being palliated still matters for expected risk when those conditions were treated or evaluated.

Interpreting a high O/E without jumping to one story

A ratio above 1 is a signal, not a verdict. Walk the differential:

• True excess mortality — complications, delays, unsafe processes. Quality and medical staff own this; CDI still makes sure the coded risk is honest so the signal is not exaggerated.
• Under-documentation / under-coding — expected too low. This is the CDI-heavy explanation and the one Domain VIII expects you to recognize.
• Mix and model — a new extracorporeal membrane oxygenation program, more transfers-in of dying patients, or a model that fits poorly can move O/E even with decent coding. Do not treat every bump as a query-rate failure.
• Over-coding elsewhere in the year — if last year’s expected was inflated by unsupported codes, this year’s “worsening” after an audit cleanup can be honesty, not worse care.

Low O/E is also not automatic proof of excellent care. It can reflect excellent outcomes, a model that over-predicts, or unsupported severity capture. Clinical validation belongs in mortality work for that reason.

Who sits at the table

Mortality review is rarely a CDI-only meeting. Quality, coding, the attending service, palliative care, and sometimes infection prevention or surgery bring different files. CDI’s distinctive contribution is whether the coded severity is a fair statistical description of the death. Coding’s contribution is guideline-compliant assignment once the documentation is clear. Quality’s contribution is whether processes failed. Collapsing those roles—CDI “fixing” O/E by any means, or quality assuming every unexpected death is a nurse-abstracted Patient Safety Indicator—produces the wrong work queue. Patient Safety Indicators are claims-based (Section 20.3); mortality O/E is a ratio; they can appear in the same committee without sharing a data collection method.

Traps

• Reading O/E < 1 as “we billed too many MCCs” or O/E > 1 as “stop documenting deaths.”
• Trying to move observed instead of expected.
• Using one death’s ROM as a hospital-wide grade.
• Mixing CMS 30-day mortality with in-hospital APR-DRG O/E in the same numerator and denominator.
• Treating a high O/E as proof of bad care without opening the record, or as proof of undercoding without looking at complications.
• Putting “improve our O/E” language in the query text.