Defect detectability as a Cross-Cutting Property of Data Quality: A Federated, Multi-Site Assessment of Routine Vital-Sign Data Across Nine Belgian Hospitals (Preprint)

Résumé fourni par la source

BACKGROUND Secondary use of electronic health records (EHR) data depends on data quality (DQ) to validate the repurposing of data. DQ assessment is usually performed one site at a time and organized by data quality dimensions. This approach is structurally blind to defects that are internally consistent within a site and become visible only when results across sites are compared. OBJECTIVE To assess DQ across a federated hospital network and to demonstrate that detectability, the scope of assessment required to reveal a defect, within a single site or only through cross-site comparison, is a property distinct from conventional data quality dimensions and independent of severity. METHODS We conducted a federated assessment of seven routinely collected vital-sign parameters across nine Belgian hospitals in three medical care departments (pediatrics, geriatrics, and surgery). Each site mapped local data to a study-specific common data model and ran locally a DQ assessment protocol, sharing only aggregate results. Completeness, consistency, correctness, and uniqueness were assessed and stratified by care department. A structured feedback loop returned site-specific anomalies for root-cause investigation. Each confirmed finding was classified by whether it was a genuine defect and, if so, whether it was self-detectable within a single site or relational, visible only through cross-site comparison. RESULTS The assessment covered 300.675 unique patients and 307.986 care episodes. Cross-site comparison surfaced defects invisible within any single site, including a heart rate and respiratory rate interchange arising from a parameter mapping error. Non-unique care-episodes identifiers were similarly detectable only against the cross-site distribution. Relational findings spanned consistency, correctness, and uniqueness simultaneously, conforming that detectability cuts across conventional dimensions. Detectability was also independent of severity: the heart interchange was severe yet invisible to single-site review, while biologically impossible values were equally severe but trivially detectable. A completeness signal in pediatric blood pressure reflected legitimate clinical practice rather than a defect, illustrating that a cross-site difference is a signal and not a data quality issue. CONCLUSIONS Detectability, the scope of assessment required to reveal a defect, within a single site or only through cross-site comparison, is a meaningful and distinct property of data quality problems in federated settings. It does not follow from the dimension label and is not determined by severity. Some defects are only visible when hospitals are compared, and distinguishing a genuine error from a difference in convention or legitimate clinical variation requires returning to the source. Cross-site comparison combined with a structured feedback loop is a practical and privacy-preserving approach to identifying data quality problems that single-site assessment cannot reach.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.