Presentation (CTI)

OBJECTIVES: Algorithms identifying patients from routinely collected health data use indirect indicators of disease status, leading to misclassification and require validation. This evaluation is constrained by gold standards restricted to confirmed cases, leaving non-cases unidentified and limiting estimation of key metrics, particularly specificity and negative predictive value. This review assesses the use of Bayesian latent class models (BLCMs) for algorithm validation in medico-administrative databases, focusing on settings with only positive cases (e.g., registries, clinical cohorts).
METHODS: A structured literature search was conducted on PubMed to identify studies using BLCMs to validate phenotyping algorithms in medico-administrative databases with gold standards restricted to confirmed cases. The equation combined keywords (BLCMs, healthcare database, gold standard, algorithms). Titles and abstracts were screened, followed by full-text review. Performance measures (e.g., sensitivity, specificity) were extracted.
RESULTS: Six studies using BLCMs in administrative health databases were found, with incomplete or absent gold standards focusing on case definitions over strict algorithm validation. Three applications for BLCMs emerged: (i) direct inference of latent disease status (n=4), (ii) construction of a latent standard for model evaluation (n=1), and (iii) evaluation of predefined algorithms using BLCM-derived reference (n=6). In one study, BLCMs improved sensitivity (95.9% vs 91.9%) and specificity (99.7% vs 90.8%) compared to rule-based approaches, with more accurate prevalence (3.5% vs 0.3%-1.5%). Another study reported higher prevalence (13.6% vs ~4%-5%) using BLCMs, improving case identification by integrating multiple data sources.
CONCLUSIONS: BLCMs provide a framework for validating phenotyping algorithms with incomplete gold standards, improving performance by accounting for misclassification assuming conditional independence between data sources. Although BLCMs have been applied in medico-administrative databases primarily in the absence of gold standards, their use is broader in diagnostic test evaluation across diverse settings. Validation of phenotyping algorithms with BLCMs in medico-administrative databases requires more studies, particularly with case-only gold standards.