Unsupervised detection of dementia based on longitudinal data from the Survey of Health, Ageing and Retirement in Europe

Abstract

AbstractBackgroundIn the Survey of Health, Ageing and Retirement in Europe (SHARE), a population‐based survey of aging in multiple countries, dementia status is self‐reported. However, more than half of dementia cases remain underdiagnosed, which may hinder the use of SHARE data to study dementia determinants and outcomes. We use unsupervised machine learning and longitudinal data from SHARE to discover clusters of patients with probable dementia.MethodsWe selected participants aged 50 or more with consecutive follow‐ups across all SHARE prospective waves. We chose variables informative of participants’ daily function and cognitive performance. We applied a Multiple Factor Analysis followed by a Hierarchical Clustering to the whole longitudinal data set. To evaluate our algorithm, we looked at its discrimination power by comparing it to the self‐reported dementia. We also checked if increased risk of transition to probable dementia was associated with non‐modifiable (age, sex) and modifiable (education, hypertension, obesity, hearing loss, diabetes, physical inactivity, smoking, drinking alcohol, social isolation, depression, air pollution) dementia risk factors. Finally, we replicated this study in the English Longitudinal Survey of Aging (ELSA).ResultsThe baseline sample consisted of 15,278 initial participants across 12 countries. The algorithm identified a higher number of probable dementia cases compared with self‐reported cases. Its discriminative power was good discriminative power across all waves (Area Under the Curve > 0.75). Sensitivity attained its highest value at wave 4 reaching 0.71 [0.66–0.77]. Specificity remained high (> 0.9) in all waves. Both non‐modifiable and modifiable dementia risk factors were associated with increased risk of transition to probable dementia. These results were replicated in ELSA.ConclusionDetection of probable dementia with longitudinal data and unsupervised clustering could be used to study dementia in SHARE and ELSA cohorts. Further validation of this method is needed to apply it in other surveys of aging.