LOINC Mapping and Unit Harmony Before a Multi-Site Lab Feed
Three hospitals are feeding a research registry with 240,000 chemistry results collected over 18 months. Local test codes include K, K+, and POTASS. Units are recorded as mmol/L, MEQ/L, and mEq/L. One site reports potassium from serum, another from plasma, and a third occasionally from urine under the code K. The vendor proposes mapping all three local codes to LOINC 2823-3 (Potassium in serum or plasma) and claims 99.8% mapping accuracy based on a random sample of 500 results. Overall, 7% of local codes remain unmapped, 2.1% of results carry the unit string MEQ/L, 0.8% have a preliminary status that is later overwritten, 12% lack a collection time, and duplicate rows appear when add-on tests are ordered. The registry team wants to go live next month.
1. Should the committee accept the vendor's 99.8% mapping-accuracy claim as sufficient for go-live? Justify your answer quantitatively and clinically.
Read the worked answer
No. A 99.8% estimate from 500 results means roughly one error in the sample, and the exact 95% confidence interval extends from about 98.8% to nearly 100%. At 240,000 results, a 0.2% error rate implies about 480 mis-mapped results, and the upper confidence bound implies far more. More important, the sample appears to exclude the 7% of codes that are unmapped, so the denominator is not the full feed. Accuracy must be reported by stratum: site, specimen type, unit string, result status, and local code. Clinically, mmol/L and mEq/L are numerically equivalent for potassium, so that unit variation is low risk if normalized explicitly. The dangerous errors are specimen and code confusion: a urine potassium mapped to serum potassium can create a false hyperkalemia, and plasma versus serum differences affect reference ranges. The vendor's aggregate figure cannot rule out a site-specific or code-specific failure. A 99.8% overall accuracy is not a sufficient acceptance criterion for a variable that drives clinical or research classification; require stratum-level positive predictive value with confidence bounds and explicit handling of the unmapped and duplicate records.
2. Design a pre-go-live validation that would detect the failure modes described.
Read the worked answer
Use stratified random sampling rather than a single random draw. Strata should include each site, each local code (K, K+, POTASS), specimen type, unit string, and result status. Draw at least 100 results per high-risk stratum, or enough to estimate positive predictive value with a lower 95% confidence bound above 99% for serum potassium. Compare the registry value against the laboratory information system as the gold standard, including LOINC concept, specimen, unit, collection time, and result status. Use two independent abstractors for a subsample and adjudicate disagreements. Add deliberate challenge cases: known urine potassium results, hemolyzed specimens, MEQ/L records, preliminary results later finalized, and add-on duplicates. Confirm that unit conversion and normalization are reproducible. Pre-specify acceptance criteria, for example no specimen misclassification, no systematic unit error, and a lower confidence bound for serum potassium mapping above 99%. Run a parallel feed for two weeks and reconcile every discrepant result before go-live. Retain the validation dataset and query logic for audit.
3. Specify how the registry should handle unmapped codes, preliminary results, and duplicate rows.
Read the worked answer
Unmapped codes should never be silently dropped. Preserve the raw local code and display text, map to a local 'unmapped' concept, and flag the record for curator review. Report the unmapped rate by site and code weekly; if a code represents a test that is out of scope, exclude it with a documented reason. Preliminary results should retain status and version; the registry should carry the final result only when a final status exists, and should keep the preliminary value in an audit table rather than overwriting it. This prevents a moving target in analyses. Duplicate rows from add-on tests should be resolved using laboratory accession number, test code, collection time, and result status. Keep the final result, and when multiple final results exist for the same accession and test, apply a pre-specified rule such as latest verified result with a flag. Never deduplicate on patient and date alone, because that can merge distinct specimens. Document all rules in the data dictionary and apply them consistently across sites so that the registry's counts are reproducible.