Tensor Kernel Learning for Classification of Alzheimer’s Conditions using Multimodal Data

Early, timely, and accurate assessment of Alzheimer’s disease (AD), particularly at its earlier stage – mild cognitive impairment (MCI) –, is central to detecting, managing, and potentially treating the disease. The biological underpinnings of AD, however, is multifaceted, from genetic variations, abnormal protein accumulation, to irregular brain functions and structure. A joint analysis of these data, therefore, may offer potentially new insights about AD-related biomarkers and AD prediction. But such explorations must confront the complexity of these data: heterogeneity, multimodality and high-dimensionality. Here, to address these challenges, we propose a new machine-learning method, namely the tensor kernel learning (TKL), leveraging tensor methods and kernel learning, to enhance AD assessment by enhancing multi-modal data integration. More specifically, TKL first uses CP/PARAFAC decomposition and graph diffusion to fuse multiple kernels learned from four complementary data modalities (MRI, PET, CSF, and SNP data). We then used a supervised kernel for a kernel SVM classifier to identify potential patients. To evaluate the effectiveness of TKL, we apply it to data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI; the sample size n = 331 subjects), including cognitively normal individuals (CN), MCI subjects, and AD patients. TKL improves AD classification performance in both linear and nonlinear combinations, achieving accuracies of 91.31% for CN vs. AD, 81.45% for CN vs. MCI, and 78.27% for AD vs. MCI, compared to 85.48%, 70.89%, and 73.51% using the best single modality. Additionally, TKL reveals clearer, more structured patterns in the data, enhancing interpretation and understanding of the relationships among different modalities.