Local-Aggregate Modeling for Multi-subject Neuroimage Data via Distributed Optimization

Abstract

Developing multi-subject predictive models based on whole-brain neuroimage data for each subject is a major challenge due to the spatio-temporal nature of the variables and the massive amount of data relative to the number of subjects. We propose a novel multivariate machine learning model and algorithmic strategy for multi-subject regression or classification that uses regularization to directly account for the spatio-temporal nature of the data. Our method begins by fitting multi-subject models to each location separately (similar to univariate frameworks), and then aggregates information across nearby locations through regularization. We develop an optimization strategy so that our so called, Local-Aggregate Models, can be fit in a completely distributed manner over the locations which greatly reduces computational costs. Our models achieve better predictions with more interpretable results as demonstrated through a multi-subject EEG example.