The complicated pathophysiology of Alzheimer's disease (AD), which can occasionally be inherited, is typified by the loss of synapses and neurons as well as the appearance of neurofibrillary tangles and senile plaques. For treatment or prevention to be effective, early detection is essential, especially in high-risk patients. This work offers a multi-model feature fusion method based on the attention mechanism as a novel way to classify Alzheimer's disease. The ADNI dataset was first used to test many pre-trained models, and the top three performances were chosen for additional testing. We created an attention-based feature fusion module to efficiently combine features from three different modalities. Our tests showed that merging features without the attention mechanism results in a significant decline in performance (accuracy=82%). However, implementing the attention mechanism before the fusion process significantly enhanced performance, with 99.31% accuracy in classifying Alzheimer's disease into five stages. Motivated by these outcomes, we expanded our approach to classify the disease into four and three stages, with 98.29% and 99.43% accuracies, respectively. Our results demonstrate how well the multi-model features with the attention mechanism work to improve Alzheimer's disease classification.