Machine learning classification of digital exclusion in ageing populations across five international cohorts

Abstract

Digital exclusion may limit access to increasingly digital health and public services, yet its cross-national structure and model transportability remain uncertain. We analysed harmonised cross-sectional data from 71,689 middle-aged and older adults in CHARLS, ELSA, HRS, MHAS and SHARE. Digital exclusion was defined as no reported internet or email use, a proxy for basic digital participation. Gradient Boosting discriminated concurrent exclusion with AUCs of 0.734–0.850, but logistic regression performed similarly (0.721–0.846; differences 0.001–0.013). Within SHARE, descriptive country-level exclusion ranged from 20.1% in Denmark to 79.2% in Poland. Leave-one-cohort-out validation showed partial transportability, with mean AUC attenuation of 0.026. Cross-fitted classification scores were associated with concurrent depression (pooled adjusted OR 2.23, 95% CI 1.83–2.71). These findings support population-level characterization of digital exclusion but not prospective or individual-level risk prediction without external validation.