Thresholded Partial Least Squares: Fast Construction of Interpretable Whole-brain Decoders

Recent neuroimaging research has shown that it is possible to decode mental states and predict future consumer behavior from brain activity data (a time-series of images). However, the unique characteristics (and high dimensionality) of neuroimaging data, coupled with a need for neuroscientifically interpretable models, has largely discouraged the use of the entire brain's data as predictors. Instead, most neuroscientific research uses "regionalized" (partial-brain) data to reduce the computational burden and to improve interpretability (i.e., localizability of signal), at the cost of losing potential information. Here we propose a novel approach that can build whole-brain neural decoders (using the entire data set and capitalizing on the full correlational structure) that are both interpretable and computationally efficient. We exploit analytical properties of the partial least squares algorithm to build a regularized regression model with variable selection that boasts (in contrast to most statistical methods) a unique "fit-once-tune-later" approach where users need to fit the model only once and can choose the best tuning parameters post-hoc. We demonstrate its efficacy in a large neuroimaging dataset against off-the-shelf prediction methods and show that our new method scales exceptionally with increasing data size, yields more interpretable results, and uses less computational memory, while retaining high predictive power.