Low-burden AI approach for cross-national early identification of cognitive impairment using real-world questionnaire response behaviours

Abstract Under-recognition of cognitive impairment remains a major global public health challenge, particularly in low- and middle-income countries. Many cases go unrecognised because conventional cognitive assessment is resource-intensive and difficult to scale across diverse populations. Here we show that easily captured indicators of reduced survey response quality, derived from how older adults answer routine psychosocial questionnaires, can identify cognitive impairment risk across countries. Using data from 45,604 older adults across five population-based ageing cohorts, we develop a cross-nationally generalisable machine-learning pipeline based on a tabular foundation model. Especially, the model jointly trained across the United States, England, India and Mexico outperforms country-specific models, achieving up to a 0.12 absolute gain in the area under the receiver operating characteristic curve (approximately 20% relative) in an external validation cohort from China. We also identify a cross-national implicit feature transfer phenomenon, in which cohorts lacking certain predictors benefit from joint training with cohorts in which these predictors are present. Decision-curve analyses indicate consistent potential public health benefit across countries by supporting population-level prioritisation of higher-risk subgroups. This low-burden approach offers an equitable and scalable strategy for generating population- and community-level evidence on cognitive impairment risk in culturally diverse and resource-constrained settings.