<p>Alzheimer’s disease (AD) is a progressive neurodegenerative disorder that affects millions worldwide, with early detection remaining a critical challenge in clinical neuroscience. While magnetic resonance imaging (MRI) offers a non-invasive avenue for visualising brain changes, existing deep learning models often struggle to capture both fine-grained regional patterns and broader global context, limiting their diagnostic precision, especially in early cognitive impairment stages. To address these limitations, we propose DLASRT, a novel Dual-Locality Attention-Based Skipped-Residual Transformer designed specifically for Alzheimer’s disease stage prediction from MRI scans. Our model integrates a multi-kernel local attention block with a whole-brain global attention mechanism to effectively learn spatial biomarkers at varying scales. Furthermore, we incorporate an absolute positional embedding strategy to encode brain-region awareness and introduce a Skipped-Layer Gated Unit (SLGU) that enhances memory flow through adaptive information skipping and retention. This architecture is further refined by a transformer-based cognitive refinement module and a lightweight classification head. We evaluate DLASRT on publicly available MRI datasets, demonstrating consistent improvements over existing deep learning models in distinguishing between multiple stages of cognitive decline. The model achieves class-wise F1-scores of 96.42% for CN, 96.18% for SMC, 96.30% for EMCI, 95.20% for LMCI, and 96.89% for AD, indicating balanced performance across both early and late-stage categories. We also validated the DLASRT model’s generalizability through cross-dataset testing on harmonized ADNI and OASIS cohorts, where diagnostic categories were aligned for consistent evaluation. In continuation of this work, we further utilize the feature representations extracted by DLASRT to support downstream analysis in a subsequent study, MACSDE, where we explore the estimation of structural patterns relevant to Alzheimer’s progression.</p>