Disagreement concerning atopic dermatitis subtypes between an English prospective cohort (ALSPAC) and linked electronic health records

Abstract

Abstract Background Subtypes of atopic dermatitis (AD) have been derived from the Avon Longitudinal Study of Parents and Children (ALSPAC) based on the presence and severity of symptoms reported in questionnaires (severe–frequent, moderate–frequent, moderate–declining, mild–intermittent, unaffected–rare). Good agreement between ALSPAC and linked electronic health records (EHRs) would increase trust in the clinical validity of these subtypes and allow inference of subtypes from EHRs alone, which would enable their study in large primary care databases. Objectives Firstly, to explore whether the presence and number of AD records in EHRs agree with AD symptom and severity reports from ALSPAC. Secondly, to explore whether EHRs agree with ALSPAC-derived AD subtypes. Thirdly, to construct models to classify ALSPAC-derived AD subtypes using EHRs. Methods We used data from the ALSPAC prospective cohort study from 11 timepoints until age 14 years (1991–2008), linked to local general practice EHRs. We assessed how far ALSPAC questionnaire responses and derived subtypes agreed with AD as established in EHRs using different AD definitions (e.g. diagnosis and/or prescription) and other AD-related records. We classified AD subtypes using EHRs, fitting multinomial logistic regression models, tuning hyperparameters and evaluating performance in the testing set [receiver operating characteristic (ROC) area under the curve (AUC), accuracy, sensitivity and specificity]. Results Overall, 8828 individuals out of a total 13 898 had been assigned an AD subtype and also had linked EHRs. The number of AD-related codes in EHRs generally increased with the severity of the AD subtype. However, not all patients with the severe–frequent subtype had AD in EHRs, and many with the unaffected–rare subtype did have AD in EHRs. When predicting the ALSPAC AD subtype using EHRs, the best tuned model had an ROC AUC of 0.65, a sensitivity of 0.29 and a specificity of 0.83 (both macro-averaged). When different sets of predictors were used, individuals with missing EHR coverage were excluded, and subtypes were combined, sensitivity was not considerably improved. Conclusions ALSPAC and EHRs disagreed not only on AD subtypes, but also on whether children had AD or not. Researchers should be aware that individuals considered to have AD in one source may not be considered to have AD in another.