Diagnosing Alzheimer's Disease with Bi-multitask Regularized Sparse Canonical Correlation Analysis and Logistic Regression

Individuals with the Alzheimer's disease (AD) go through multiple stages from health to illness. The pathogenesis of AD remains uncertain, and there may be different biomarkers in different diagnostic groups. In the field of brain imaging genetics, it has become a significance challenge to utilize the brain genotype-phenotype correlations to probe the pathogenesis of AD. To solve these problems, a novel approach named bi-multitask regularized sparse canonical correlation analysis and logistic regression (BRSCCALR) is proposed, which can identify AD related biomarkers and classify subjects. Specifically, multitask sparse canonical correlation analysis focuses on learning genotype-phenotype associations. Yet the newly constructed multitask regularized logistic regression that prevents overfitting is responsible for identifying diagnosis-specific biomarkers. In addition, the connectivity-based penalty term is also introduced to enrich the prior information and enhance the biological significance of the method. Under the five-fold cross-validation experiment, the proposed method is compared with several state-of-the-art methods on a real brain imaging genetic dataset. The canonical correlation coefficients demonstrate that BRSCCALR method achieves outstanding performance. Finally, the learned biomarkers are applied to the classification experiment, and results show that the biomarkers are valid.