Chronic disease prevalence estimates from electronic medical records: Applying capture-recapture to clinical data

Background Prevalence estimates increasingly use electronic medical records (EMR), but the dominant, registry-based approach underestimates the relative prevalence of hard-to-diagnose and those prone to treatment delay, e.g., substance use disorders (UD) and asthma Capture-recapture (C-R) can improve estimates by using the relationship between multiple indicators of disease status to estimate missed as well as identified cases. However, most public health C-R applications have employed sporadic indicators, making it difficult to evaluate the reliability of the estimates. Methods A systematic approach to applying C-R to time series data derived from the EMR is used to estimate UD prevalence among 31,861 members of a membership based health plan who responded to a healthcare survey. Plausible assumptions about continuity of care and the relationship of diagnosis to treatment are tested using 2-period time series for diagnosis and treatment. The resulting, parsimonious 4-indicator model is used to estimate prevalence and bootstrap confidence intervals are calculated. Survey data and an independent screening study are used to evaluate the C-R estimates. Results C-R estimates ascertainment corrected prevalence of 5.8 (5.7-9.6)times the number registered medically, which agrees with intensive screening estimate Estimated prevalence per 100 is 89 for subjects reporting UD on the survey and 7-8 for others. Conclusions Repeated measures readily available with EMR yield stable and parsimonious C-R based prevalence estimates of UD. The ascertainment corrected estimates can improve planning by clarifying the true relative prevalence of more and less readily ascertained diseases.