Multi-Kernel Generalized Additive Models: A Predictive Framework for Multimodal Imaging Data

• Activity Number: 170
• Type: Topic Contributed
• Date/Time: Monday, August 10, 2015 : 10:30 AM to 12:20 PM
• Sponsor: ENAR
• Author(s): Wen-Yu Hua and Philip Reiss* and David Lawrence Miller
• Companies: New York University School of Medicine and New York University School of Medicine and University of St. Andrews
• Keywords: multiple kernel learning ; generalized additive models ; multimodal imaging ; principal coordinate regression ; multidimensional scaling ; kernel ridge regression

Abstract

Increasingly, large-scale studies collect brain images of multiple modalities from each participant, often longitudinally, and a key desideratum is to use all these images to predict clinically relevant responses. This talk will show how, by combining the well-known distance/kernel duality with current software for generalized additive models, essentially arbitrary distances among images can be used to construct regression models that incorporate multiple image types, non-image covariates and random effects. This framework can be viewed as a flexible new approach to multiple kernel learning.