Performance of probable dementia classification in a European multi-country survey

Feasibility constraints limit availability of validated cognitive assessments in observational studies. Algorithm-based identification of 'probable dementia' is thus needed, but no algorithm developed so far has been applied in the European context. The present study sought to explore the usefulness of the Langa-Weir (LW) algorithm to detect 'probable dementia' while accounting for country-level variation in prevalence and potential underreporting of dementia. Data from 56 622 respondents of the Survey of Health, Ageing and Retirement in Europe (SHARE, 2017) aged 60 years and older with non-missing data were analyzed. Performance of LW was compared to a logistic regression, random forest and XGBoost classifier. Population-level 'probable dementia' prevalence was compared to estimates based on data from the Organisation for Economic Co-operation and Development. As such, application of the prevalence-specific LW algorithm, based on recall and limitations in instrumental activities of daily living, reduced underreporting from 61.0 (95% CI, 53.3-68.7%) to 30.4% (95% CI, 19.3-41.4%), outperforming tested machine learning algorithms. Performance in other domains of health and cognitive function was similar for participants classified 'probable dementia' and those self-reporting physician-diagnosis of dementia. Dementia classification algorithms can be adapted to cross-national cohort surveys such as SHARE and help reduce underreporting of dementia with a minimal predictor set.