Independent-Component-Based Encoding Models of Brain Activity During Story Comprehension

Researchers at MIT, led by Kamya Hari, have introduced a novel independent component (IC)-based encoding framework to better understand how the brain processes stories. Traditional voxelwise encoding models struggle with noise and variability across subjects. The new method decomposes continuous fMRI data from naturalistic story listening into independent components using one subset of data, then trains encoding models on separate data to predict IC time series from large language model (LLM) representations of linguistic input.

Across subjects, a subset of ICs showed consistently high predictivity, corresponding to cognitive networks like auditory and language regions. Auditory component time series strongly correlated with acoustic features, demonstrating interpretability. Noise components identified by ICA-AROMA had poor predictive performance, confirming the framework's ability to isolate genuine neural signals. This approach enables analyses at the functional network level, accommodating individual variability and providing interpretable, comparable results across subjects.