Brain age prediction in a multiethnic Asian population: A comparison of machine learning algorithms and their application for early-stage cognitive impairment diagnosis

Background Neuroimaging-derived brain age is a promising biomarker of early neurodegeneration, but methodological variation in machine learning (ML) algorithms and input features as well as scarce evidence from various ethnic populations limit clinical translation. Objective To identify an accurate and interpretable machine learning–based brain age model for a multiethnic Asian population and examine its utility as a biomarker of early cognitive decline Methods Nine brain age prediction models were developed using 406 cognitively normal individuals (45–86 years) from two population-based studies using structural MRI features. Prediction performance was evaluated using mean absolute error (MAE) and Pearson's correlation coefficient (R 2 ). Feature importance was assessed using the SHapley Additive exPlanations (SHAP) analysis based on best performing model. The model was applied to an independent cohort with no cognitive impairment (NCI), mild and moderate cognitive impairment no dementia (CIND), and dementia. Differences in BrainAGE across cognitive groups were examined using an ANOVA test. Results The chosen ensemble model, comprised of linear regression, lasso and SVR, was trained on 17 features (11 subcortical volumes and 6 lobe-level cortical thickness measures) and achieved an overall bias-corrected MAE and R 2 of 4.04 years and 0.59 respectively. Feature importance analysis found thalamic, lateral ventricle, accumbens area and gray matter volume as important features for brain age prediction. Conclusions An interpretable ensemble ML model using structural MRI provides a robust BrainAGE biomarker capable of detecting early cognitive decline in multiethnic Asian populations.