Movie viewing elicits rich and reliable brain state dynamics

The study investigates how naturalistic sensory stimulation remodels whole-brain state dynamics compared to unconstrained rest. Although resting-state fMRI reveals dynamic transitions between discrete brain states, their functional relevance and relation to perception and evaluation remain unclear. The authors hypothesize that movie viewing reshapes intrinsic brain dynamics into a reliable sequence of well-defined states whose timing aligns with movie features, associates with autonomic physiology, and reflects subjective engagement. By comparing resting state with movie viewing across two sessions separated by three months, they aim to disambiguate stimulus-driven from endogenous dynamics and evaluate reliability and behavioral relevance of brain state transitions inferred via a hidden Markov model (HMM).

Prior work shows the brain exhibits coordinated, dynamic patterns even at rest, with dynamic connectivity linked to behavioral traits such as processing speed and intelligence. HMM approaches model discrete, transitory brain states and have related resting-state dynamics to genetics and personality. However, resting-state variability and non-neural confounds (motion, cardiac, respiratory) complicate behavioral inference. Traditional task paradigms lack ecological validity and continuous narrative engagement. Naturalistic stimuli (movies, narratives) impose structure and drive intersubject synchronization, with hierarchical timescales of state transitions observed during movie viewing, reflecting environmental statistics. The present work builds on these findings to compare resting-state bistability with potentially richer, stimulus-aligned dynamics during movie viewing, and to link state expression to meta-analytic functional profiles and physiological indices.

Participants: 21 healthy adults (11 females, right-handed, 21–31 years, mean 27 ± 2.7). Seventeen completed two sessions; three excluded due to head motion. Main analyses included 18 participants (session A); 14 participants completed both sessions and were used for group HMM estimation enabling cross-session comparisons. Design: Two MRI sessions 3 months apart. Each session: 8 min eyes-closed rest followed by 20 min viewing of the short film “The Butterfly Circus.” Post-session questionnaire assessed boredom, enjoyment, emotional feelings, and audio quality (1–5 scales), plus control questions. Data acquisition: Siemens TIM Trio, 12-channel head coil. Gradient-echo EPI: TR = 2200 ms, 3 mm isotropic. Rest: 220 volumes; Movie: 535 volumes. Structural T1: 1 mm isotropic. Concurrent physiology: heart rate (HR) via MR-compatible amplifier (5 kHz), respiration (50 Hz), pupil diameter (PD) via Eyelink (1 kHz). Preprocessing: fMRIPrep 1.1.5: intensity non-uniformity correction, MNI normalization, tissue segmentation, slice-time and motion correction, co-registration, smoothing (6 mm), ICA-AROMA (non-aggressive). Temporal preprocessing with Nilearn 0.5.0: band-pass 0.01–0.15 Hz, regression of WM and CSF global signals. HR processed with FMRIB FastR; resampled to fMRI TR. PD cleaned (blink/bad segments removal), interpolated, downsampled to TR. Parcellation: 14 canonical brain networks (BNs) from Shirer et al. For each BN: mean signal within ROIs; first 5 volumes discarded; demeaned and scaled to unit variance per participant/session, yielding 220×14 (rest) and 535×14 (movie) matrices per session. HMM modeling: Time series concatenated across 14 participants who completed both sessions (rest A 215 vols; movie A 530; rest B 215; movie B 530), forming a 20,860×14 matrix. Variational Bayes inversion with HMM-MAR (500 training cycles). Number of states chosen as 10 via AIC and free-energy/permutation checks (tests with 6, 8, 12, 14, 25 states showed redundancy ≥12). Each state defined by a multivariate Gaussian (means and covariances over BNs). Outputs: per-participant state paths (Viterbi), fractional occupancy (FO), dwell times, and transition probability matrices. Functional decoding: Reverse/forward association using Neurosynth topic maps (16 topics including anxiety, language, task switching, inhibition, sensory domains, emotion, face perception) by correlating state spatial profiles with topic maps. Consistency analyses: Inter-subject consistency estimated via sliding window counts of participants expressing a state; inter-session consistency via Jaccard overlap of state presence vectors across sessions. State dynamics compared between rest and movie using paired t-tests on FO and dwell times. State transition differences assessed with Network-Based Statistics (NBS, PFWE < 0.05). Physiology and annotations: State-specific deviations in HR and PD computed relative to movie-wide means; tested with one-sample two-sided t-tests, FWE-corrected. Movie annotations (faces ±, scene valence ±, language, scene changepoints) were binarized and overlapped with state vectors using Szymkiewicz–Simpson index; significance via 5000 permutations (PFWE < 0.05). Behavioral linkage: Post-movie questionnaire summarized via multidimensional scaling. Inter-subject representational similarity analysis (IS-RSA) related distances in questionnaire ratings to distances in FO, transition matrices, and state paths; significance via 5000 permutations, Pearson correlations on lower-triangular distance matrices.

- Movie viewing increased inter-subject and inter-session consistency of brain state expression relative to rest, with multiple time points showing 100% across-participant alignment to specific scenes. Average Jaccard overlap across sessions: movies 0.18 ± 0.04 vs rest 0.08 ± 0.07 (p = 0.002). - Resting-state dynamics were dominated by bistable transitions between two relatively uniform states (states 5 and 9), with low alignment across participants. - Movie viewing elicited a richer repertoire: significantly higher fractional occupancy for states 1–4 and 6–8 (linked to visual, auditory, and language networks), shorter dwell times, and more diverse transitions among specific states (notably 1, 2, 3, 7, 8). Rest > movie showed increased transitions from state 9 to state 5; movie > rest showed increased transitions among states 1, 2, 3, 7, 8 (NBS PFWE < 0.05). - Functional decoding (Neurosynth) yielded distinct profiles: State 3 strongly associated with language, emotion, auditory; State 6 with emotion (positive and negative); States 4 and 7 with task switching and sensorimotor; State 10 with sensorimotor, pain, inhibition. - Physiological coupling: State 3 associated with lower HR (ΔHR = −0.42 bpm, p = 0.021). PD deviations: State 1 +45.7 a.u. (p = 0.014), State 2 +90.3 a.u. (p = 0.023), State 4 −11.4 a.u. (p = 0.001, PFWE < 0.05). PD strongly anti-correlated with scene luminance (r = −0.69, p = 10^−75). - Annotation overlaps (PFWE < 0.05): State 1 aligned with positive faces (t = 13.17) and positive scenes (t = 14.22); State 6 with negative scenes (t = 12.58) and negative faces (t = 9.31); State 3 with language (t = 11.40), positive faces (t = 13.08), negative faces (t = 7.09), and changepoints (t = 5.58); State 4 with language (t = 39.42), faces positive/negative and positive scenes; State 7 with changepoints (t = 9.70). States 8–10 showed no significant annotation overlaps. - Individual differences: IS-RSA showed significant correlations between questionnaire-based distances and FO distances (r = 0.174, p = 0.031) and transition distances (r = 0.182, p = 0.034), indicating that more similar subjective engagement/emotion corresponded to more similar movie-evoked brain state dynamics. - Despite precise timing misalignments, pairwise Jaccard dissimilarities of state paths ranged 0.72–0.88, indicating moderate commonality of sequence structure across participants.

The findings demonstrate that an engaging, naturalistic stimulus reorganizes whole-brain dynamics from resting-state bistability into a reliable sequence of well-defined functional states. These states are temporally aligned with perceptual, semantic, and narrative features, and are coupled to autonomic physiology (HR, PD), supporting their functional relevance for exteroceptive and interoceptive processing. Movie viewing enhances both across-subject and test–retest reliability of brain state dynamics, addressing limitations of unconstrained rest for linking dynamics to behavior. The richer transitions and stronger deviations from global mean activity suggest deeper attractor engagement under stimulus-driven conditions, consonant with theoretical frameworks of metastability and multistability. Associations with subjective ratings indicate that idiosyncratic aspects of state occupancy and transitions reflect the subjective experience of engagement and emotion during narrative processing.

This study shows that movie viewing elicits a rich, reliable sequence of discrete whole-brain states, with transitions aligned to narrative structure, linked to autonomic indices, and predictive of subjective engagement. Compared to resting-state bistability, naturalistic stimuli provide stronger test–retest reliability and functional interpretability of brain state dynamics. These results advocate for ecologically valid paradigms and time-resolved modeling (e.g., HMM) in cognitive and clinical neuroscience. Future work should employ faster neurophysiological modalities (M/EEG) to resolve sub-second relationships between stimulus change points and brain state transitions, explore finer-grained or adaptive parcellations to uncover hierarchical timescales, expand sample sizes and stimuli diversity, and assess clinical utility for monitoring disease-related alterations in dynamic brain states.

- HMM imposes discretely expressed states and is not a biophysical model; results depend on model assumptions and chosen state number (10 states selected via AIC and free-energy checks). - Coarse spatial scale (14 canonical networks) may have restricted detection of finer resting-state dynamics and precluded identifying hierarchically nested timescales seen with regional HMMs. - Sample size is modest; three participants excluded for motion; only one movie stimulus used, potentially limiting generalizability. - BOLD temporal resolution (TR 2.2 s) limits inference about precise temporal coupling between rapid scene changes and state transitions. - Potential physiological confounds remain, though mitigated by structured, emotionally salient stimuli and concurrent physiological measures.