Fast construction of interpretable whole-brain decoders

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning in Neuroscience

Background:

• Decoding mental states from functional MRI (fMRI) data is a key goal in neuroscience.
• Accurate whole-brain decoding models are desirable but often computationally intensive and lack interpretability.
• Existing statistical methods pose challenges for whole-brain analysis due to computational burden and poor interpretability.

Purpose of the Study:

• To develop a computationally efficient and interpretable method for building whole-brain neural decoders.
• To address the limitations of current statistical approaches in whole-brain functional MRI analysis.
• To facilitate wider implementation of interpretable whole-brain predictors in neuroimaging research.

Main Methods:

• Extension of the partial least squares (PLS) algorithm to create a regularized model with variable selection.
• Introduction of a "fit once, tune later" strategy for model fitting and parameter selection.
• Development of a method that scales effectively with increasing data size.

Main Results:

• The proposed method demonstrates computational efficiency and scalability with larger datasets.
• The developed whole-brain decoders are interpretable, providing insights into brain activity patterns.
• The approach yields accurate predictions of mental states from functional MRI data.

Conclusions:

• The novel method offers an interpretable and computationally efficient solution for whole-brain decoding in fMRI.
• The "fit once, tune later" approach simplifies the application of complex neural decoders.
• Public availability of the algorithm aims to promote its adoption in the broader neuroimaging community.