Alzheimer’s disease (AD) exhibits marked inter-individual heterogeneity in progression trajectories. While machine learning enables personalized prediction, existing models often fail to capture long-term biomarker dynamics due to short observation windows, which limits the ability to capture long-term trends. In this work, an improved Transformer dynamic prediction method based on long-term follow-up to 48-month data was proposed. First, the disease progression space is constructed based on non-negative matrix factorization, and then by mapping the input data into multiple heads of the ProbSparse self-attention mechanism, the model by calculating the sparsity measure of the attention score, the key-value pairs that have a significant impact on the disease progression are dynamically screened, and the topk. This approach enhancing the modeling ability of dynamic dependencies across time steps and provides more accurate personalized prediction of AD progression trajectory predictions. The proposed method is evaluated on Alzheimer’s Disease Neuroimaging Initiative (ADNI) datasets, and the experimental results show that the proposed method is superior to the traditional methods, achieving MSE of 0.0115 ± 0.0007. This method provides a reliable decision-making tool for precision medicine for AD.