Can medication mentions in electronic health record databases be used for researching medication use in older people with dementia? Comparing the output of the CRIS natural language processing app for medicines to GP prescribing.

BackgroundThe efficacy of long-term preventative medication in people who have dementia and other comorbidities is unknown. Electronic health record-based observational studies may play a role in evaluating medicines, and SLaM-CRIS is one such resource. Medication in SLaM-CRIS is extracted from structured and unstructured fields using a natural language processing app. We aimed to compare the results from the medication app with GP prescribing, using an innovative data link with primary care (Lambeth DataNet, LDN) that covers around one-quarter of the SLaM-CRIS population. MethodsA cohort was created of people with both LDN record and SLaM-CRIS record who had received a diagnosis of dementia in either record, in which ascertainment of medication could be compared. Ten classes of medication commonly taken long-term for preventative purposes were studied (aspirin, statins, ACE inhibitors, anticoagulants, beta-blockers, bisphosphonates, calcium with vitamin D, non-aspirin antiplatelets, proton pump inhibitors and antidepressants), plus medications taken for dementia itself. Mentions of these medication classes in SLaM-CRIS from around the time of dementia diagnosis were extracted using a natural language processing application. Prescription issue for the same medications was extracted from LDN in the year of dementia diagnosis and compared with that from SLaM-CRIS on a cohort and individual level. ResultsOur sample included 4410 with documentation of dementia in either SLaM-CRIS or LDN. Estimation of the prevalence of the use of each medication in CRIS was compared to LDN, and was within +/-3 per 100, except for calcium with vitamin D supplement, where SLaM-CRIS prevalence was 13 people per 100 lower than LDN (8.6% vs 21.2%). Medication ascertainment of all groups except calcium with vitamin D supplement showed good agreement (kappa above 0.7), and very good agreement for antidepressants and dementia drugs (kappa above 0.8). Sensitivity was highest for dementia drugs and antidepressants (above 90%), 85% for statins and 75% for aspirin. Restricting to those with a memory clinic referral did not change the levels of agreement. DiscussionRoutinely collected data cannot provide a gold-standard measure of what medications are truly taken by patients with dementia, but several sources can provide a proxy measure. This analysis supports the use of the natural language processing application for medication in SLaM-CRIS to extract medication mentions of relevance to people with dementia, as compared to prescribing from the GP at around the same time. However, some medications show low sensitivity, possibly due to low recording or inconsistency in the text used to record, and enhancement will be needed before studying these medications.