Brain age prediction is a neuroimaging framework for estimating biological brain aging and identifying deviations from expected aging trajectories. However, the value of brain age models depends not only on predictive accuracy, but also on interpretability, longitudinal reliability, and clinical relevance. This thesis investigates explainable machine learning approaches for brain age prediction using FreeSurfer-derived structural MRI features from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) dataset. Five regression models were developed and compared: Random Forest, Support Vector Regression, Elastic Net, Ridge Regression, and Gaussian Process Regression. Model performance was evaluated using Mean Absolute Error, Pearson correlation, and Brain Age Gap dispersion. In addition, the study examined hyperparameter sensitivity, explainability using SHAP values and complementary feature-importance methods, short-term and long-term longitudinal stability, and normative modelling in cognitively normal, Mild Cognitive Impairment, and Dementia subjects. The evaluated models achieved comparable predictive performance, with test-set errors below five years. Random Forest achieved the lowest prediction error, Elastic Net produced the strongest correlation with chronological age, and SVR showed the lowest Brain Age Gap dispersion. Explainability analyses identified biologically plausible predictors, including white matter hypointensities, ventricular measures, hippocampal and thalamic structures, and temporal and parietal cortical regions. SHAP feature rankings were highly stable across random seeds. Longitudinal analyses showed strong short-term reliability, although individual long-term aging trajectories were more difficult to capture. Normative modelling revealed progressively elevated Brain Age Gap values in Mild Cognitive Impairment and Dementia subjects, with group differences remaining significant after adjustment for age and sex. The findings of this thesis support the importance of evaluating brain age models beyond cross-sectional accuracy alone by incorporating explainability, longitudinal stability, and normative clinical analysis.