(AH) Algorithm to Identify Patients with Hypereosinophilic Syndrome (HES) in Electronic Health Record Databases

Publication Number: 1032

Background: Diagnosing patients with HES is difficult due to disease rarity, varied presentation and the need to exclude related conditions. Further, there is often a lack of administrative codes recorded in electronic health record (EHR) databases making it difficult to identify HES patients.

Objectives: Develop an algorithm using data from the UK Clinical Practice Research Datalink (CPRD) to identify patients with HES which could be used in other EHR databases, to help detect missed HES cases.

Methods: This study used data from CPRD Aurum database and linked Hospital Episode Statistics database (Jan 2012–June 2019). Patients with HES had a diagnosis code for HES during the study (index) and were matched (age, sex and index date) 1:29 with a non-HES cohort, who had no HES code at any time and ≥1 blood eosinophil test code (index) during the study. Algorithm development was in three stages: the identification and categorisation of pre-defined variables differing between HES cases and controls; model fitting using Firth logistic regression, with selection of five models based on measures of Akaike’s information criterion (AIC), area under the receiver operating characteristics curve, and sensitivity at 50% predicted probability threshold; internal validation using Leave-One-Out Cross Validation.

Results: 88 HES and 2552 matched cases were included. From 270 models run, the key predictive variables were: potential HES treatments (6 months post-index); asthma (ever); white blood cell conditions (ever); blood eosinophil count ≥1500 cells/µL (24 months pre-index). At a predicted 80% probability threshold, the final model had a sensitivity of 69% (95% confidence interval: 59,79), >99% specificity, positive predictive value of 2%, negative predictive value of >99% and one of the top AIC measures. Of note, specificity (and thus PPV and NPV) varied marginally and were not informative in picking the ‘best’ prediction model due to the rarity of HES (1.15/100,000 persons).

Conclusions: By combining medical codes, prescribed treatment data and laboratory results, the algorithm developed can be used to identify patients with HES from the CPRD Aurum database; validation in other EHR databases is needed.

Funding: GSK [213879]