Schedule of Events | Symposia

Predicting brain responses from short movies: challenges and opportunities

Poster Session D - Monday, March 31, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom

Avery Van De Water¹ (avery-vandewater@uiowa.edu), Lisa Byrge², Dan Kennedy³, Dorit Kliemann¹, James Traer¹; ¹The University of Iowa, ²University of North Florida, ³Indiana University

The human brain must flexibly integrate complex, dynamic information describable by high-dimensional features. Passive movie-viewing paradigms (moviefMRI) have been shown to reliably engage dynamic information processing across brain networks and predict individual differences. Encoding models have shown promise in revealing distributed feature organization within high-dimensional feature spaces. Combining encoding models and moviefMRI to understand dynamic brain function offers particular promise for characterizing atypical information integration in psychiatric conditions. However, the robustness of this approach in short(er) research protocols (<30 minutes) remains unclear. We applied cross-validated encoding models (ridge, banded ridge regressions) to moviefMRI data (typically developed adults, n=79; mean age 26.8; six 5-23 minute movies) using cortical parcellations derived from multimodal or resting-state data. Extracting feature spaces (visual, auditory, abstract representations e.g., socio-emotional) from the movie stimuli, we assessed model prediction accuracy, quantified feature contributions via variance decomposition, and introduced novel null models. Results showed robust signal predictions, for a sparse set of parcels within sensory and multimodal networks, driven by a sparse set of auditory and abstract features. Model performance was sensitive to hyperparameter selection. With increasing numbers of stimulus features, null models performed increasingly well, approaching standard feature model levels (r~0.3 across most of the brain), limiting visual features’ utility (the largest feature space). In summary, future work using shorter paradigms (e.g., for clinical populations) must carefully address overfitting beyond standard cross-validation. Restricting encoding models to smaller, carefully selected feature sets may yield more robust results.

Topic Area: METHODS: Neuroimaging