Alzheimer’s disease is a neurological disorder that can be diagnosed by using PET or MRI scans. Reduction of images’ resolution can be suitable to solve resource-limited tasks, such as Alzheimer’s disease prediagnosis. This paper proposes the EDoViT-Alz model, an approach based on a vision transformer that utilizes MRI scans downscaled to an 8 \(\times \) 8 resolution. This method significantly reduces computational demands while keeping other important features. EDoViT-Alz is evaluated using a comprehensive dataset that demonstrates high performance across various dementia stages. In addition to the patch embedding layer and positional encoding, the model has transformer encoders with multi-head self-attention mechanism on a feed-forward network to increase the efficacy of the approach.