Abstract
Alzheimer's disease, an irreversible neurodegenerative disorder, manifests through the progressive deterioration of memory and cognitive functions. While magnetic resonance imaging has become an indispensable neuroimaging modality for Alzheimer's disease diagnosis and monitoring, current diagnostic paradigms predominantly rely on single-time-point data analysis, neglecting the inherent longitudinal nature of neuroimaging applications. Therefore, in this paper, we propose a multi-task feature selection algorithm for Alzheimer's disease classification based on longitudinal imaging and hypergraphs (THM2TFS). Our methodology establishes a multi-task learning framework where feature selection at each temporal interval is treated as an individual task within each imaging modality. To address temporal dependencies, we implement group sparse regularization with two critical components: (1) a hypergraph-induced regularization term that captures high-order structural relationships among subjects through hypergraph Laplacian modeling, and (2) a fused sparse Laplacian regularization term that encodes progressive pathological changes in brain regions across time points. The selected features are subsequently integrated via multi-kernel support vector machines for final classification. We used functional magnetic resonance imaging and structural functional magnetic resonance imaging data from Alzheimer's Disease Neuroimaging Initiative at four different time points (baseline (T1), 6th month (T2), 12th month (T3), and 24th month (T4)) to evaluate our method. The experimental results show that the accuracy rates of 96.75%, 93.45, and 83.78 for the three groups of classification tasks (AD vs. NC, MCI vs. NC and AD vs. MCI) are obtained, respectively, which indicates that the proposed method can not only capture the relevant information between longitudinal image data well, but also the classification accuracy of Alzheimer's disease is improved, and it helps to identify the biomarkers associated with Alzheimer's disease.