Finding the (many) missing cases of chronic disease using linkage: an illustration from arthritis research

Background: Diagnoses are identified in routinely collected data (e.g. primary care EHR databanks) by applying code lists. Although validation studies allow the discovery and exclusion of false positives, it is near impossible to find false negatives – patients who have the disease but are missing a code. The digitisation of secondary care EHRs and the ability to process free text using natural language processing opens the opportunity for linkage and an improved understanding of and correction for misclassification.

Objectives: The aim of this study was to examine the prevalence of psoriatic arthritis (PsA) by 1) using primary care records only, 2) quantifying false positives AND false negatives in a sub-population with linked hospital outpatient letters, and 3) recalculating the prevalence for the whole population using the resultant sensitivity and specificity.

Methods: We used primary care EHR data from adults in Salford, Greater Manchester, from Jan2014 to Dec2019. Diagnoses from rheumatologists were available as an outpatient letter for a subset. MedCAT software identified and mapped diagnoses from the letters to SNOMED CT.
The annual prevalence was first calculated using primary care data alone. In the sub-population with linked primary and secondary care diagnoses, we identified the number of false positives and false negatives, considering the hospital diagnosis as ‘gold standard’. Using Bayesian modelling, we incorporated the sensitivity and specificity of primary care PsA diagnosis and recalculated the annual prevalence adjusted for misclassification.

Results: 245 cases of PsA were identified among 188,286 adults registered with primary care, a prevalence of 0.04% [95% CI: 0.03%-0.05%] in 2014 and 0.13% [0.12%-0.15%] in 2019. Among a subgroup of 7,532 primary care patients with linkage to secondary care, 192 cases were confirmed as true PsA, while 10 cases were deemed false positives based on their secondary care records. The primary care code list failed to identify 234 patients diagnosed with PsA in hospital settings, indicating a 48.85% underestimation using primary care codes alone. The primary care code list showed a sensitivity of 45.1% and specificity of 99.85%, respectively. After adjusting for sensitivity and specificity, the estimated true prevalence of PsA was 0.09% [0.07%-0.11%] in 2014 to 0.30% [0.26%-0.34%] in 2019.

Conclusions: The linkage of primary care data with text-mined hospital outpatient letters allows the identification of false positive and false negative cases and the correction of wider prevalence estimates. In our example of PsA, whilst most cases identified from primary care had ‘true’ disease, we found almost as many cases that weren’t picked up from the primary care code list.