Cross-modal brain decoding: using fMRI to decode video stimuli from integrated sensory streams

Motivation: Understanding how the human brain decodes complex video stimuli can improve our insights into multisensory integration and advance brain-computer interfaces. Goal(s): This study aims to enhance video decoding from fMRI data by integrating visual, textual, and audio streams, hypothesizing that each stream independently contributes to brain representation. Approach: Using the BOLDMoments dataset of video-fMRI pairs, we developed cross-subject decoding models able to retrieved seen videos from fMRI activity, estimating multimodal embeddings from pre-trained models (XCLIP, CLAP) directly from preprocessed fMRI data. Results: Multimodal integration (Video+Text+Audio) achieved the highest retrieval and identification accuracy, underscoring the advantage of combined sensory information for decoding. Impact: This work opens pathways for more accurate brain decoding in multisensory contexts, potentially advancing brain-computer interfaces and aiding clinical applications in sensory processing disorders.