Early diagnosis of Alzheimer’s disease and related dementias (AD/ADRD) is critical but often constrained by limited access to fluid and imaging biomarkers, particularly in low-resource settings.

To develop and evaluate a predictive model for cognitive decline using survey-based data, with attention to model interpretability and fairness.

Using data from the Mexican Health and Aging Study (MHAS), a nationally representative longitudinal survey of adults aged 50 and older ( N = 4095), we developed a machine learning model to predict future cognitive scores. The model was trained on survey data from 2003 to 2012, encompassing demographic, lifestyle, and social determinants of health (SDoH) variables. A stacked ensemble approach combined five base models—Random Forest, LightGBM, XGBoost, Lasso, and K-Nearest Neighbors—with a Ridge regression meta-model.

The model achieved a root-mean-square error (RMSE) of 39.25 (95 % CI: 38.12–40.52), representing 10.2 % of the cognitive score range, on a 20 % held-out test set. Features influencing predictions, included education level, age, reading behavior, floor material, mother’s education level, social activity frequency, the interaction between the number of living children and age, and overall engagement in activities. Fairness analyses revealed model biases in underrepresented subgroups within the dataset, such as individuals with 7–9 years of education.

These findings highlight the potential of using accessible, low-cost SDoH survey data for predicting risk of cognitive decline in aging populations. They also underscore the importance of incorporating fairness metrics into predictive modeling pipelines to ensure equitable performance across diverse groups.