Sepsis Compliance Fell Eleven Points: Validating Before the Board Sees It

Sequence matters here because the cheapest, most likely explanations should be eliminated first.

1. Recompute prior quarters under the new denominator definition. If the historical rate drops similarly, the change is a measurement artifact; if the drop persists only in Q3, the definition is not the explanation.
2. Check data latency and completeness. The nightly refresh means late-arriving ED records and open encounters can sit in the denominator without their elements - a false decline that resolves in two weeks.
3. Trace one patient end-to-end through the EHR to the warehouse. The flowsheet row rename is a prime suspect: if the interface mapping was not updated, completed elements land in an unmapped row and are scored as missing. This would produce a uniform, system-wide drop of exactly the kind seen.
4. Re-abstain a sample from Q2 and Q3 with a second reviewer. The vendor change introduces inter-rater variability; confirm the new abstractors were calibrated.
5. Stratify by unit and check case mix and denominators. A 28-bed unit with a handful of eligible patients produces volatile rates; the progressive care conversion changes acuity and eligibility.
6. Review the order set change for whether it altered time zero or element definitions in a way that shifts who counts.

Steps 1 to 4 test measurement failure; steps 5 and 6 test real change. Do not proceed to naming units until at least steps 1 to 3 are complete.

Present an annotated statistical process control chart rather than two bars, because the board needs to see whether a change is signal or noise. Include:

• Monthly or weekly rates with control limits and annotations for each known change (order set, vendor, denominator, upgrade) so the audience can see candidate causes in time order.
• The rate computed under both the old and new denominator, side by side, with a short note on why the definition changed.
• Denominators and confidence intervals printed with every rate. A unit with 14 eligible patients should not be ranked against one with 90.
• Stratification by unit and by service line, with case-mix notes where relevant.
• Explicit data-confidence language: what is validated, what is still under review, and the date of the next report.

The framing is: here is what we know, here is what we are still testing, here is the action already underway. The board is entitled to the caveats, not just the trend. If validation is incomplete, present the validated Q2 figures as the last reliable point and state the Q3 status as provisional. This protects credibility: a number retracted next quarter costs more trust than a delay of ten days.

First, separate real from causal. Check whether those two units have process differences the others lack: the new order set not set as default, lower order set adoption, a staffing change, lactate draw timing moving from triage to later in the encounter, or a nurse educator vacancy. Pull a focused chart audit of 20 to 30 cases per unit to see which bundle element is failing and when in the sequence - a uniform failure of one element suggests a workflow obstacle, while scattered failures suggest knowledge or staffing.

Second, bring the units in before the board does. Share the validated data with the managers and a sample of staff, ask what changed on their unit in Q3, and let them correct you. They often know the answer, and they will not own a number that arrives as an accusation.

Third, run a short rapid-cycle improvement on the identified element with a weekly measure, and report the two units as a system-supported improvement project rather than a performance failure.

Finally, tell the board the same thing in the same words: two units, one element, this cause, this action, this review date. Naming units without case-mix adjustment and root cause invites defensiveness and gaming of the documentation, which corrupts the metric going forward.