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

Area of Science:

• Gerontology
• Epidemiology
• Cognitive Science

Background:

• Validated cognitive assessments are often infeasible in large observational studies.
• Existing dementia detection algorithms lack application in European contexts.
• Underreporting of dementia is a significant challenge in population health research.

Purpose of the Study:

• To evaluate the Langa-Weir (LW) algorithm for identifying 'probable dementia' within the European context.
• To assess the algorithm's ability to account for country-level variations in dementia prevalence and underreporting.
• To compare the LW algorithm's performance against machine learning classifiers.

Main Methods:

• Analysis of data from 56,622 older adults (60+ years) in the Survey of Health, Ageing and Retirement in Europe (SHARE, 2017).
• Comparison of the Langa-Weir algorithm with logistic regression, random forest, and XGBoost classifiers.
• Evaluation of 'probable dementia' prevalence against OECD estimates.

Main Results:

• The prevalence-specific LW algorithm reduced dementia underreporting from 61.0% to 30.4%.
• The LW algorithm outperformed tested machine learning algorithms in reducing underreporting.
• Participants classified with 'probable dementia' by LW showed similar health and cognitive performance to those with self-reported physician diagnoses.

Conclusions:

• Algorithm-based dementia detection, specifically the LW algorithm, is adaptable for cross-national cohort surveys like SHARE.
• The LW algorithm effectively reduces dementia underreporting using a minimal predictor set.
• This approach enhances the accuracy of dementia prevalence estimates in large-scale observational studies.