Speaker: Dr. Sandra Safo, Associate Professor, University of Minnesota

Title: Unlocking Insights: eXplainable Kernel and Deep Learning Methods for Multimodal Data Integration

Abstract: Statistical and machine learning methods for multimodal data integration have garnered significant attention in recent literature. In this talk, I will discuss the challenges and opportunities associated with integrating data from diverse sources (or views). I will introduce both supervised and unsupervised nonlinear methods for data integration, emphasizing their application in biomarker identification and data reconstruction. The first method I will present is a scalable randomized kernel approach designed to model nonlinear relationships within multimodal data while predicting clinical outcomes. This method identifies key variables or groups of variables that contribute to the relationships among different views. By leveraging random Fourier bases to approximate shift-invariant kernel functions, we construct nonlinear mappings for each data view. These mappings, in conjunction with the outcome variable, enable us to learn view-independent low-dimensional representations and identify the variables driving shared nonlinear structures. The second method combines the flexibility of deep learning with the statistical advantages of data-driven and knowledge-driven feature selection, resulting in an interpretable deep learning framework for data integration. Through simulation studies and real data applications in COVID-19 and cancer, we demonstrate that these proposed methods outperform various linear and nonlinear approaches for multimodal data integration. This is especially notable in small-sample scenarios, making them particularly valuable for biomedical research involving limited data.