Readmissions Across Three EHRs and a 90-Day Claims Lag
You support analytics for a three-hospital accountable care organization. You are asked to build a 30-day all-cause readmission measure for 2023 discharges. Hospital A runs Cerner, Hospital B runs Meditech, and Hospital C is a critical access hospital on a legacy system that supplies only a claims extract. Claims from two payers covering 62% of attributed lives arrive with a 90-day lag. A regional HIE ADT feed captures emergency and inpatient events from 71% of facilities in the catchment area. About 22% of the ACO's discharges go to a skilled nursing facility. The medical director needs the number 'by Friday' for a payer contract negotiation, and has already told the payer the rate is 'around 12%.'
1. Define the numerator and denominator, and justify which data source you use for each component given the lag and coverage gaps.
Read the worked answer
Denominator (index discharges): live inpatient discharges from the three hospitals in the measurement year, excluding transfers to another acute hospital, excluding planned readmissions per the CMS planned-readmission algorithm (a specified set of procedures and diagnoses that do not count), and excluding patients who left against advice in ways that make follow-up impossible - document whichever exclusions you adopt, because the exclusion set changes the rate more than the model does.
Numerator: any acute inpatient readmission to any hospital within 30 days. Observation stays do not count as readmissions unless converted to inpatient; decide and document this, because the ACO's own hospitals may differ in observation practice.
Source strategy - use both, for different jobs:
- Claims are the only source that sees out-of-network readmissions, but the 90-day lag means 2023 data are still maturing. Use claims as the eventual gold standard.
- HIE ADT is timely and widens capture beyond the ACO's walls, but covers only 71% of facilities. Use it for interim reporting and to catch readmissions that never generate a claim in your two payers.
- EHR data are complete only for readmissions back to the same system - precisely the failure mode that biases rates downward for patients with high out-of-migration.
Reconcile, don't choose. Union the sources, flag the source of each readmission, and report the rate as provisional with a stated capture percentage. Validate against a chart-review sample of 100 discharges to estimate the residual undercount.
2. What data quality checks and mapping decisions must be resolved before the number is credible?
Read the worked answer
Run these as a documented checklist, and show the results to the medical director - data quality findings are the most persuasive reason to delay a number.
- Coverage/completeness by facility. Hospital C's claims-only extract will systematically undercount its readmissions and may not capture its own index discharges consistently. Compute completeness by site and report facility-specific rates.
- Identity resolution. Three source systems plus an HIE feed means duplicate and split records. Check the enterprise master patient index match rate; manually adjudicate a sample of unmatched records that appear in both ADT and claims.
- Date and sequence logic. Test for admit ≤ discharge, discharge date within the measurement year, and readmission date strictly after the index discharge. Watch for UTC/local time-zone mismatches on ADT feeds, which push same-day events across the 30-day boundary.
- Code set versioning. ICD-10-CM/PCS updates take effect October 1. Confirm every source's discharge dates are mapped to the correct fiscal-year code set; a stale mapping silently drops new codes.
- Disposition mapping. Each hospital's discharge disposition codes must map to a common vocabulary (home, home health, SNF, expired) - especially the 22% SNF population, whose readmissions are real but whose attribution rules differ.
- Measure specification lock. Record the steward specification version and the run date; a payer using a different specification version will produce a different number from the same data.
3. The medical director has already told the payer 'around 12%' and needs a number Friday. How do you handle this?
Read the worked answer
Deliver a number Friday, but deliver it as a bounded estimate with stated assumptions, not as a point rate.
What you produce:
- The best current estimate with a confidence interval (with roughly 1,200 discharges and a 14% observed rate, the interval is on the order of ±2 points).
- A sensitivity analysis across the three plausible capture scenarios: claims-only, ADT-plus-within-system, and the union estimate. Present the range, e.g. '12.4%–16.8% depending on how many out-of-network readmissions we can see.'
- A one-page data provenance note: sources, lag status, capture percentages, exclusions applied.
What you say to the medical director, directly and early: the figure already quoted to the payer is likely an undercount, not an overcount, because out-of-network readmissions are the least visible. It is far better to correct a low number in negotiation than to have the payer's own claims analysis produce a higher one.
What you do not do: quietly produce the number that matches 12%. That is the failure mode this whole exercise is designed to prevent, and it destroys the credibility of the measure permanently.
If the payer needs a contractual figure, propose reporting the 2022 measurement year (fully mature claims) as the benchmark, with 2023 shown as provisional.