Two Readmission Numbers and $4.2 Million at Risk
Lakeview Health is a three-hospital system preparing its first enterprise data warehouse. Two readmission reports disagree: the quality department's congestive heart failure 30-day readmission rate is 21.4%, while the payer's monthly report shows 17.8%. The difference is attributed to inclusion of observation stays, transfers between the system's own hospitals, and patients who leave against medical advice. Lakeview's CFO has signed a shared-savings contract under which $4.2 million is at risk if readmissions exceed a 19.5% threshold, and the payer uses its own measure specifications. Meanwhile, the CMIO has been asked to approve a vendor sepsis alert validated externally but never tested on Lakeview's population; the vendor's contract limits access to the model's logic. Clinicians already receive 42 alerts per day per physician and override 88% of them. The board's quality committee meets in three weeks and wants one number it can trust.
1. What governance structure and processes should Lakeview put in place before the committee meets, and who should own each?
Read the worked answer
Build governance before choosing a number.
- Charter a data governance council with decision rights. Chair it with the CMIO or CNO and include quality, finance, compliance and privacy, IT, and clinical service representation. Its job is to approve certified metrics and resolve disputes, not to build reports.
- Assign data stewardship by domain. Quality owns the readmission definition; finance owns cost and contract measures; IT owns the platform, pipeline, and access controls. The most common failure is holding IT accountable for what a number means, when IT can only guarantee that the code matches a specification someone else approved.
- Create a metric catalog with a unique identifier, plain-language definition, numerator and denominator, exclusions, source system, owner, version, and effective date. Every dashboard must cite the identifier.
- Institute change control. No definition changes without council approval and a version note; either restate historical trends or annotate a break in series.
- Monitor data quality: completeness, timeliness, validity, and periodic reconciliation checks, with a log of known discrepancies.
For the three weeks available, convene the council once, appoint a readmission steward, and commit to presenting a variance bridge rather than a single reconciled number. Promising one number before the definitions are reconciled guarantees the same dispute in six months.
2. How should the readmission discrepancy be resolved, and what should be presented to the quality committee?
Read the worked answer
Reconcile specifications first; the arithmetic is the easy part.
Build a side-by-side specification table: index admission cohort (age, payer, diagnosis), planned versus unplanned readmissions, observation stays, transfers within the system, discharges against medical advice, in-hospital deaths, multiple readmissions per patient, and risk adjustment. The 3.6-point gap is fully explainable by observation stays and internal transfers; each is a defensible policy choice, but they are different measures.
Rerun both specifications on the same data to isolate definitional differences from pipeline differences. If the gap disappears, the problem is definitional. If a residual remains, there is a data problem to trace through the warehouse.
Separate the two uses:
- The contract number governs the $4.2 million and must follow the specification the contract actually references, with a written reconciliation and a formal discrepancy process. Confirm the contract language before assuming the payer's 17.8% is authoritative.
- The improvement number should be a clinically meaningful measure used consistently by care teams, typically all-cause, 30-day, risk-adjusted, excluding planned readmissions.
To the committee: present one recommended measure with its definition, both figures with a variance bridge, the financial exposure under the contract specification, and the remediation plan. Naming the discrepancy openly protects credibility more than a premature single number that collapses under challenge.
3. Should Lakeview deploy the vendor sepsis alert? Justify your decision with a decision framework.
Read the worked answer
Do not deploy it yet. Approve a staged validation with pre-specified criteria.
Local performance. External validation is necessary but not sufficient. Sepsis prevalence, coding, and documentation vary by site, and calibration typically degrades. Run a retrospective silent evaluation on 12 to 24 months of local data and report discrimination (AUROC), calibration across risk deciles, positive predictive value at the proposed threshold, and subgroup performance by age, race, language, and renal status. A model that underperforms in a subgroup creates an equity problem that will surface later and publicly.
Alert burden is a clinical risk, not a convenience issue. At 42 alerts per physician per day and an 88% override rate, clinicians are already trained to dismiss alerts. Adding a low-specificity alert can degrade response to the alerts that matter. Model incremental alert volume at several thresholds and choose one that produces a manageable volume with high positive predictive value.
Vendor opacity. A model whose logic cannot be inspected limits your ability to validate, monitor drift, explain a miss, or meet regulatory and safety expectations. Request calibration documentation, intended-use labeling, and a right to performance data; refusal is a governance finding, not merely a procurement detail.
Decision: approve a time-limited silent-mode pilot in two units with a pre-specified stop rule (for example, positive predictive value below 20%, or more than 10 alerts per clinician per shift), clinician co-design of the workflow, and a monitoring plan for drift and subgroup performance. Reject full deployment until the pilot meets criteria.