Alzheimer’s disease (AD) is a persistent neurological disorder with almost no current cure, although available medications can mitigate its progression. The timely identification and staging of AD is pivotal in impeding and managing its advancement. This study aims to develop a comprehensive framework for the early detection of AD and the classification of medical images across four AD stages. We have proposed a weighted ensemble deep transfer learning framework using two pretrained CNN architectures, namely ResNet152V2 and DenseNet201. Additionally, gradient-weighted class activation mapping (Grad-CAM) is utilized to enhance interpretability by capturing gradients associated with AD in brain MRI images, propagating them to the final convolutional layer. This technique narrows the interpretability gap in deep learning models, enhancing accessibility and understanding of their decision-making processes, particularly in the context of AD diagnosis. Two different MRI datasets have been used to train and evaluate the performances of the proposed framework focusing on generalization capability with diverse data. The performances of the framework are analyzed meticulously under different case studies. The experimental results demonstrate that the weighted ensemble architecture exhibits superior performance characteristics over base models showing accuracy of 98%, which is compatible with related state of the art algorithms.