Longitudinal Alzheimer’s Disease Progression Modelling via Hybrid Vision Transformers and Recurrent Neural Networks With Cross‐Modal Feature Fusion

ABSTRACT Modelling the evolution of Alzheimer's disease (AD) requires a thorough spatiotemporal study of longitudinal neuroimaging data. We propose in this paper a novel deep learning framework that uses a parallel combination of Recurrent Neural Networks (RNNs) and Vision Transformers (ViT) to extract temporal disease dynamics and spatial structural changes from serial MRI data. While the RNN evaluates sequential dependencies across timepoints, the ViT branch uses self‐attention to derive hierarchical brain‐region characteristics. A stacked auto‐encoder (SAE) fuses these features into a compact representation, enhancing discriminative power while reducing redundancy. Fully connected layers are given the fused features in order to predict progression and classify AD (CN/MCI/AD). We used the ADNI dataset to test our proposed methodology. In terms of disease stage differentiation, our approach reaches state‐of‐the‐art accuracy of 92.3%. Compared to CNN or RNN‐only models, it considerably improves the prediction of the early conversion of MCI to AD (AUC = 0.94). When processing heterogeneous neuroimaging data, the SAE‐based fusion outperforms attention methods. With potential uses in customised treatment planning, this hybrid approach provides a clinically interpretable tool for longitudinal AD.