Comparison of Algorithms for Dementia Classification in the Survey of Health, Ageing and Retirement in Europe

Abstract

AbstractBackgroundLow dementia rates, reflecting underdiagnosis in representative cohort studies, limit statistical power of etiological and preventative research. Although several algorithms for automated classification of presence or absence of dementia have been validated in the Health and Retirement Study (HRS), no such algorithm has yet been applied to the Survey of Health, Ageing and Retirement in Europe (SHARE).MethodThe Langa‐Weir classification (LW) was adapted to readily available indicators in SHARE, including immediate and delayed recall. Adapted algorithms additionally included instrumental activities of daily living (IADL) and used cut‐offs defined by either sample‐ or population‐level distributions. Performance was compared to logistic and bayesian‐logistic regression models and a gradient boosting machine (XGBoost) with the same indicators, adjusting for age groups, gender and educational level. The bayesian‐logistic regression used priors for sociodemographic indicators and global dementia incidence. Accuracy, specificity and sensitivity were compared with a train‐test split approach in SHARE wave 7 (2017).ResultIn total, N = 72,329 participants (57% female) above age 50 had no missing data on self‐reported dementia diagnosis, immediate or delayed recall and IADLs. LW based on immediate and delayed recall with a score cutoff based on dementia population‐incidence performed best overall (Accuracy = .92, Balanced Accuracy = .75, Sensitivity = .58, Specificity = .92), and showed greatest similarities to participants with self‐reported dementia diagnosis regarding risk factors and comorbidities (i.e., gripstrength, numerical performance, verbal fluency). Results from XGBoost suggested comparable performance however with risk of overfitting.ConclusionLW adaptations outperformed regression models regarding sensitivity. Comparisons of risk factor and comorbidity distributions suggest meaningful differences in comorbidities and risk factors in participants classified with and without dementia. With a lack of proxy assessments in SHARE, a suspected healthy volunteer bias and the absence of standardized cognitive assessments, probable dementia detection in SHARE necessarily comes with less confidence compared to algorithms tested in HRS. Nonetheless, performance of LW adaptations in SHARE is in line with previous validation studies in HRS. Future research should validate the algorithms through more extensive cognitive assessments once available.