Bipolar disorder in oncology: Considerations for the psychiatrist.

Naturalistic stimuli—such as movies or spoken narratives—have become increasingly popular in functional MRI (fMRI) research because they offer tools to investigate brain function under ecologically valid conditions (; ; ). Unlike traditional, highly controlled experimental paradigms, naturalistic stimuli capture the richness and temporal continuity of real-world experiences, thereby engaging a wide range of perceptual, cognitive, and affective processes simultaneously. This shift has opened new avenues for studying how the brain integrates complex, dynamic information over time. A seminal line of work demonstrated that when multiple individuals watch the same movie or listen to the same story, their brain activity exhibits significant inter-subject correlations (ISC), revealing shared neural responses to naturalistic events (). Such data-driven analyses have been instrumental in identifying brain regions involved in high-level perception, attention, and social cognition (). However, while these approaches effectively capture synchronized patterns of activity across individuals, they do not directly reveal which specific features of the stimulus drive these neural responses. To address this, researchers have increasingly focused on extracting quantitative features from naturalistic stimuli and mapping their temporal dynamics onto brain activity measured by fMRI (; ; ; ; ). A wide variety of stimulus-derived features can be used for such mapping. These range from low-level sensory features—such as visual luminance, contrast, and auditory pitch—to higher-level variables derived from physiological signals (e.g., pupil size) or subjective behavioral ratings (e.g., perceived emotional valence or theory of mind, ToM). In recent years, feature extraction has also benefited from advances in deep neural networks, where model-based representations capture complex visual, auditory, or semantic content (; ; ). These features can be analyzed individually using general linear models (GLMs) or jointly in encoding-model frameworks to predict regional brain activity.