Accuracy of diagnostic codes and algorithms used to identify connective tissue diseases in electronic health records and administrative databases: systematic review and meta-analysis

Objective

To assess the accuracy of codes and algorithms used to identify selected connective tissue diseases (CTDs) in electronic health records (EHRs) and administrative databases.

Methods

We searched MEDLINE, Embase, and CENTRAL databases for studies that validated case definitions in EHRs against a reference standard, including rheumatologist-confirmed diagnosis or clinical classification criteria. Title/abstract screening, full-text review, data extraction, and quality appraisal were independently performed in duplicate. Findings were synthesised narratively and through a bivariate random-effects meta-analysis of sensitivity and specificity.

Results

A total of 38 studies were included. Systemic lupus erythematosus (SLE) was the most frequently studied disease, followed by systemic sclerosis (SSc). Across diseases, progressively restrictive algorithms generally improved positive predictive value (PPV), although the optimal strategy varied by disease. In SLE, multiple diagnostic codes provided the most favorable overall performance, whereas adding clinical data provided limited incremental benefit. In SSc, disease-specific clinical features, particularly Raynaud’s phenomenon, improved case identification, while specialist or inpatient-based algorithms performed better for polymyositis/dermatomyositis, although evidence was limited. For SLE, meta-analysis yielded pooled sensitivity of 0.90 (95% Cis 0.70-0.97), specificity of 0.83 (95% CIs 0.54-0.95), and PPV of 0.75 (95% CIs 0.59-0.86). Evidence for Sjögren’s Disease was limited to one study, and no studies evaluated MCTD/UCTD.

Conclusions

Optimal EHR case-identification strategies vary by disease and should be tailored to disease-specific clinical and diagnostic characteristics rather than algorithm complexity alone. Future research should develop and externally validate robust, standardised algorithms in representative populations, using rigorous validation methods, and assess the added value of longitudinal EHR data, biomarkers, and computational phenotyping approaches.

Significance and innovations

• Current evidence supports disease-specific approaches to EHR-case identification: repeated diagnostic codes provided the most favorable performance for SLE, whereas disease-specific clinical features improved identification of SSc and specialist or inpatient-based definitions showed better performance for PM/DM, although evidence for the latter was limited to four studies.

• Increasing algorithm complexity did not consistently improve case-identification accuracy, highlighting the importance of selecting data elements that reflect the clinical phenotype and diagnostic characteristics of each disease.

• Important evidence gaps remain, particularly for Sjögren’s Disease and MCTD/UCTD, as well as for externally validated algorithms incorporating longitudinal clinical information and computational approaches.

• This systematic review extends previous evidence by providing the most comprehensive synthesis to date of validated EHR algorithms across selected CTDs, incorporating newer multicomponent and computational approaches and evaluating the incremental contribution of laboratory, and medication data that had been identified as important gaps in earlier reviews, although evidence for these more complex algorithms remains limited.