Alzheimer’s disease (AD) represents a significant global health challenge, affecting millions of individuals worldwide through progressive cognitive decline and behavioral changes. The burden extends beyond patients to caregivers and healthcare systems. While traditional diagnostic methods pose financial obstacles, emerging non-imaging techniques show promise. Machine learning has emerged as a transformative approach for enhancing both diagnosis and management. This study aims to develop a robust multi-class classification model using random forest (RF) and extreme gradient boosting algorithms on non-imaging data from the Australian AD Neuroimaging Initiative, with emphasis on the Australian Imaging, Biomarkers, and Lifestyle Study of Aging. Extensive data analysis was conducted, including feature importance and selection, to improve interpretability and classification accuracy. Synthetic oversampling was applied to address class imbalance. The findings indicate the superiority of the tuned RF model, achieving 90% in accuracy, precision, recall, and F1 scores. In addition, cost-effective diagnostic variables were explored, with neuropsychology assessment variables demonstrating exceptional accuracy (90%). This research contributes to early AD detection, personalized treatment, and optimized resource allocation.