Data-driven disease subtypes identification based on multimodal brain images using non-negative matrix factorization

Abstract

This paper proposes the Clustering Multiblock Sparse Multivariable Analysis (CMSMA) method to identify disease subtypes in multimodal brain image analysis. The method implements a multi-layered scoring method using voxel values as input data. The first layer computes each modality score, and the next layer uses nonnegative matrix factorization to combine these scores, allowing robust data-driven subject classification and efficient dimension reduction. Experiments with real brain disease data show that the proposed method serves as a biomarker for data-driven disease classification based on brain images and that the derived scores are useful for subgroup classification of patients with mild cognitive impairment. In addition, relevant brain regions were visualized, suggesting the importance of the hippocampus and surrounding areas. The association between cluster results and prognostic classification was evaluated, and it was revealed that cluster 2 is a subgroup with a relatively poor prognosis. In addition to neuroimaging, this technique is useful for predicting treatment response and potential side effects in the medical field, providing comprehensive knowledge on disease classification, influencing medical research and practice, and contributing to improved diagnostic accuracy and treatment strategies.