Registry Data Quality in 2025: Still More Art Than Science, but AI Is Knocking

Seyyedeh Fatemeh Mousavi Baigi, Seyyed Mohammad Tabatabaei, Abbas Sheikhtaheri, et al.

Summary: This PROSPERO-registered systematic review synthesized 119 studies examining how data quality (DQ) in health and disease registries is conceptualized, measured, and improved, with particular attention to AI and machine learning approaches. Completeness and accuracy dominated as assessed dimensions, while representativeness and coverage were largely neglected. Conventional rule-based and audit methods remain the workhorses of registry DQ management, and AI/ML applications—though growing—remain selective and constrained by structural and interoperability barriers. The authors conclude that registry DQ management is still predominantly retrospective and fragmented rather than proactive and lifecycle-integrated.

Why it matters: Registries are increasingly central to regulatory decision-making, post-market surveillance, and real-world evidence generation, making data quality not a housekeeping concern but a scientific and policy-critical issue. The finding that representativeness and coverage—arguably the dimensions most consequential for generalizability and bias—are underassessed is a meaningful gap with direct implications for how registry-derived evidence should be interpreted. For rare diseases, where registries may capture a substantial fraction of the known population, completeness and representativeness failures can silently distort epidemiological estimates and treatment-effectiveness analyses.

Expert perspective: This is a methodologically conscientious review—PRISMA 2020 adherence, broad multi-database search through late 2025, and Mixed Methods Appraisal Tool quality appraisal all suggest reasonable rigor. That said, systematic reviews of heterogeneous registry literature inherit the limitations of their constituent studies: variable DQ definitions, inconsistent reporting standards, and a publication landscape that skews toward positive or novel findings. The thematic synthesis approach is appropriate but inherently qualitative, and the 119-study corpus likely reflects substantial heterogeneity in registry type, disease domain, and healthcare context that resists clean aggregation.

The finding that AI/ML remains selectively applied is credible and unsurprising. Anomaly detection and record linkage are natural entry points for ML precisely because they map onto well-defined computational tasks; fuzzier quality dimensions like clinical plausibility or longitudinal consistency are harder to automate without robust reference standards—which, as the review notes, are often absent. The call for "lifecycle-integrated quality ecosystems" is directionally correct but operationally underspecified, which is a common limitation of synthesizing aspirational literature.

For registry professionals and informaticists, the most actionable takeaway is the persistent neglect of representativeness and coverage assessment—dimensions that require external data linkage and population denominators, not just internal audits. Advancing these requires investment in interoperability infrastructure and data governance, not just better algorithms. The AI enthusiasm in this space is real but should not obscure the more foundational work still needed in standardized data definitions and sustainable audit processes.

Relevance: Registry and Real-World Data Methods

Source: https://pubmed.ncbi.nlm.nih.gov/42813616/

---