The Merged Clinic's Inconsistent Diabetes Registry
A regional primary care network has just merged three clinics onto a single electronic medical record. Each clinic used its own way of recording diabetes: Clinic A used a structured problem-list code (SNOMED 44054006); Clinic B used a free-text problem-list entry reading 'DM2'; Clinic C recorded only the HbA1c values in a flowsheet with no diagnosis code. The quality team wants to report the network's diabetes registry size and the proportion with a current HbA1c (within 12 months). The data analyst runs a query on the structured code alone and reports 1,240 patients. The clinical lead believes the true number is closer to 1,800. There are 2,050 patients with at least one HbA1c result in the last 18 months. The board wants the number by Friday for a funding application.
1. Identify the data-quality problems in this registry and recommend how the analyst should build a defensible case definition.
Read the worked answer
The core problem is definitional inconsistency across source systems, which is a data-quality issue of comparability and completeness, not a coding error. A query on SNOMED 44054006 captures only Clinic A, so the 1,240 figure measures one clinic's coding habit, not the network's diabetic population. Clinic B's free-text 'DM2' is not computable without mapping; Clinic C has clinical evidence (HbA1c) but no diagnosis assertion.
A defensible approach:
- Define the phenotype explicitly and document it: e.g. 'any patient with a structured diabetes code OR a mapped free-text diagnosis OR two HbA1c results ≥6.5% at least 3 months apart.'
- Map free text using a validated terminology (SNOMED CT) with a manual review sample to estimate mapping error.
- Validate against a reference standard - chart review of a random sample (e.g. 100 patients) to estimate sensitivity and positive predictive value of each rule.
- Report uncertainty, not a single point estimate: give the range and state which sources are included.
The 2,050 figure is an upper bound on tested patients, not on diagnosed patients - some will be prediabetes or screening. Never present the structured-only count as the registry size without a caveat.
2. The board wants a single number by Friday. Draft the framing you would give, and justify it.
Read the worked answer
I would give the board a number with an explicit definition and a stated range, and refuse to give a bare point estimate that implies false precision.
Suggested framing: 'Using a harmonised definition across all three clinics, we estimate 1,650–1,850 patients with diabetes (best estimate 1,750). The previously quoted 1,240 reflected only one clinic's structured coding and undercounts by roughly 30%. Of the estimated cohort, X% have an HbA1c within 12 months.'
Justification:
- Funding decisions are consequential; an undercount could cost the network resources, and an overcount could misallocate them. Both errors are avoidable by stating the range.
- Transparency about method protects the analyst if the figure is later revised; the definition is auditable.
- A deadline is not a reason to drop validation. A rapid chart-review sample of 50–100 records can be done in a day and materially tightens the estimate.
I would also flag the known gap: Clinic C patients identified only by HbA1c are provisional until diagnosis is confirmed, so the upper bound should be labelled provisional.