Exploring the Neural Processes behind Narrative Engagement: An EEG Study

The study investigates how engagement with narratives relates to brain activity when narrative structure (the dramatic arc) is explicitly considered. Prior work often treated narratives holistically and focused on cognitive or affective correlates without decomposing narratives into structural phases. The authors ask: (1) how engagement fluctuates across the six phases of a dramatic arc; (2) whether self-reported engagement can be predicted from neural activity; and (3) whether distinct phases of the arc elicit idiosyncratic brain responses and distinct engagement–brain relationships. The motivation lies in theories that narrative arcs build and release tension to drive engagement and that shared neural responses reflect stimulus-driven processing and engagement. Addressing these questions advances cognitive narratology and neuroscience by linking narrative structure, engagement, and neural dynamics.

The paper reviews studies on brain responses to narratives, including event boundary processing, narrative comprehension, perspective taking, and the relationship between engagement and neural synchrony. Prior findings show: functional connectivity varies with narrative transportation; specific neural patterns during movie viewing reflect self-reported engagement; graph-theoretic features relate to attention and emotional moments; and intersubject correlation (ISC) captures engagement as emotionally engaging content increases brain synchronization across individuals. However, prior work did not examine engagement–brain relationships in the context of narrative structure phases. The authors position their work to fill this gap by combining narrative arc segmentation with EEG-based measures (dISC, dFC, graph features) and engagement ratings.

Stimulus and design: An 8 min 27 s excerpt from the movie Pride and Prejudice and Zombies, chosen to instantiate six phases of a modern Freytag dramatic arc (exposition, rising action, crisis, climax, falling action, denouement; inciting incident merged with rising action), was presented. Groups and tasks:

• EEG and recall: 32 right-handed participants (13 females; mean age 26.84, SD 4.33, range 20–37) viewed the excerpt in VR (HTC headset) while EEG was recorded from 32 channels (10–20 system; Brain Products; 1000 Hz; impedances <25 kΩ). About 50 minutes post-viewing, participants provided free verbal recall (audio recorded).
• Self-report: A separate group of 20 participants (9 females; mean age 24.80, SD 2.26, range 22–30) continuously rated engagement via a 1–9 slider during viewing (PsychoPy v3.0). Ratings were resampled to 200 Hz, aligned to stimulus time, and z-normalized.
• Event segmentation: A third group of 19 participants (8 females; mean age 26.10, SD 2.74, range 23–32) attended a workshop on narrative arc phases and timestamped the start/end of each phase during viewing. Statistical aggregation (Silva et al., 2019; 3 s coincidence window; 1000 shuffles; p<0.05) determined consensus phase boundaries. Table 1 reports phase starts/durations. Ethics: Approved by Aalborg University Technical Faculty of IT and Design ethics committee; informed consent obtained; compensated. EEG preprocessing: Third-order Butterworth bandpass 1–40 Hz; noisy channels detected by mean ±3 SD and visual inspection removed (avg rejected channels 0.93±0.60). One participant excluded (>4 bad channels). ICA (second-order blind identification) used to remove ocular and other artifacts based on spectra, time courses, and topographies (avg rejected components 4.62±0.86). Bad channels interpolated via spherical spline; re-referenced to average. Current source density (CSD) transform (m=4) applied to mitigate volume conduction. Data z-normalized across time and downsampled to 200 Hz. Event segmentation analysis: Consensus phase onsets computed via coincidence analysis and permutation testing; start times and durations obtained. Engagement synchronization: Pairwise correlations among raters’ continuous engagement ratings computed; Fisher z-averaged; FDR correction across pairs to assess significance; group-average engagement time course derived. Dynamic intersubject correlation (dISC): Using a tapered sliding window (15 samples at 200 Hz; ~70 ms; step 5 ms), Fisher z-transformed Pearson correlations computed between homologous channels for every participant pair to obtain time-resolved ISC per channel. Averaged dISC per channel correlated (Pearson r) with group-averaged engagement. Significance assessed via phase-randomized permutation (1000 surrogates), one-tailed p=(1+ #null r > empirical r)/(1+Nperm), with FDR correction. EEG amplitude prediction: Sliding windows (as above) used to compute mean EEG amplitude features. Support vector regression (SVR; nonlinear) with leave-one-subject-out (LOO) cross-validation trained on all-but-one participants’ features to predict group-averaged engagement; tested on held-out participant. Performance metric: Fisher z-transformed correlation between observed and predicted engagement; permutation testing with phase-randomized engagement (1000) for null distribution. Feature selection variant included channels significantly correlating with engagement (one-sample t-test p<0.01) prior to SVR. Dynamic functional connectivity (dFC): Within-subject 32×32 Fisher z-transformed Pearson correlation matrices computed per window (496 unique edges). LOO-SVR predicted engagement from time series of selected edges (feature selection by one-sample t-test p<0.01 within fold). Significance via permutation as above. Regional analysis: Edges categorized into frontal (F), central (C), parietal (P), temporal (T), occipital (O) networks; proportion of predictive edges computed and tested (one-tailed nonparametric; FDR p<0.001) vs chance. Graph features: From each window’s weighted connectivity matrix, node degree (ND), clustering coefficient (CC), and betweenness centrality (BC) computed per electrode (FastFC toolbox). LOO-SVR used each feature type separately to predict engagement; permutation tested. Channels consistently selected as predictors assessed via proportion vs chance (FDR corrected). Frequency-resolved dISC (dISC-bands): For each window and pair, magnitude-squared coherence between homologous channels computed to yield spectral ISC (Maffei, 2020). Averaged across pairs to obtain per-channel spectra, then band-limited to δ (1–4 Hz), θ (4–7 Hz), α (8–13 Hz), β (14–30 Hz). LOO-SVR and permutation testing assessed predictive power of each band’s dISC vs engagement; significant channels reported with FDR-corrected p-values. Recall analysis: Audio recalls transcribed; filler/irrelevant utterances removed. Sentences assigned to phases based on event timing from segmentation. Latent semantic analysis (LSA) computed semantic similarity across participants’ recalls within each phase. Documents tokenized; common words extracted; singular value decomposition with 20 components (chosen via LDA perplexity across 5/10/15/20). Pairwise cosine distances yielded between-subject similarity matrices; phase-wise means and SDs compared via ANOVA and Dunn–Sidak post hoc. Code/data: Data on Zenodo; analysis code on GitHub; computational specs reported.

• Event segmentation: Consensus phase boundaries identified (Table 1): Exposition start 0 s (duration 151 s; 100% agreement), Rising 151 s (149 s; 68.42%), Crisis 300 s (45 s; 52.63%), Climax 345 s (115 s; 63.16%), Falling 460 s (22 s; 89.47%), Denouement 482 s (24 s; 94.74%).
• Engagement ratings: Raters’ engagement time courses were significantly synchronized (mean pairwise Pearson r=0.40±0.20; 89.47% of pairs significant after FDR, p<0.05). Group-averaged engagement followed the dramatic arc, peaking at climax and declining thereafter.
• Recall: Falling action+denouement recalls had highest similarity (M=0.97, SD=0.07) and significantly lower SD than other phases (ANOVA F(4,45)=2.53, p=0.015; Dunn–Sidak post hoc p<0.05). Word counts differed across phases (ANOVA F(4,155)=15.11, p<0.001), with exposition longer than others.
• dISC vs engagement (whole narrative): Significant channel-wise correlations: FC5 r=0.11, p=0.023 (positive), FC1 r=-0.11, p=0.034 (negative), CP2 r=-0.10, p=0.019 (negative), FDR-corrected. Indicates cross-subject neural synchrony in fronto-central regions tracks engagement (in opposite directions for some channels), aligning with the dramatic arc.
• EEG amplitude prediction: Not significant (p=0.521, r=-0.015, MSE=1.064, R²=-0.064). Even with feature selection: p=0.742, r=-0.005, MSE=1.032, R²=-0.032.
• dFC prediction (whole narrative): Significant (p=0.029, r=0.049, MSE=1.095, R²=-0.095). Predictive edges concentrated within/between central regions and connections to other regions (except occipital), with prominent central involvement (FDR p<0.001 for several within/between-region proportions).
• Graph features (whole narrative): Betweenness centrality (BC) significantly predictive (p=0.035, r=0.012, MSE=1.101, R²=-0.101). ND and CC not significant (ND: p=0.423, r=0.009; CC: p=0.831, r=-0.010). Predictive BC channels: CP2 (positive), CP1 and F7 (negative), each exceeding chance selection frequency (CP2 proportion 0.066; F7 0.433; CP1 0.033; all FDR p<0.001).
• dISC frequency bands: δ (1–4 Hz), θ (4–7 Hz), and β (14–30 Hz) bands significantly predicted engagement; α (8–13 Hz) did not. Significant channels and correlations (FDR-corrected): • δ: FT7 r=0.17, p=0.022; P7 r=0.24, p=0.039; P4 r=0.20, p=0.029; FC2 r=0.27, p=0.004; F8 r=0.16, p=0.021. • θ: FT7 r=0.22, p=0.036; P4 r=0.27, p=0.015; C4 r=0.16, p=0.043; FC2 r=0.25, p=0.038; F8 r=0.18, p=0.011. • β: P4 r=0.25, p=0.024; FC2 r=0.31, p=0.027. All positive, indicating greater cross-subject spectral synchrony in more engaging moments.
• Phase-specific predictions: • dISC significantly predicted engagement in rising (CP5 r=0.065, p=0.015; FP2 r=0.074, p=0.032), crisis (Pz r=-0.116, p=0.033; Cz r=-0.113, p=0.011; FP2 r=0.135, p=0.014), and climax (FT8 r=-0.098, p=0.010; FC6 r=-0.216, p=0.011). Not significant for exposition, falling, denouement (representative nonsignificant channels reported). • dFC significantly predicted falling (p=0.028, r=-0.197, MSE=1.798, R²=-0.798) and denouement (p=0.009, r=-0.202, MSE=1.812, R²=-0.812); not significant for exposition, rising, crisis, climax. • BC significantly predicted falling (p=0.051, r=0.045, MSE=1.586, R²=-0.586) and denouement (p=0.041, r=0.097, MSE=1.387, R²=-0.387); not significant for exposition, rising, crisis, climax. In falling, F7 (negative; proportion 0.033, p=0.029) was predictive. In denouement, C4 (negative; proportion 1.0, p<0.018), P3 (negative; 0.066, p=0.018), O2 (positive; 0.030, p=0.031), T8 (positive; 0.200, p<0.001) were predictive.
• Overall pattern: dISC captured higher-engagement phases (rising, crisis, climax) often with lower shared understanding (higher variability in recall), whereas dFC and BC captured lower-engagement, higher-shared-understanding phases (falling, denouement). Frontal and central regions and their inter-regional connections played key predictive roles across analyses.

Findings demonstrate that self-reported engagement synchronizes across viewers and mirrors the dramatic arc’s tension trajectory, peaking at climax and decreasing thereafter. Neural measures sensitive to stimulus-locking and network dynamics predicted these engagement fluctuations. Channel-wise dISC in fronto-central regions tracked engagement, with both positive and negative associations across electrodes, indicating that shared neural responses vary with narrative-driven engagement. Extending dISC into frequency bands revealed that δ, θ, and β synchrony increased during more engaging moments, aligning with literature linking low-frequency ISC to attentional and affective processing during naturalistic viewing. Dynamic functional connectivity and graph betweenness centrality also predicted engagement, highlighting the central and frontal regions as hubs and connectors whose interactions covary with engagement. Importantly, decomposing the narrative into phases revealed complementary sensitivities: dISC better predicted engagement in phases eliciting high engagement and lower shared understanding (rising, crisis, climax), while dFC and BC predicted engagement in phases with lower engagement and higher shared understanding (falling, denouement). This phase specificity suggests that different neural metrics capture distinct cognitive processes engaged by narrative structure, such as integration, attention, and the stability of situation models. The convergence across dISC, dFC, and BC indicates that engagement with narrative structure is reflected in shared neural responses, network co-activations, and hub dynamics, particularly in frontal and central regions.

This study is the first to characterize neural correlates of engagement across the phases of a narrative dramatic arc. It shows that group-averaged engagement aligns with arc structure and can be predicted from EEG-based features: dISC (including δ, θ, β bands), dynamic functional connectivity, and betweenness centrality. Phase-wise analyses reveal that shared neural synchrony (dISC) captures highly engaging phases (rising, crisis, climax), while connectivity and hub metrics (dFC, BC) capture less engaging phases (falling, denouement), with frontal and central regions playing prominent roles. These findings advance cognitive narratology and neuroscience by linking narrative structure to neural dynamics and perceived engagement. Future work should: employ complete narratives to probe expectancy/closure; test additional narratives and genres for generalizability; compare alternative ISC methods (e.g., CCA-based), and intersubject functional connectivity; consider additional graph metrics of centrality and small-worldness; and design stimuli with balanced phase durations and explore linear vs nonlinear/interactive narratives.

• ISC computation used a relatively simple correlation-based dISC rather than CCA-based ISC; future work could compare methods.
• Functional connectivity was computed within-subject across electrodes; intersubject functional connectivity could offer robustness to endogenous patterns and should be examined.
• Only three graph features (ND, CC, BC) were evaluated; other centrality and small-world measures may provide additional insights.
• Narrative phases had unequal durations (short falling/denouement; long exposition), which may affect phase-wise analyses; future stimuli should balance phase lengths.
• The study used a single movie excerpt; results may be content-specific and require replication across diverse narratives and full-length stories.