Paper deep dive

Transition-Related Potentials as Markers of Narrative Comprehension in Continuous EEG

Bálint Csanády, Péter Vedres, Kristóf Zsolt Makó, Orsolya Papp-Zipernovszky, Márta Volosin, Dávid Apagyi, András Lukács, András Bálint Kovács, Zoltan Nadasdy

Year: 2026Venue: arXiv preprintArea: q-bio.NCType: PreprintEmbeddings: 85

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 95%

Last extracted: 7/24/2026, 2:09:18 AM

Summary

This study investigates Transition-Related Potentials (TRPs) as markers of narrative comprehension in continuous EEG recordings. Participants watched short films (coherent vs. incoherent/scrambled) while EEG was recorded. The authors demonstrate that TRPs aligned to cinematic cuts exhibit ERP-like structures (P2, P3, LPC) and are significantly modulated by narrative coherence, with incoherent narratives eliciting stronger posterior negativity. Furthermore, a compact Deep Neural Network (DNN) successfully detected cut-related EEG signatures directly from continuous recordings, generalizing across films and subjects, offering a semi-automated framework for analyzing naturalistic cognitive processing.

Entities (7)

Electroencephalography · methodology · 99%Transition-Related Potentials · neurophysiological-signal · 98%Deep Neural Network · computational-model · 97%Event-Related Potential · neurophysiological-signal · 96%Narrative Coherence · experimental-condition · 95%Cinematic Cuts · stimulus-event · 94%Narrative Engagement Scale · psychometric-tool · 92%

Relation Signals (5)

Cinematic Cuts → triggers → Transition-Related Potentials

confidence 97% · we collected continuous EEG while participants watched short films and extracted potentials aligned to sharp cinematic transitions (cuts).

Narrative Coherence → modulates → Transition-Related Potentials

confidence 96% · By comparing coherent films with scene-scrambled versions... we find that these responses are systematically shaped by narrative context.

Transition-Related Potentials → exhibitsstructureof → Event-Related Potential

confidence 95% · We demonstrate that such transition-related potentials (TRPs) exhibit canonical ERP-like temporal structure associated with significant information processing.

Deep Neural Network → detects → Cinematic Cuts

confidence 93% · We then show that the cut-related EEG signature can be recovered directly from group-averaged continuous recordings with a compact deep neural network (DNN).

Narrative Coherence → correlateswith → Narrative Engagement Scale

confidence 90% · At the individual level, coherent film viewing was strongly correlated with higher engagement scores...

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Harnessing the potential of electroencephalography (EEG) for brain research is fundamentally limited by intrinsic noise and the diffuse projection of brain-generated activity over the scalp. The standard event-related potential (ERP) paradigm addresses this limitation by relying on repeated independent trials, albeit at the cost of moving away from naturalistic experimental conditions. As a more naturalistic alternative, we collected continuous EEG while participants watched short films and extracted potentials aligned to sharp cinematic transitions (cuts). We demonstrate that such transition-related potentials (TRPs) exhibit canonical ERP-like temporal structure associated with significant information processing. By comparing coherent films with scene-scrambled versions containing matched post-cut sensory input, we find that these responses are systematically shaped by narrative context. We then show that the cut-related EEG signature can be recovered directly from group-averaged continuous recordings with a compact deep neural network (DNN). The detector generalized across films and subject groups, and the resulting TRPs reproduced the main context-dependent effects observed for manually annotated cuts. These results indicate that narrative context leaves a measurable signature in EEG responses, that this signature can be detected directly in continuous recordings, and that such detections provide a semi-automated framework for analyzing how viewers process and understand film narratives. We propose that the method outlined here can be adapted to parse EEG responses to other forms of continuous stimulation, providing a general tool for probing experimental conditions that are closer to natural human experience.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

84,001 characters extracted from source content.

Expand or collapse full text

Transition-Related Potentials as Markers of Narrative Comprehension in Continuous EEG Bálint Csanády a,∗ , Péter Vedres b,c , Kristóf Zsolt Makó a , Orsolya Papp-Zipernovszky a,d , Márta Volosin f , Dávid Apagyi a , András Lukács a , András Bálint Kovács a , Zoltan Nadasdy a,e,g,∗ a ELTE Eötvös Loránd University, Budapest, Hungary b Budapest University of Technology and Economics, Budapest, Hungary c HUN-REN Wigner Research Centre for Physics, Budapest, Hungary d Semmelweis University, Budapest, Hungary e University of Miskolc, Miskolc, Hungary f University of Szeged, Szeged, Hungary g The University of Texas at Austin, Austin, TX, USA Abstract Harnessing the potential of electroencephalography (EEG) for brain research is fundamentally limited by intrinsic noise and the diffuse projection of brain-generated activity over the scalp. The standard event-related potential (ERP) paradigm addresses this limitation by relying on repeated independent trials, albeit at the cost of moving away from naturalistic experimental condi- tions. As a more naturalistic alternative, we collected continuous EEG while participants watched short films and extracted potentials aligned to sharp cinematic transitions (cuts). We demonstrate that such transition-related potentials (TRPs) exhibit canonical ERP-like temporal structure associated with significant information processing. By comparing coherent films with scene-scrambled versions containing matched post-cut sensory input, we find that these responses are systematically shaped by narrative context. We then show that the cut-related EEG signature can be recovered directly from group-averaged continuous recordings with a compact deep neural network (DNN). The detector generalized across films and subject groups, and the resulting TRPs reproduced the main context-dependent effects observed for manually annotated cuts. These results indicate that narrative context leaves ∗ Corresponding author: csbalint@protonmail.ch ∗ Corresponding author: zoltan@utexas.edu arXiv:2607.20720v1 [q-bio.NC] 22 Jul 2026 a measurable signature in EEG responses, that this signature can be detected directly in continuous recordings, and that such detections provide a semi- automated framework for analyzing how viewers process and understand film narratives. We propose that the method outlined here can be adapted to parse EEG responses to other forms of continuous stimulation, providing a general tool for probing experimental conditions that are closer to natural human experience. Keywords: EEG, ERP, deep learning, LSTM, cut detection, short films 1. Introduction Our conscious experience is dominated by the parsing of continuous sen- sory streams into discrete, meaningful segments, through which we construct the underlying narrative structure of the world around us. This segmentation is often triggered by sensory events that elicit recursive perceptual cycles of feature extraction, scene segmentation, object recognition, and integration with prior knowledge (Stojić and Nadasdy, 2024). Understanding the neu- ronal processes underlying such events remains a challenge in cognitive neu- roscience, especially when relying on non-invasive and readily available tools such as electroencephalography (EEG). What triggers a cognitive event in one individual may not trigger the same event in another, and even when an event is shared, the timing of the corresponding cognitive EEG markers may differ across people. This variability is compounded by the fact that scalp-recorded activity has a low signal-to-noise ratio (SNR) and reflects the simultaneous activation, interference, and superposition of multiple neural sources. Moreover, the mapping from neural sources to scalp potentials is not uniquely invertible, so recovering the underlying neural events from con- tinuous EEG constitutes an ill-posed inverse problem (Nunez and Srinivasan, 2006). A standard way to separate stimulus-related activity from noise is to improve the SNR by averaging EEG segments aligned to repeated triggers. In the event-related potential (ERP) paradigm, this trigger is usually a stimulus onset. EEG segments aligned to repeated presentations of the same stimulus or stimulus category allow related activity to combine constructively, while activity unrelated to the stimulus is attenuated (Luck, 2014; Kappenman and Luck, 2016; Boudewyn et al., 2018). This yields reproducible ERP waveforms and has become a central method for tracking cognitive processes with EEG. 2 However, the method typically relies on a trial-by-trial structure in which stimuli are repeatedly presented in isolation, usually in a randomized order, so that trials can be treated as approximately independent. This requirement is in tension with cognition in its natural form, where each event unfolds in relation to those preceding it. In this way, the continuous experience that motivates the question is partly sacrificed for methodological control. In contrast, allowing subjects to watch short films as a continuous, task- free visual experience provides a more naturalistic window into brain activ- ity (Sonkusare et al., 2019). Although professionally produced films differ markedly from the trial-by-trial stimulus presentations typical of ERP stud- ies, they are far from unstructured stimuli. Films contain precisely planned transitions (cuts), and salient yet unpredictable events that closely emulate real-life scenarios. We recorded EEG from participants while they watched a film, treated the film as a continuous naturalistic stimulus, and used the cuts separating shots as event triggers. The method preserves the continuity and contextual depen- dence of naturalistic experience, but introduces transitions that are discrete, frame-accurate, and shared across viewers. While individual differences in in- terpretation may introduce substantial inter-individual EEG variance (Mars et al., 2008), cuts serve as precisely timed markers around which average waveforms can be computed, approximating the kinds of event boundaries that are thought to structure ongoing experience (Zacks et al., 2007, 2010; Kurby and Zacks, 2008; Magliano and Zacks, 2011). Related event-boundary and cinematic-transition studies have shown that such transitions can elicit characteristic ERP-like responses (Silva et al., 2019; Sanz-Aznar et al., 2023). As such, cuts provide temporal anchors for EEG analysis through transition- related potentials (TRPs), marking moments at which attention may shift and new information may be processed. If cuts can serve as temporal anchors for naturalistic EEG, the next ques- tion is what kind of neural response they elicit. The EEG responses to cuts themselves have been characterized in some detail. Cut-locked responses con- tain both early components and later components (Matran-Fernandez and Poli, 2015); they are sensitive to whether a cut joins related or unrelated ma- terial (Francuz and Zabielska-Mendyk, 2013), and they vary with the editing technique used (Heimann et al., 2017). What has not yet been examined is whether and how these responses depend on narrative coherence, that is, on the extent to which transitions are supported by the preceding narrative con- text. Furthermore, to date, these responses have primarily been measured 3 by averaging around cuts whose timing was already known. It remained an open question whether cut-related responses can be detected within the continuous EEG recordings without prior knowledge of the cut timings. In this study, we investigate both questions. First, we characterize the effect of narrative coherence on cut-locked TRP signals. Second, we pro- pose a population-level deep-learning model for detecting cut-related EEG responses in continuous recordings. In addition to the model’s ability to iden- tify cuts with great precision, we demonstrate that the detected timestamps are suitable substitutes for the original cut annotations, in the sense that the resulting TRPs still exhibit the same characteristics in relation to narrative coherence as their hand-annotated counterparts. Finally, we compare our cut-detection approach with alternative methods to better understand the factors driving model performance. Together, these analyses provide a semi- automated framework for using EEG to study how viewers segment, update, and comprehend film narratives during naturalistic viewing. 2. Experimental Setup The experiment was conducted at the Institute of Psychology, Univer- sity of Szeged, between October 2022 and May 2023. Participants sat in a comfortable chair in a dimly lit, sound-attenuated, and electrically shielded room. A 20-inch LCD monitor (LG Flatron L206WTQ-SF; LG Electronics) and two speakers were placed on a table in front of them. Participants were presented with one of two short romantic films with minimal dialogue and a clear narrative structure: The Art of Love (Farinella, 2013) or City Lights (Wiles, 2016). Hereafter, we refer to them as Art and City, respectively. In addition to the original film with a coherent narrative, an incoher- ent version was created by a professional editor, who cut the film at scene boundaries and randomly reordered the scenes. Because simple randomiza- tion could leave two consecutive takes in place, we enforced a scrambling procedure that preserved no consecutive takes (Figure 1). Each participant viewed the coherent version of one film and the incoherent version of the other. Thus, participants watched either the coherent version of Art and the incoherent version of City, or vice versa. The order of film presentation (film title and narrative coherence) was counterbalanced across participants. After each film, participants completed the validated Hungarian version of the Narrative Engagement Scale (Papp-Zipernovszky et al., 2024). The Nar- rative Engagement Scale is a 12-item self-report measure of audience engage- 4 ment with narrative content, originally developed by Busselle and Bilandzic (2009). It contains four subscales (Narrative Understanding, Attentional Fo- cus, Emotional Engagement, and Narrative Presence), and each response is given on a 7-point Likert scale. Coherent narrative (original) Incoherent narrative Figure 1: Editing the films with incoherent narratives. The incoherent versions were created by shuffling the takes separated by cuts. Each participant watched one version of each film, one with a coherent narrative and one with an incoherent narrative. 2.1. Participants The experiment included 34 university students aged between 18 and 47 years (10 men, 24 women; mean age: 22.82 years, SD: 5.48; 5 left-handed). Participants were recruited through university courses and mailing lists, and received no financial compensation. All reported normal or corrected-to- normal vision, normal hearing, and no history of psychiatric or neurological disorders. None were studying cinematography or related fields. All partic- ipants provided written informed consent after receiving a full explanation of the experimental procedure. The study was approved by the United Re- view Committee for Research in Psychology (Hungary, EPKEB – number 2021-123), and was conducted in accordance with the Declaration of Helsinki (World Medical Association, 2025). 2.2. EEG Recording Continuous EEG was recorded at a sampling rate of 512 Hz using a BioSemi ActiveTwo system (BioSemi B.V., 2001). A total of 34 active elec- trodes (32 scalp electrodes and 2 additional electrodes) were mounted on an elastic cap according to the 10/20 system (Nuwer et al., 1998). Electrode impedances were kept below 30 kΩ. Stimulus presentation was controlled 5 using MATLAB (Version 9.0.0, R2016a; The MathWorks Inc.) and Psy- chophysics Toolbox Version 3.0.18 (Brainard, 1997; Kleiner et al., 2007), running under the Windows 10 operating system. 2.3. EEG Preprocessing The EEG data were processed in MNE-Python (Gramfort et al., 2013) using the BioSemi 32-channel montage. Cuts were manually annotated at frame accuracy and stored as millisecond timestamps corresponding to the start of the first frame following each cut. Bad channels were automatically identified in MNE on the basis of a local outlier factor criterion and sub- sequently reconstructed by spatial interpolation. Line noise was removed by a 50 Hz notch filter, and a 0.1–60 Hz band-pass filter was applied to re- duce low-frequency drift and high-frequency noise. Extracted EEG segments were normalized per channel by dividing by the standard deviation com- puted over the full epoch window (−200 to 1200 ms), then baseline-corrected by subtracting the mean amplitude of the−200 to 0 ms pre-stimulus interval. To visualize the spatial distribution of the cut-TRP response, electrode- wise grand-average waveforms were computed separately for each film by averaging across participants. Figure 2 shows the resulting butterfly plots for the coherent versions of the films, together with the corresponding anterior and posterior averages. A separation between the anterior (negative) and posterior (positive) electrode groups is clearly visible, motivating the use of anterior and posterior electrode-group averages in the TRP analyses below. 0.20.00.20.40.60.81.01.2 Time (s) 0.2 0.0 0.2 0.4 0.6 0.8 Amplitude Posterior average Anterior average Fp AF F FC C CP P PO O Electrode (a) Art 0.20.00.20.40.60.81.01.2 Time (s) 0.2 0.0 0.2 0.4 0.6 0.8 Amplitude Posterior average Anterior average Fp AF F FC C CP P PO O Electrode (b) City Figure 2: Butterfly plots of the average TRP signals aligned to cuts in the coherent versions of the films, by individual electrodes and anterior/posterior averages. The color of the electrode traces is based on their anteriority. 6 3. Results 3.1. TRP Analysis Rather than partitioning the experiment into discrete trials, we annotated the timing of the cuts in the films and treated them as naturally occurring temporal markers in the continuous EEG recordings. Because cinematic cuts can take several forms, ranging from gradual dimming or translucent tran- sitions to instantaneous frame changes without an intervening blank frame, we defined each cut as the earliest detectable transition between two scenes. If cuts function as event boundaries at which significant new information is introduced, the corresponding cut-locked TRPs should contain ERP-like components, interpreted within the standard ERP framework (Luck, 2014; Picton et al., 2000). As anticipated, the TRP displayed in Figure 2 reproduced a prototypi- cal posterior positivity in the P2 range (∼ 250–400 ms), followed by promi- nent late components including the P3/P300 (∼ 400–600 ms) and LPC/P600 (≳ 600 ms) over the posterior domain. This was complemented over the an- terior electrodes by a modest N270-like deflection and a sustained late neg- ativity between 400 and 1000 ms. One striking feature of the TRPs was a divergence between anterior and posterior activity. This divergence began with a steep increase in occipital positivity at approximately 300 ms after cut onset, leading to the peak formation of P2, and was accompanied by the simultaneous evolution of anterior negativity reaching the minimum at ∼ 500 ms, followed by a convergence complete at ∼ 1 s. Conspicuously, the earlier exogenous visual components, such as P50 and N100, were not expressed. This pattern is consistent with previous ERP studies using dy- namic video stimuli, in which early visual components were weak or absent under continuous stimulus presentation, whereas later components related to scene comprehension, semantic integration, and post-cut updating remained robustly detectable (Sitnikova et al., 2003, 2008; Sanz-Aznar et al., 2023). These early sensory components are likely attenuated, refractory, or tempo- rally smeared in continuous film viewing, because the visual system is already engaged by ongoing stimulation and because the exact perceptual registra- tion of a cut may vary across edits and observers. Despite the possible arbitrariness of onset detection, the ERP-like components were reproduced with temporal precision characteristic of event-related responses. To investigate whether and how the cut-locked response was shaped by the coherence of the film’s narrative, we compared TRPs derived from the 7 coherent and incoherent versions of each film. Since the cut-locked TRPs demonstrate a clear separation of anterior and posterior activity, we com- pared the anterior and posterior average TRP signals (Figure 3). Clear sta- tistical differences were observed between TRP amplitudes following coherent vs. incoherent cuts, highlighted by green shaded intervals. 0.20.00.20.40.60.81.01.2 Time (s) 0.4 0.2 0.0 0.2 0.4 0.6 Amplitude Coherent anterior Coherent posterior Incoherent anterior Incoherent posterior (a) Art 0.20.00.20.40.60.81.01.2 Time (s) 0.4 0.2 0.0 0.2 0.4 0.6 Amplitude Coherent anterior Coherent posterior Incoherent anterior Incoherent posterior (b) City Figure 3: Anterior and posterior TRPs of the cuts within coherent and incoherent narrative versions of the same film. Cuts are matched one-to-one. Green bars indicate significant clusters. Statistical significance was assessed using a spatio-temporal cluster per- mutation test across subjects (Maris and Oostenveld, 2007). Candidate clus- ters were formed across adjacent time samples and neighboring EEG sensors using a one-way F-statistic with a cluster-forming threshold of p < 0.0027 (3σ). Clusters were considered significant at a threshold of p < 0.0456 (2σ), estimated using 10 000 permutations. This conservative cluster-forming threshold was chosen to favor focal, high-confidence effects over broad clus- ters driven by weaker distributed differences. In both films, the posterior average showed a stronger late negative deflection when the narrative was incoherent. Additionally, in Art, there was a statistically significant nega- tive bump in the incoherent posterior average, and the amplitude attenuated more slowly during cuts related to an incoherent narrative context. The exact details of the clusters are shown in Table 1. It is worth pointing out that in these comparisons the film segments fol- lowing the cuts were paired one-to-one with frame precision. This means that the TRPs compared aggregate responses from time windows in which the immediate post-cut audiovisual information presented to participants was identical across the coherent and incoherent versions. As noted above, 8 Filmt min t max p-valueChannels Art0.230 0.271 0.0106 CP1,CP2,CP5,CP6,O1,O2,Oz,P3,P4,P7,P8,PO3,PO4,Pz Art0.350 0.469 0.0018 AF3,AF4,C4,CP5,CP6,F3,F7,F8,FC5,FC6,Fp1,Fp2,P4,P7,T7,T8 Art0.627 1.008 0.0011 CP1,CP2,Cz,F3,F7,FC1,FC5,P3,P4,PO3,Pz,T7 City0.381 0.506 0.0042 CP1,CP2,CP5,CP6,F8,FC6,O1,Oz,P3,P4,P7,P8,PO3,PO4,Pz,T7,T8 Table 1: Significant clusters for the Art and City films. coherence assignment was counterbalanced across films, so the observed ef- fects cannot be attributed to a fixed difference between participant groups. Consequently, the measured effects reflect the influence of preceding context, including narrative comprehension. To assess whether the differences observed over posterior and anterior electrode groups were related to participants’ engagement with the narra- tive, we analyzed the post-experiment questionnaire completed by each par- ticipant. Table 2 shows the aggregate Narrative Engagement scores from the questionnaire completed after each film. At the individual level, coherent film viewing was strongly correlated with higher engagement scores (Art: 0.77; City: 0.73). Overall, the coherent versions of the two films received signifi- cantly higher engagement scores than their incoherent versions, providing a plausible behavioral correlate of the TRP differences. Moreover, people who watched the incoherent version of Art scored somewhat lower on the scale, which is consistent with the stronger TRP difference; however, the confidence intervals overlap, so this effect cannot be taken as conclusive. NarrativeCoherentIncoherent Film Art CityArt City Mean score64.2 66.342.6 45.3 CI 95% 4.8 3.14.5 7.5 Table 2: Narrative Engagement mean scores and confidence intervals by film. To elucidate the main factors underlying the modulation of TRP ampli- tudes, we performed a temporal Principal Component Analysis (PCA) on cut-locked TRP waveforms (Picton et al., 2000). For a given film (e.g. Art), we extracted the coherent and incoherent TRP channel averages from 200 to 1200 ms. This yielded 64 observations in a 512-dimensional temporal feature space. PCA on these observations produced 512-dimensional temporal eigen- vectors and principal component (PC) scores consisting of 32 + 32 weights, 9 one for each channel in each condition. Figure 4 summarizes the results of this analysis, performed on Art and City independently. Notably, the result- ing eigenvectors were similar across films, and the first PC scores highlight the main anterior–posterior polarity separation we have already remarked on. This evidence supports the anterior–posterior polarity separation as the strongest topographic pattern associated with the processing of cuts in both coherent and incoherent narratives. The second PC, in contrast, captured a posterior topographic pattern related to narrative coherence (Figure 4). We can see that PC2 modulates the extensive temporal-parietal negativity of the occipito-parieto-temporal dipole for the incoherent narratives for both films while it was completely absent during watching the coherent films. A spatial variant of the PCA analysis is available in the Appendix (Figure B.13). 0.20.40.60.81.01.2 Time (s) 0.06 0.04 0.02 0.00 0.02 0.04 0.06 0.08 Amplitude PC1 (88.8%) PC2 (6.3%) (a) Art principal eigenvectors. 0.20.40.60.81.01.2 Time (s) 0.06 0.04 0.02 0.00 0.02 0.04 0.06 0.08 Amplitude PC1 (90.1%) PC2 (4.8%) (b) City principal eigenvectors. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (c) Art coherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (d) Art incoherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (e) City coherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (f) City incoherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (g) Art coherent PC2. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (h) Art incoherent PC2. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (i) City coherent PC2. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (j) City incoherent PC2. Figure 4: Temporal PCA analysis of the TRPs. 10 3.2. Cut Detection We have shown that TRPs are sensitive markers of cognitive events that not only reproduce critical features of ERPs, but also reveal important as- pects of film narrative processing without participants being explicitly in- structed to pay attention to them. This finding motivated us to use deep learning to detect significant cognitive events that occur naturally during the processing of the complex sensory stream typical of film viewing. More specifically, we trained a deep neural network (DNN) to detect manually an- notated cinematic cuts from EEG signals alone and then applied the model to recover cut-like EEG patterns, including patterns not necessarily elicited by cinematic cuts. DNNs are well established for detecting events in EEG, most prominently for seizure detection (Roy et al., 2019) and sleep-stage scoring (Supratak et al., 2017), and compact architectures developed for brain-computer inter- faces have shown that such models can generalize across tasks and subjects (Lawhern et al., 2018; Schirrmeister et al., 2017). These events, however, are defined by the neural state itself: seizures and sleep stages can be identified from EEG by professionals, which guarantees the presence of a detectable neural signature. In contrast, cinematic cuts are defined objectively by the film and independently of brain activity. In addition, cuts are audiovisual transitions that may elicit sensory and cognitive processes related to the clo- sure of the previous event and the beginning of a new one, consistent with event-segmentation accounts of film comprehension and cinematic continuity (Zacks et al., 2007, 2010; Magliano and Zacks, 2011; Smith, 2012), thereby modulating late ERP-like components. Studies using naturalistic stimuli have applied inter-subject correlation to identify stimulus-driven EEG re- sponses shared across viewers (Hasson et al., 2004). Our analysis also relied on common components of EEG signals, but focused on a more complex objective. We asked not only whether the events could be detected with accuracy exceeding correlation-based methods, but also whether their onsets could be localized with sufficient temporal precision to recover high-quality TRPs. To fulfill these requirements, we trained a DNN to recover the timing of cinematic cuts directly from the synchronized EEG signals. The population- level detector operated on short sliding windows of the averaged EEG signal, producing a continuous cut-likelihood score that was converted into discrete cut predictions by detecting peaks in the score trace (Figure 5). To clarify the 11 detector, we first outline the DNN model used to compute the cut-likelihood score. Figure 5: Snapshot from the EEG cut detection on the film City. 3.2.1. DNN Model Let X ∈R S×32×T denote the cut-synchronized EEG recordings for one version of one film, where S is the number of subjects, 32 is the number of channels, and T is the number of samples. To ensure comparable signal contribution from each participant, we first normalized each response using a causal rolling baseline Z-score at the participant level. Specifically, each EEG sample was standardized channel by channel relative to the mean and stan- dard deviation of the preceding 200 ms of the same signal. The normalized signals were then averaged across subjects to obtain ˆ X ∈R 32×T . This across- subject averaging improved the signal-to-noise ratio by emphasizing activity that was temporally aligned and shared across subjects. ˆ X was divided into overlapping windows W i ∈R 32×512 of 512 samples (1 s). Each extracted window was assigned a ground-truth label y(W i ) based on its temporal proximity to annotated events. For each window onset t i and nearest cut annotation t ∗ i (both measured in samples), we defined y(W i ) as: y(W i ) = min max 1.1− |t i − t ∗ i | 320 , 0 , 1 . 12 Architecture. We built the cut detector as a compact classifier with recurrent convolutional architecture. A 1D convolutional embedding first compressed the input EEG window to a shorter temporal representation. We employed an embedding consisting of two convolutional stages with output channel sizes 32 and 64, using kernel size and stride of 3 in the first and 4 in the second layer. The embedded sequence was then processed by a two-layer bidirectional residual LSTM with 32 hidden units. The processed sequence was average-pooled resulting in a single 64 dimensional vector. Finally, this vector was processed by a dense classification head with 16 hidden neurons, producing a scalar logit for each input window. The convolutional layers captured short-timescale temporal patterns, while the LSTM examined how these patterns evolved across the window. Training. The models were trained on randomly sampled mini-batches of size 64, with binary cross-entropy loss on logits with the continuous target value y(W i ), and AdamW optimization (learning rate: 0.0005, weight decay: 0.001). For sampling purposes, windows where y(W i ) > 0 were considered positive, and windows with y(W i ) = 0 were considered negative. To balance training, negative windows were randomly under-sampled in each epoch to match the number of positive windows. For regularization, we trained the networks using batch normalization and dropout (0.25). To augment train- ing, we added random white noise (SD: 0.1) to each sample before presenting it to the network. Unless otherwise stated, models were trained for 20 epochs. Inference Methodology. At inference time, the model was evaluated at every eighth window, providing a time series of model predictions throughout the film. This model score trace was smoothed using a 20-sample moving average window. Cuts were predicted by peak finding on the smoothed score trace. A peak was accepted only if it satisfied the chosen minimum height (0.5), minimum width (3), and minimum inter-peak distance (0.625 s). A detected cut was classified as a true positive (TP) if it fell within an acceptance window of 0.625 s centered on a true cut. In other words a de- tected peak corresponding to EEG window W i was accepted if y(W i )≥ 0.575. Similarly, detected cuts were classified as false positives (FPs) if they fell outside all acceptance windows, and acceptance windows with no detected peak were counted as false negatives (FNs). Double-counting of true posi- tives was prevented by setting the minimum inter-peak distance equal to the acceptance-window width. 13 3.2.2. Results In the following analysis, we compared different ways of partitioning the four available datasets (2 films, 2 narrative coherence modes) into training and evaluation sets, as summarized in Table 3. Each row defines one gener- alization setup. The DNN was trained on Group A and evaluated on Group B, and the same comparison was also performed in the reverse direction. These splits tested whether the detector generalized across film identity, par- ticipant group, narrative coherence, or the joint change of film identity and participant group. GeneralizationGroup A Group B FilmArt, ::: ArtCity, :::: City Subjects Art, :::: City ::: Art, City Coherence Art, City ::: Art, :::: City Film & subjects ArtCity :::: Art :::: City Table 3: Dataset splits used for DNN generalization tests. Models were trained on one group and evaluated on the other, with both directions tested. The wavy underline marks the incoherent version of a film (e.g. ::: Art). We performed five independent training sessions for each setup. For each training session, we summarized performance by averaging the F1 scores for detected peaks across epochs 10 to 20. Figure 6 shows the box plot of the re- sults. As expected, the most challenging training setup was the one in which the model had to generalize across both film and subject group. This was likely due to both the amount of training data being halved and the harder generalization task. However, this setup prevented the detector from exploit- ing film- or subject-specific idiosyncrasies and tested whether the learned EEG signature of cuts transferred across stimulus material. The remaining experimental results in this section are based on this training setup. The difference in performance between the two films was also notable. Although all training setups achieved higher F1 scores when tested on the incoherent version of Art, this pattern did not hold for City. Error Analysis. The primary goal of EEG-based cut detection was to test whether TRPs triggered by visual and cognitive events are robust enough relative to noise to support cut identification. At the same time, detection errors, especially false positives, can reveal the types of patterns and narra- tive events that share EEG characteristics with cinematic cuts. Therefore, 14 Art (coherent)Art (incoherent)City (coherent)City (incoherent) Recording 0.775 0.800 0.825 0.850 0.875 0.900 0.925 0.950 F1 Generalize on: Film Subjects Coherence Film & Subjects Figure 6: Box plot of the neural model’s performance on the different EEG recordings, by the generalization objective. these errors may help to clarify what kinds of stimuli TRPs capture from the EEG. Upon initial inspection, false negatives did not appear to form a random subset: at least some corresponded to cuts in which the audiovisual transitions were too subtle, or in which transients occurred in rapid succes- sion, making the signal difficult to resolve. For example, cuts corresponding to camera zoom or angle shifts seemed hard to notice. Likewise, false posi- tives did not seem to appear at random either: many were auditory and/or visual transitions in the film, such as a light switching on or off, a picture be- ing shown on a computer monitor, or a change of framing where a narratively important object or person appears. These may share common sensory or narrative features with true cuts. In order to analyze the errors by the above aspects, we categorized them manually (Table 4). This error analysis used all annotated cinematic cuts, rather than only the matched subset used for the TRP comparisons. False positives were classified as transients when an intense auditory or visual tran- sient was present other than a cinematic cut, as significant events when the detection also coincided with a narratively relevant event, or as undetermined otherwise. False negatives were classified as weak transients when the audio- 15 FilmArtCity NarrativeCoherent IncoherentCoherent Incoherent% True positive123158147118 False positive428175 Transient0128546.3% Significant event4135040.7% Undetermined 034013.0% False negative4472859 Weak transient7081018.1% Sub-window125162844.2% Undetermined 25242137.7% Table 4: Cut detection outcomes by EEG recording, manual error assessment. visual transient associated with the cut was subtle (e.g. cut involved only a change of camera angle), as sub-window when a neighboring cut was present within a 1 s window, as undetermined when neither of these applied. Over- all, the error structure suggests that the detector learned a physiologically meaningful cut-related response rather than an arbitrary mapping from EEG fluctuations to annotations. The majority of false negatives (62.3%) occurred when the cinematic transition was perceptually weak or temporally crowded by neighboring cuts, whereas false positives mostly coincided (87%) with strong audiovisual changes or narratively meaningful events. This pattern supports the interpretation that the detector was sensitive to EEG activity resembling event-boundary processing, even when such processing was not always tied to a formally annotated cinematic cut. TRP Validation. To validate the efficacy of the model outputs, we repeated the TRP analysis from Section 3.1 using the detected cuts (both true and false positives). The accuracy of the recovered TRPs is remarkable, considering that the model was blind to the cut-segmentation of the films. Figure 7 shows the resulting TRPs, and Table 5 lists the cluster permutation test results. Most importantly, similar clusters were recovered as in the hand- annotated TRPs (Figure 3 and Table 1) with marginally smaller confidence (larger p-values). These results not only corroborate the model’s accuracy in finding cuts, but also highlight its temporal precision, which was essential for obtaining TRP waveforms with clear ERP-like features. Additionally, temporal PCA based on the detected cut TRPs yielded similar results to PCA based on the original cuts (Figure B.14 in the Appendix). 16 Filmt min t max p-valueChannels Art0.240 0.293 0.0128 CP1,O1,O2,Oz,P3,P4,P7,P8,PO3,PO4,Pz Art0.369 0.479 0.0045 AF3,AF4,C3,CP1,CP5,F3,F7,FC5,Fp1,Fp2,P3,P7,T7 Art0.703 1.070 0.0025 CP1,CP2,Cz,F3,F7,FC1,FC5,P3,P4,PO4,Pz City0.367 0.502 0.0053 CP1,CP5,CP6,F8,FC6,O1,O2,Oz,P3,P4,P7,P8,PO3,PO4,T7,T8 Table 5: Significant clusters for the detected cuts in Art and City. 0.20.00.20.40.60.81.01.2 Time (s) 0.4 0.2 0.0 0.2 0.4 Amplitude Coherent anterior Coherent posterior Incoherent anterior Incoherent posterior (a) Art 0.20.00.20.40.60.81.01.2 Time (s) 0.4 0.2 0.0 0.2 0.4 Amplitude Coherent anterior Coherent posterior Incoherent anterior Incoherent posterior (b) City Figure 7: Anterior and posterior average TRPs of the detected cuts within the coherent and incoherent narrative versions of the same film. Green bars indicate significant clusters. Classifying Narrative Coherence. Finally, we asked whether narrative coher- ence could be decoded from single-subject EEG responses. To address this question, we computed subject-level TRPs separately for each participant and each film condition by averaging the cut-locked EEG segments of that participant. Each subject-level TRP was then cropped to the 200–1000 ms post-cut interval, and the resulting 32 × 409 channel-by-time matrix was flattened into a single feature vector. We used these feature vectors to clas- sify whether the corresponding TRP came from the coherent or incoherent version of the film. Classification was evaluated with leave-one-subject-out cross-validation, so that the subject being tested was never included in the training set. Because each subject contributed only one TRP per film, the classifier could not rely on repeated samples from the same subject within a film-specific dataset. Thus, successful classification would indicate that subject-level TRPs retained information about narrative coherence, rather than merely reflecting subject-specific idiosyncrasies. We repeated the anal- ysis using both the original manually annotated cuts and the cut events detected by the DNN model. In the manually annotated version, cuts were 17 matched one-to-one across coherent and incoherent conditions. This ensured that the classifiers could not rely on systematic differences in the immediate post-cut sensory input. FilmArtCity Cut EventOriginal DetectedOriginal Detected Gradient Boost0.710.740.820.74 LDA0.850.740.740.77 Logistic Regression 0.820.850.850.77 Random Forest0.880.850.740.71 Ridge Classifier0.790.910.770.77 SVM0.820.850.820.77 XGBoost 0.820.740.770.79 Average0.8150.8110.7860.756 Table 6: Classification accuracy scores of narrative coherence based on true cuts and DNN- detected cut-like events. The numbers represent accuracy scores achieved by different ML models using leave-one-subject-out cross-validation. Narrative coherence was classified using different machine learning (ML) approaches. As Table 6 indicates, average classifier performance decreased only marginally when detected cuts were used instead of original cut anno- tations. The classifiers correctly identified whether subject-level TRPs were derived from coherent or incoherent narratives for the majority of held-out subjects. Although the average accuracy of the ML models was > 75%, substantially above the empirical chance level, the main point of this anal- ysis was not to maximize classification performance. Instead, the aim was to demonstrate the efficacy of this method to extract subject-level cogni- tive information from EEG, and to further corroborate the quality of the DNN-based cut detections. Ablations. A baseline comparison is reported in Appendix A. In brief, the main DNN detector outperformed the anterior–posterior divergence, TRP- template-fit, and correlation-based baselines in the film-generalization set- ting. Importantly, the DNN localized cut onsets more accurately: on aver- age, true positives were within one frame of the true cut onsets, while simpler peak-detection baselines approached the timing error expected from random detections within the acceptance window. Corroborating that temporal pre- cision was critical for recovering high-quality TRPs from detected events. 18 4. Discussion EEG is the most accessible noninvasive tool for studying the brain’s macroscopic electric potentials, which correlate with cognitive functions at millisecond precision. To unlock EEG’s full potential in diverse experimen- tal settings (Wang et al., 2018; Sonkusare et al., 2019; Lau-Zhu et al., 2019; Kaushik et al., 2022) and access deeper layers of information processing in the brain we need to improve the sensitivity and specificity of the data processing method. Even when extrinsic noise is managed, the biggest challenge is the large-amplitude intrinsic noise frequently obscuring the relationship between neuronal responses, particularly when stimuli are continuous and events oc- cur at irregular intervals. Conventional ERP analysis addresses this by av- eraging time-locked responses across repeated stimulus presentations (Luck, 2014), but this repetition renders most experiments unnatural and contrived. One conventional solution developed to overcome this limitation was single- trial analysis, replacing waveform averaging with spectral or spatial filters and classifiers (Blankertz et al., 2011). Instead, we developed an alterna- tive approach to extract ERP-like transition-related potentials (TRPs) from continuous EEG recorded while subjects watched short films, time-locked to well-defined sensory events, such as cinematic cuts. The first part of the study demonstrated that TRPs reproduce essen- tial cognitive components of ERPs, specifically P2, P3 and N4 (Figure 2). Among those, we showed that TRPs are systematically modulated by nar- rative context. Group-level TRPs drawn from EEGs recorded while par- ticipants watched scene-scrambled versions of films, exhibited reproducibly different characteristics relative to the TRPs computed from the EEGs while watching the original version. Specifically, we observed significant modula- tion of P2 and P3 in the posterior and N400 amplitudes in the anterior elec- trodes consistent with the coherent-incoherent versions (Figure 3). These ef- fects were controlled by the counterbalanced allocation of participants across films and by one-to-one scene matching, which ensured identical audiovisual input in the post-cut TRP source-windows. Furthermore, the TRP differ- ences aligned with participants’ subjective assessments, as scene-scrambled films were consistently rated as less narratively engaging. The observed dif- ferences were further corroborated by temporal PCA, which separated two main components underlying the modulation of TRPs (Figure 4). The first and strongest principal component, explaining ≈ 90% of signal variance, captured a broadly distributed posterior positivity: identified with the P2 19 and P3 components. It also confirmed that the large topographic separation between anterior negativity and posterior positivity was the dominant char- acteristic of cut-elicited TRPs in general, and was not specific to the narrative differences between the films. In contrast, the second principal component displayed a pronounced difference between coherent and incoherent narra- tives, with maximal expression over parietal and temporal areas. Together, these findings support two conclusions: first, that cuts set a cascade of cog- nitive components into motion, originating over occipital cortex and followed by the evolution of a sustained frontal negativity; and second, that the mod- ulation of these components by visual transients carries the fingerprints of highly abstract cognitive processing, specifically narrative coherence, a key determinant of film comprehension and, more broadly, of how meaning is constructed during continuous real-world experience. It is worth to contextualize these findings within the ERP literature. Regarding the positive components of the TRPs, both P2 and P3 are asso- ciated with the reduction of subjective uncertainty. P2 (Luck and Hillyard, 1994) is commonly evoked in auditory oddball and visual priming paradigms, and is most pronounced centrally in auditory speech perception tasks (Sig- noret et al., 2013), with a more attenuated expression in visual paradigms (Capilla et al., 2016). Its manifestation in our experiment was nonethe- less evident: given that P2 is generally associated with novelty (Barcelo and Knight, 2007), it was expected to be most prominent following a cut when the new scene is less predictable. Likewise, the P3 component is sensitive to the information-theoretic surprise value of a stimulus, with P3a over frontal and central areas reflecting attentional engagement (Polich, 2003) and P3b, with its posterior maximum, indexing stimulus unexpectedness (Duncan-Johnson and Donchin, 1977). Two competing hypotheses can be articulated regarding the P3b in our experiment. The first predicts a more prominent P3b in incoherent-narrative EEGs, since the randomized sequence of shots produces a higher rate of unexpected transitions. The second, however, draws on the relationship be- tween P3b amplitude and information-theoretic unexpectedness, as captured by Rényi entropy (Rényi, 1961), and leads to the opposite prediction. In a coherent film, predictable transitions predominate, meaning that the rare unpredictable cut carries greater surprise and may therefore elicit a larger P3b than any individual transition in an incoherent film, where all cuts are equally unexpected and no strong prior exists to be violated. This depen- dence on baseline stimulus probability is precisely what Rényi entropy for- 20 malizes, and trial-by-trial P3 modulation by such prior probabilities has been demonstrated in Bayesian paradigms (Mars et al., 2008). In addition to the positive components, we observed a prominent ante- rior negativity spanning from 250 to 1000 ms, which we associate with the cognitive process of anticipation resolution. This frontally-dominant nega- tive potential is consistent with feedback-related negativity documented in the literature (Maruyama et al., 2026). Although N400 is most commonly associated with linguistic tasks, it is not strictly a semantic component; it arises in any task involving the processing of meaning or conceptual infor- mation (Ganis et al., 1996). This has been demonstrated in image-based sentence completion paradigms, where the final word is replaced by a pre- dictable or unpredictable image rendering the sentence semantically congru- ous or anomalous, respectively (Nigam et al., 1992; Ganis et al., 1996). To our knowledge, the only prior study to elicit an N400-like response using video stimuli is that of Sitnikova et al. (2003), who presented contextually congruous and incongruous film endings. They found that incongruous end- ings evoked an enhanced anterior negativity resembling N400, as well as a more positive late positivity relative to congruous endings of short video presentations of human interactions with objects. This pattern is perfectly concordant with our own finding that cuts demarcating incoherent narratives elicit larger N400 deflections and increased late P3b positivity. Notably, the above referenced study called for exactly the paradigm we employed (p. 163): “To get a better estimate of the timing of semantic integration in videos, it might be useful to collect ERPs to critical items that have a clear point of appearance in the scene (e.g., at a scene change).” Our experiment directly answers that call. Next we turned the task around and approached it from the opposite direction. Leveraging the power of sequence-processing artificial neural net- works, we trained DNNs to exploit the characteristic EEG signature of TRPs and localize cinematic cuts in continuous EEG based solely on their electro- graphic signatures (Figure 5). More specifically, we developed a DNN-based cut-detection methodology that located the vast majority of cinematic cuts from group-averaged EEG signals with a low error rate, and most impor- tantly, with a sub-frame onset accuracy (< 40 ms). With the exception of a single outlier, cut detection achieved consistently strong performance (0.8 < F1 < 0.95; Figure 6); even in the most challenging data parti- tioning scenario, when generalizing across both, film and participant group, using only one group-averaged EEG recording for training. Nevertheless, 21 performance differences across all generalization conditions remained within ∆F1 = 0.05, underscoring the robustness of the learned representations. Al- though this approach has practical value as a method for simply detecting cinematic cuts only to a limited extent, failures of the detector may point to events with cognitive properties similar to those of cuts. Thus, investigating the detection errors proved highly informative. The majority of false negatives (62.3%) were accounted for by cuts involving in- sufficiently salient visual transients (such as transparent transition), or by cuts occurring in rapid succession within the 1-second window during which multiple detections were suppressed. False positives, by contrast, were of con- siderable interest, as they shed light on the cognitive underpinnings of cut detection. Conceptually, a false positive arises when an EEG pattern shares the electrographic fingerprint of a TRP-template without being preceded by an actual cut. Analyzing the false positives revealed a general pattern. Some false positives coincided with audiovisual transients that were very similar to cuts while technically not being cinematic cuts, such as a light in the room being switched on/off, or a camera shutter effect. Naturally, the transients associated with false positives (87%), were on a spectrum in terms of in- tensity. A general trend we found was that some of the weaker transients were also associated with narratively important moments, and this associ- ation was a likely factor in their detection. For example, in one scene, on his computer monitor, the protagonist reviews photos he took earlier, and the transitions between these photos resemble cinematic cuts, but impact only a portion of the frame. Out of the four pictures, the one that coincided with a false positive detection was the most narratively important, a por- trait of the girl the protagonist is interested in. While such events may lack a strong audiovisual transient and thus the early sensory components might have been too weak, their late cognitive components – P2, P3, and N4 – were nonetheless strong enough to be triggered by narrative events similar to cuts with respect to demanding an update to the viewer’s internal model of the story. The coincidence of false positives with such narrative events suggests that the DNN had learned to generalize TRP features beyond literal cuts, effectively detecting cognitively salient moments in the continuous EEG. The timing precision of the detector was further demonstrated by the fact that detected events could be used directly to generate TRP waveforms that retained the main narrative-context-dependent differences observed for the original cut annotations (Figure 7). In addition, a participant-level classifi- cation of narrative coherence (inferring from individual TRPs if a participant 22 watched the coherent or the incoherent version of the film), suffered only a marginal decrease in accuracy when TRPs were generated from the algorith- mically detected cuts rather than from the original annotations. These results appear even more compelling in light of the errors produced by the detec- tor: many false negatives occurred in challenging cases, such as fast-paced cuts or weak audiovisual changes, whereas many false positives coincided with strong audiovisual transients, often with narrative significance. Taken together, these results provide strong evidence that the DNN learned to iden- tify narratively significant cognitive events from EEG waveform signatures derived from cinematic cuts. 5. Conclusion Our findings collectively demonstrate that narrative comprehension leaves a measurable electrographic signature at moments of event transition, man- ifest as transition-related potentials (TRPs). Moreover, by leveraging the structure of TRPs, a DNN can be trained to extract analogous signatures from continuous EEG group-averages by capturing events that share physi- cal or cognitive attributes with cinematic cuts or other well-defined stimulus boundaries. Furthermore, with the appropriate algorithmic approach, these detections are sufficiently accurate to automatically reproduce high-quality TRP waveforms. Importantly, these results indicate, that a robust mapping exists between continuous EEG dynamics and the timing of externally de- fined narrative transitions. Such electrographic signatures offer a principled means of parsing continuous EEG for cognitive events representing episodes of information updating, uncertainty reduction, or the resolution of suspense: hallmarks of the dramatic architecture of produced films and, more broadly, of natural experience (Dini et al., 2023). This discovery opens the door to detecting meaningful transitions in other continuous stimuli, where event boundaries may not be directly observable or cannot be annotated by hand with comparable precision. Such applications would extend the TRP frame- work beyond cinematic cuts and enable the analysis of information processing across a broader range of continuous, naturalistic experiences. Future efforts will be devoted to test the generalization of TRPs and to parse real-time individual EEG. 23 6. Limitations Several limitations should be considered when interpreting these results. First, the experiment used only two short films with clear narrative structure and relatively sparse dialogue. The generality of the findings should therefore be tested on a wider range of films, genres, editing styles, and audiovisual materials. Second, the DNN detector operated on subject-averaged EEG. It should therefore be interpreted as a population-level detector rather than a real-time single-subject detector. Future work should test whether subject- adaptive models can recover comparable event timing from individual EEG recordings. Third, the interpretation of false positives and false negatives re- lied partly on manual categorization. Combining EEG-based detection with richer stimulus annotations, including auditory changes, object appearances, semantic shifts, and narrative turning points were out of scope of this study, and would be subject of further investigation. Fourth, due to the small num- ber of experimental subjects, the TRP results are based on signals averaged across both individual cuts and participants. Future studies aiming to an- alyze responses to individual cuts should therefore use substantially larger participant samples. Finally, the analyses reported here are scalp-level anal- yses. They identify reproducible temporal and topographic patterns in the EEG, but they do not localize the neural generators of these effects. Source- level analyses or multimodal recordings would be needed to determine the neural systems underlying the observed TRP components. Ethics Statement All participants provided written informed consent after receiving a full explanation of the experimental procedure. The study was approved by the United Review Committee for Research in Psychology (Hungary, EPKEB 2021-123). The study was conducted in accordance with the Declaration of Helsinki. Participants received no financial compensation. Data and Code Availability Statement The EEG dataset is available at Mendeley Data (Csanády et al., 2026). Code is available at GitHub (Csanády, 2026). Declaration of Competing Interest The authors declare that they have no conflict of interest. 24 Credit Authorship Contribution Statement Bálint Csanády: conceptualization, methodology, software, formal anal- ysis, investigation, writing – original draft, data visualization. Péter Ve- dres: conceptualization, methodology, formal analysis, investigation. Kristóf Zsolt Makó: software, investigation. Orsolya Papp-Zipernovszky: research design, psychometric assessment. Márta Volosin: research design, investiga- tion, EEG data recording. Dávid Apagyi: software, investigation. András Lukács: supervision, methodology. András Bálint Kovács: conceptualization, research design, methodology. Zoltán Nádasdy: conceptualization, methodol- ogy, supervision, writing – review and editing. Acknowledgments This research was supported by the EKÖP-25-4-I-ELTE-1093 Univer- sity Research Fellowship Program of the Ministry for Culture and Innova- tion from the Source of the National Research, Development and Innovation Fund. The research was additionally supported by the Hungarian National Research, Development and Innovation Office within the framework of the Thematic Excellence Program 2021 – National Research Sub programme: “Artificial intelligence, large networks, data security: mathematical founda- tion and applications” and the Artificial Intelligence National Laboratory Program (MILAB). This research was also funded by the OTKA research funding program under project number NKFI-138108. References Barcelo, F., Knight, R.T., 2007. An Information-Theoretical Approach to Contextual Processing in the Human Brain: Evidence from Prefrontal Lesions. Cerebral Cortex 17, i51–i60. doi:10.1093/cercor/bhm111. BioSemi B.V., 2001. ActiveTwo/Active2.5: 280-Channel, DC Amplifier, 24- Bit Resolution Biopotential Measurement System with Active Electrodes. https://w.biosemi.com/Products_ActiveTwo.htm. Accessed: 2026- 06-02. Blankertz, B., Lemm, S., Treder, M., Haufe, S., Müller, K.R., 2011. Single- trial analysis and classification of ERP components – a tutorial. NeuroIm- age 56, 814–825. doi:10.1016/j.neuroimage.2010.06.048. multivariate Decoding and Brain Reading. 25 Boudewyn, M.A., Luck, S.J., Farrens, J.L., Kappenman, E.S., 2018. How many trials does it take to get a significant ERP effect? It depends. Psy- chophysiology 55, e13049. doi:10.1111/psyp.13049. Brainard, D.H., 1997. The Psychophysics Toolbox. Spatial Vision 10, 433– 436. doi:10.1163/156856897X00357. Busselle, R., Bilandzic, H., 2009. Measuring Narrative Engagement. Media Psychology 12, 321–347. doi:10.1080/15213260903287259. Capilla, A., Melcón, M., Kessel, D., Calderón, R., Pazo-Álvarez, P., Carretié, L., 2016. Retinotopic mapping of visual event-related potentials. Biological Psychology 118, 114–125. doi:10.1016/j.biopsycho.2016.05.009. Csanády, B., 2026. Analysis of Transition-Related Potentials in EEG During Film Viewing. URL: https://github.com/aielte-research/EEG_cuts. Csanády, B., Volosin, M., Zipernovszky, O.P., Kovács, A.B., Vedres, P., 2026. Continuous EEG Recordings During Coherent and Scene-Scrambled Film Viewing. doi:10.17632/3kd62t4dt3.1. Dini, H., Simonetti, A., Bruni, L.E., 2023. Exploring the Neural Processes be- hind Narrative Engagement: An EEG Study. eNeuro 10, ENEURO.0484– 22.2023. doi:10.1523/ENEURO.0484-22.2023. Dmochowski, J.P., Sajda, P., Dias, J., Parra, L.C., 2012. Correlated Com- ponents of Ongoing EEG Point to Emotionally Laden Attention – A Pos- sible Marker of Engagement? Frontiers in Human Neuroscience 6, 112. doi:10.3389/fnhum.2012.00112. Duncan-Johnson, C.C., Donchin, E., 1977. On Quantifying Surprise: The Variation of Event-Related Potentials With Subjective Probability. Psy- chophysiology 14, 456–467. doi:10.1111/j.1469-8986.1977.tb01312.x. Farinella, L., 2013. The Art of Love. Short film. URL: https://w.imdb. com/title/t26250136/. Francuz, P., Zabielska-Mendyk, E., 2013. Does the Brain Differentiate Be- tween Related and Unrelated Cuts When Processing Audiovisual Mes- sages? An ERP Study. Media Psychology 16, 461–475. doi:10.1080/ 15213269.2013.831394. 26 Ganis, G., Kutas, M., Sereno, M.I., 1996. The Search for “Common Sense”: An Electrophysiological Study of the Comprehension of Words and Pictures in Reading. Journal of Cognitive Neuroscience 8, 89–106. doi:10.1162/jocn.1996.8.2.89. Gramfort, A., Luessi, M., Larson, E., Engemann, D.A., Strohmeier, D., Brod- beck, C., Goj, R., Jas, M., Brooks, T., Parkkonen, L., Hämäläinen, M.S., 2013. MEG and EEG data analysis with MNE-Python. Frontiers in Neu- roscience 7, 1–13. doi:10.3389/fnins.2013.00267. Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., Malach, R., 2004. Intersubject Synchronization of Cortical Activity During Natural Vision. Science 303, 1634–1640. doi:10.1126/science.1089506. Heimann, K.S., Uithol, S., Calbi, M., Umiltà, M.A., Guerra, M., Gallese, V., 2017. “Cuts in Action”: A High-Density EEG Study Investigating the Neural Correlates of Different Editing Techniques in Film. Cognitive Science 41, 1555–1588. doi:10.1111/cogs.12439. Kappenman, E.S., Luck, S.J., 2016. Best Practices for Event-Related Po- tential Research in Clinical Populations. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging 1, 110–115. doi:10.1016/j.bpsc.2015. 11.007. Kaushik, P., Moye, A., van Vugt, M., Roy, P.P., 2022. Decoding the cog- nitive states of attention and distraction in a real-life setting using EEG. Scientific Reports 12, 20649. doi:10.1038/s41598-022-24417-w. Ki, J.J., Kelly, S.P., Parra, L.C., 2016. Attention strongly modulates re- liability of neural responses to naturalistic narrative stimuli. Journal of Neuroscience 36, 3092–3101. doi:10.1523/JNEUROSCI.2942-15.2016. Kleiner, M., Brainard, D., Pelli, D., 2007. What’s new in Psychtoolbox-3? Perception 36, 14. doi:10.1177/03010066070360S101. Kurby, C.A., Zacks, J.M., 2008. Segmentation in the perception and memory of events. Trends in Cognitive Sciences 12, 72–79. doi:10.1016/j.tics. 2007.11.004. 27 Lau-Zhu, A., Lau, M.P.H., McLoughlin, G., 2019. Mobile EEG in research on neurodevelopmental disorders: Opportunities and challenges. Devel- opmental Cognitive Neuroscience 36, 100635. doi:10.1016/j.dcn.2019. 100635. Lawhern, V.J., Solon, A.J., Waytowich, N.R., Gordon, S.M., Hung, C.P., Lance, B.J., 2018. EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces. Journal of Neural Engineering 15, 056013. doi:10.1088/1741-2552/aace8c. Luck, S.J., 2014. An Introduction to the Event-Related Potential Technique. 2 ed., MIT Press, Cambridge, MA. Luck, S.J., Hillyard, S.A., 1994. Electrophysiological correlates of feature analysis during visual search. Psychophysiology 31, 291–308. doi:10.1111/ j.1469-8986.1994.tb02218.x. Magliano, J.P., Zacks, J.M., 2011. The Impact of Continuity Editing in Narrative Film on Event Segmentation. Cognitive Science 35, 1489–1517. doi:10.1111/j.1551-6709.2011.01202.x. Maris, E., Oostenveld, R., 2007. Nonparametric statistical testing of EEG- and MEG-data. Journal of Neuroscience Methods 164, 177–190. doi:10. 1016/j.jneumeth.2007.03.024. Mars, R.B., Debener, S., Gladwin, T.E., Harrison, L.M., Haggard, P., Roth- well, J.C., Bestmann, S., 2008. Trial-by-trial fluctuations in the event- related electroencephalogram reflect dynamic changes in the degree of sur- prise. Journal of Neuroscience 28, 12539–12545. doi:10.1523/JNEUROSCI. 2925-08.2008. Maruyama, Y., Aoyama, K., Fukayama, O., 2026. Modulation of feedback- related negativity by objective and subjective response correctness. Scien- tific Reports 16, 13574. doi:10.1038/s41598-026-43451-6. Matran-Fernandez, A., Poli, R., 2015. Event-Related Potentials induced by cuts in feature movies and their exploitation for understanding cut efficacy, in: 2015 7th International IEEE/EMBS Conference on Neural Engineering (NER), IEEE. p. 74–77. doi:10.1109/NER.2015.7146563. 28 Nastase, S.A., Gazzola, V., Hasson, U., Keysers, C., 2019. Measuring shared responses across subjects using intersubject correlation. Social Cognitive and Affective Neuroscience 14, 667–685. doi:10.1093/scan/nsz037. Nigam, A., Hoffman, J.E., Simons, R.F., 1992. N400 to semantically anoma- lous pictures and words. Journal of Cognitive Neuroscience 4, 15–22. doi:10.1162/jocn.1992.4.1.15. Nunez, P.L., Srinivasan, R., 2006. Electric Fields of the Brain: The neuro- physics of EEG. 2 ed., Oxford University Press, New York. doi:10.1093/ acprof:oso/9780195050387.001.0001. Nuwer, M.R., Comi, G., Emerson, R., Fuglsang-Frederiksen, A., Guérit, J.M., Hinrichs, H., Ikeda, A., Luccas, F.J.C., Rappelsburger, P., 1998. IFCN standards for digital recording of clinical EEG. Electroen- cephalography and Clinical Neurophysiology 106, 259–261. doi:10.1016/ S0013-4694(97)00106-5. Papp-Zipernovszky, O., Deák-Kovács, T., Bálint, K.É., Kovács, A.B., 2024. A Narratív Bevonódás Skála magyar nyelvű adaptációja. Alkalmazott Pszi- chológia 26, 29–50. doi:10.17627/ALKPSZICH.2024.3.29. Picton, T.W., Bentin, S., Berg, P., Donchin, E., Hillyard, S.A., Johnson, R., Miller, G.A., Ritter, W., Ruchkin, D., Rugg, M.D., Taylor, M.J., 2000. Guidelines for using human event-related potentials to study cognition: Recording standards and publication criteria. Psychophysiology 37, 127– 152. doi:10.1111/1469-8986.3720127. Polich, J., 2003. Theoretical Overview of P3a and P3b, in: Detection of Change: Event-Related Potential and fMRI Findings. Springer US, Boston, MA, p. 83–98. doi:10.1007/978-1-4615-0294-4_5. Rényi, A., 1961. On Measures of Entropy and Information, in: Proceed- ings of the Fourth Berkeley Symposium on Mathematical Statistics and Probability, University of California Press, Berkeley, CA. p. 547–561. Roy, Y., Banville, H., Albuquerque, I., Gramfort, A., Falk, T.H., Faubert, J., 2019. Deep learning-based electroencephalography analysis: a system- atic review. Journal of Neural Engineering 16, 051001. doi:10.1088/ 1741-2552/ab260c. 29 Sanz-Aznar, J., Bruni, L.E., Soto-Faraco, S., 2023. Cinematographic continu- ity edits across shot scales and camera angles: an ERP analysis. Frontiers in Neuroscience 17, 1173704. doi:10.3389/fnins.2023.1173704. Schirrmeister, R.T., Springenberg, J.T., Fiederer, L.D.J., Glasstetter, M., Eggensperger, K., Tangermann, M., Hutter, F., Burgard, W., Ball, T., 2017. Deep learning with convolutional neural networks for EEG decoding and visualization. Human Brain Mapping 38, 5391–5420. doi:10.1002/ hbm.23730. Signoret, C., Gaudrain, E., Perrin, F., 2013. Similarities in the neural signature for the processing of behaviorally categorized and uncatego- rized speech sounds. European Journal of Neuroscience 37, 777–785. doi:10.1111/ejn.12097. Silva, M., Baldassano, C., Fuentemilla, L., 2019. Rapid memory reactiva- tion at movie event boundaries promotes episodic encoding. Journal of Neuroscience 39, 8538–8548. doi:10.1523/JNEUROSCI.0360-19.2019. Sitnikova, T., Holcomb, P.J., Kiyonaga, K.A., Kuperberg, G.R., 2008. Two neurocognitive mechanisms of semantic integration during the comprehen- sion of visual real-world events. Journal of Cognitive Neuroscience 20, 2037–2057. doi:10.1162/jocn.2008.20143. Sitnikova, T., Kuperberg, G., Holcomb, P.J., 2003. Semantic integration in videos of real-world events: An electrophysiological investigation. Psy- chophysiology 40, 160–164. doi:10.1111/1469-8986.00016. Smith, T.J., 2012. The Attentional Theory of Cinematic Continuity. Projec- tions 6, 1–27. doi:10.3167/proj.2012.060102. Sonkusare, S., Breakspear, M., Guo, C., 2019. Naturalistic Stimuli in Neu- roscience: Critically Acclaimed. Trends in Cognitive Sciences 23, 699–714. doi:10.1016/j.tics.2019.05.004. Stojić, S., Nadasdy, Z., 2024. Event as the central construal of psychological time in humans. Frontiers in Psychology 15. doi:10.3389/fpsyg.2024. 1402903. Supratak, A., Dong, H., Wu, C., Guo, Y., 2017. DeepSleepNet: a Model for Automatic Sleep Stage Scoring based on Raw Single-Channel EEG. 30 IEEE Transactions on Neural Systems and Rehabilitation Engineering 25, 1998–2008. doi:10.1109/TNSRE.2017.2721116. Wang, M.Y., Luan, P., Zhang, J., Xiang, Y.T., Niu, H., Yuan, Z., 2018. Concurrent mapping of brain activation from multiple subjects during so- cial interaction by hyperscanning: a mini-review. Quantitative Imaging in Medicine and Surgery 8, 819–837. doi:10.21037/qims.2018.09.07. Wiles, E., 2016. City Lights. Short film. URL: https://w.imdb.com/ title/t4660164/. World Medical Association, 2025. World Medical Association Declaration of Helsinki: Ethical Principles for Medical Research Involving Human Par- ticipants. JAMA 333, 71–74. doi:10.1001/jama.2024.21972. Zacks, J.M., Speer, N.K., Swallow, K.M., Braver, T.S., Reynolds, J.R., 2007. Event Perception: A Mind–Brain Perspective. Psychological Bulletin 133, 273–293. doi:10.1037/0033-2909.133.2.273. Zacks, J.M., Speer, N.K., Swallow, K.M., Maley, C.J., 2010. The Brain’s Cutting-Room Floor: Segmentation of Narrative Cinema. Frontiers in Human Neuroscience 4, 168. doi:10.3389/fnhum.2010.00168. 31 Appendix A. Ablation Studies To gain insight into which aspects of the cut-related EEG waveforms the DNN cut detector is sensitive to, we compared it with alternative ba- sic cut-detection approaches informed by the observed TRP waveforms and the previous literature. These detectors were designed to capture different properties of the EEG response expected around cuts. They served as in- terpretable reference points for the DL approach and helped illustrate why accurate, temporally precise event detection is important for TRP analysis. The TRP waveforms derived from the DNN model’s false-positive and false-negative detections revealed the following patterns (Figure A.8). First, anterior–posterior separation between 400 and 800 ms appeared significantly less prominent in both the false positive and false negative TRPs, than in the TRPs generated from the original (gold) cuts. Second, the false positive TRP lacks the distinct components such as P2 in the posterior average, while the false negative TRP still resembles the expected TRP waveform. Regarding the anterior–posterior divergence, when we shift the mean-correction interval from 0–200 ms to 100–300 ms, the false positive TRP shows stronger anterior– posterior separation. 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5 Amplitude FP anterior FP posterior Gold anterior Gold posterior (a) False Positives 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5 Amplitude FN anterior FN posterior Gold anterior Gold posterior (b) False Negatives Figure A.8: Anterior and posterior TRPs drawn from false positive and false negative triggers produced by the DNN model in the film-generalization setting. Signals averaged over all films. These differences between the false positive and false negative TRPs may arise because false positives are model-detected events and may therefore be less precisely time-locked, whereas false negatives are a subset of the true cut annotations. Nevertheless, these patterns motivate two basic cut-detection 32 strategies. First, if the DNN model is less likely to detect cuts when the anterior–posterior separation is less prominent, whereas its false positives still display such a separation, then a detector based on anterior–posterior divergence may provide a useful baseline for identifying film events that trig- ger cognitive updating. In other words, by such a detector we asked to what extent the DNN model’s performance could be explained by simply detect- ing anterior–posterior separation. Second, if the model misses cuts that still express ERP-like features, while producing false positives that do not, then a detector based on faithful TRP-template matching is worth investigating as a potential alternative. We defined the anterior–posterior (A–P) divergence detector as follows. Electrodes were grouped into anterior 1 and posterior 2 regions, and the signal within each region was averaged across channels. The metric was computed within normalized (1 s) windows as the signed root mean square difference between the anterior and posterior regional averages. This score takes high values during periods in which the two regions diverge with the polarity pat- tern expected for cut-related responses. We then applied peak detection to the resulting score signal. Because the peaks corresponding to the strongest anterior–posterior separation occurred at a consistent latency after cut onset, we corrected the detected peak times by shifting them backward by a fixed temporal offset. The second basic approach, the TRP-fit detector was calculated by fitting the cut-locked TRP template on the continuous EEG signal. We used cross- validation generalizing on films: the TRP template calculated from Art was fitted to the EEG signal of City, and vice versa. As in the A–P divergence cut-detector, the signals were first averaged within the anterior and posterior electrode groups. Separate TRP templates were then fitted to the anterior and posterior regional averages. The resulting fit score was evaluated over time on the test signal. Cut predictions were then obtained from this metric by peak-detection, as in the A–P divergence detector. To make the compar- ison less sensitive to slow drifts, the matching was performed on the discrete difference of the signals rather than on the raw waveform. 1 Fp1, Fp2, AF3, AF4, F7, F3, Fz, F4, F8, FC5, FC1, FC2, FC6, C3, C4. 2 T7, T8, CP5, CP1, CP2, CP6, P7, P3, Pz, P4, P8, PO3, PO4, O1, Oz, O2. 33 Appendix A.1. Correlation-Based Cut-Detectors The correlation-based cut detectors were motivated by the assumption that salient events such as cinematic cuts, which may represent event bound- aries in the film’s narrative (Zacks et al., 2010; Magliano and Zacks, 2011), evoke synchronized responses across viewers (Hasson et al., 2004; Dmochowski et al., 2012; Ki et al., 2016; Nastase et al., 2019). Under this assumption, cut-related time periods should be accompanied by increased similarity be- tween the EEG signals of different subjects. The detectors therefore measure inter-subject similarity over time and use this signal to identify candidate cut locations. We compared multiple variants of the correlation-based cut detection. • First, we measured the average inter-subject Pearson correlation within 0.5 s rolling windows, applied peak detection to the resulting time series, and shifted the peaks to the left by a constant amount, similar to the A–P divergence cut-detector. Additionally, we trained the same peak-detection DNN architecture as dis- cussed in Section 3.2, but instead of training on averaged EEG signals, we trained on correlation-derived signals: • We trained one version on a 32-channel signal in which each channel contained the inter-subject Pearson correlation of the corresponding EEG channel within a 1 s moving window. • Moreover, we trained a detector on a two-channel signal in which the two channels corresponded to the average correlations within the ante- rior and posterior electrode groups. Appendix A.2. Ablation Results We compared the basic cut detectors to the main DNN cut detector. All detectors were evaluated in the film-generalization cross validation scenario, meaning that where applicable, training was performed on both versions of one film and evaluation on both versions of the other film. As a reference, we also report a no-skill baseline, which provides an upper bound for a random detector constrained by the same minimum inter-peak distance. Because the minimum inter-peak distance was fixed to 0.625 s, and the acceptance window also spanned 0.625 s in total, this baseline distributed detections uniformly every 0.625 s throughout the recording. This construction achieves perfect 34 ModelF1 TP FP FN Avg. distance DNN0.863 139 16 2840 ms Correlation DNN 0.610 111 86 5690 ms Correlation DNN AP0.465 91 133 76132 ms Correlation0.472 94 137 73151 ms A-P divergence 0.442 74 94 93147 ms TRP-fit0.413 99 214 68153 ms No Skill0.278 167 868 0156 ms Table A.7: Evaluation metrics on Art coherent by model. ModelF1 TP FP FN Avg. distance DNN0.891 151 14 2323 ms Correlation DNN0.707 116 37 5974 ms Correlation DNN AP0.555 93 67 82107 ms Correlation0.559 109 106 66139 ms A-P divergence 0.519 87 73 88155 ms TRP-fit0.481 117 195 58150 ms No Skill0.355 175 636 0156 ms Table A.8: Evaluation metrics on City coherent by model. ModelF1 TP FP FN Avg. distance DNN0.953 161 12 429 ms Correlation DNN0.712 142 92 2360 ms Correlation DNN AP0.509 110 157 55122 ms Correlation0.490 106 162 59144 ms A-P divergence0.516 89 91 76164 ms TRP-fit 0.397 104 255 61163 ms No Skill0.277 165 860 0156 ms Table A.9: Evaluation metrics on Art incoherent by model. ModelF1 TP FP FN Avg. distance DNN0.868 145 11 3322 ms Correlation DNN0.710 114 29 6466 ms Correlation DNN AP0.550 90 59 88110 ms Correlation0.529 106 117 72133 ms A-P divergence0.460 74 70 104154 ms TRP-fit 0.508 113 154 65147 ms No Skill0.381 178 575 0156 ms Table A.10: Evaluation metrics on City incoherent by model. 35 recall and yields maximal F1 score among no-skill detectors whose successive detections must be at least 0.625 s apart. Its purpose was to define a simple reference level against which the data-driven methods could be compared. Tables A.7 to A.10 summarize the model performance across films and narrative versions. Among the models compared, the main DNN model de- tected cuts most accurately, achieving the highest F1 scores and the fewest errors across all conditions. The average distance (temporal error) of the de- tected true positive cuts by the DNN model was within one frame (40 ms at 25 frames per second). The simple peak-detection-based models performed better than random (no-skill baseline), however, their average temporal er- rors were close to 156.25 ms, the value expected if detections were distributed within the acceptance windows uniformly at random. 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5 Amplitude TP anterior TP posterior Gold anterior Gold posterior (a) All 32 channel correlations. 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5 Amplitude TP anterior TP posterior Gold anterior Gold posterior (b) A–P group avg. correlations. Figure A.9: Anterior and posterior TRPs drawn from the true positive triggers by the DNN models trained on channel correlations. Models generalized on the film, signals averaged over all films. Frame-accurate cut detection is crucial for reliably reproducing TRP waveforms. Figure A.9 shows the TRPs generated from true positive detec- tions of the DNN models trained on correlation data. The TRP generated from true positive detections of the 32-channel correlation DNN model is already heavily attenuated at an average temporal error of 71.67 ms, which is less than two frames (Figure A.9a). Moreover, at an average temporal er- ror of less than three frames (117.95 ms), the anterior–posterior correlation- based DNN model already produces a true positive TRP waveform with no noticeable ERP-like features apart from basic anterior–posterior divergence (Figure A.9b). 36 Appendix A.3. Error Analysis FilmArtCity NarrativeCoherent IncoherentCoherent Incoherent% True positive139161151145 False positive16121411 Transient5513756.6% Significant event961234.0% Undetermined21029.4% False negative2842333 Weak transient207819.3% Sub-window 93142052.3% Undetermined 1712528.4% Table A.11: DNN cut detection outcomes by EEG recording in the film-generalization setting. FilmArtCity NarrativeCoherent IncoherentCoherent Incoherent% True positive111142116114 False positive85923729 Transient193214930.5% Significant event 513981044.4% Undetermined1521151025.1% False negative55235964 Weak transient4091312.9% Sub-window1510272839.8% Undetermined 3613232347.3% Table A.12: Cut detection outcomes by EEG recording for the 32-channel correlation DNN model in the film-generalization setting. Although the correlation-based cut detectors performed worse than the main DNN model and operated with lower temporal precision, we hypothe- sized that they might identify more narratively relevant false positives. Ta- bles A.11 and A.12 compare the manual error-analysis results for the main DNN model and the 32-channel correlation-based DNN model. As antici- pated, the correlation-based model identified a larger number of narratively relevant events. However, these events cannot be directly used in the stan- dard TRP framework, because they do not exhibit strong ERP-like features 37 when averaged around the detected peaks (Figure A.10). This suggests that correlation-based alternatives may be useful for discovering potentially inter- esting moments in the stimulus, but not directly for recovering meaningful event-related waveforms. Leveraging the additional potential of these sig- nificant false positive detections, would, therefore, require analysis methods beyond the scope of this study. 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5 Amplitude FP anterior FP posterior Gold anterior Gold posterior (a) All 32 channel correlations. 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 0.5 Amplitude FP anterior FP posterior Gold anterior Gold posterior (b) A–P group avg. correlations. Figure A.10: Anterior and posterior TRPs drawn from the false positive triggers by the DNN models trained on channel correlations. Models generalized on the film, signals averaged over all films. Appendix B. Additional Figures 0.20.00.20.40.60.81.01.2 Time (s) 0.4 0.2 0.0 0.2 0.4 0.6 Amplitude Posterior average Anterior average Fp AF F FC C CP P PO O Electrode (a) Art (incoherent). 0.20.00.20.40.60.81.01.2 Time (s) 0.4 0.2 0.0 0.2 0.4 0.6 Amplitude Posterior average Anterior average Fp AF F FC C CP P PO O Electrode (b) City (incoherent). Figure B.11: Butterfly plots of the average anterior and posterior TRP signals aligned to all cuts in the incoherent versions of the films. 38 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 Amplitude Posterior average Anterior average Fp AF F FC C CP P PO O Electrode (a) Art (coherent− incoherent). 0.20.00.20.40.60.81.01.2 Time (s) 0.3 0.2 0.1 0.0 0.1 0.2 0.3 0.4 Amplitude Posterior average Anterior average Fp AF F FC C CP P PO O Electrode (b) City (coherent− incoherent). Figure B.12: Butterfly plots of the coherent− incoherent average TRP signals. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 0.3 0.2 0.1 0.0 0.1 0.2 0.3 (a) Art eigenvector 1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 0.3 0.2 0.1 0.0 0.1 0.2 0.3 (b) Art eigenvector 2. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 0.3 0.2 0.1 0.0 0.1 0.2 0.3 (c) City eigenvector 1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 0.2 0.1 0.0 0.1 0.2 (d) City eigenvector 2. 1.00.50.00.51.0 Principal Component 1 (72.5%) 1.0 0.5 0.0 0.5 1.0 Principal Component 2 (19.6%) Coherent Incoherent (e) Art principal components. 1.00.50.00.51.01.5 Principal Component 1 (78.1%) 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 1.00 Principal Component 2 (15.4%) Coherent Incoherent (f) City principal components. Figure B.13: Spatial PCA analysis of the scene-matched hand-annotated cut-TRPs. 39 0.20.40.60.81.01.2 Time (s) 0.06 0.04 0.02 0.00 0.02 0.04 0.06 0.08 Amplitude PC1 (90.1%) PC2 (6.7%) (a) Art principal eigenvectors. 0.20.40.60.81.01.2 Time (s) 0.06 0.04 0.02 0.00 0.02 0.04 0.06 0.08 0.10 Amplitude PC1 (92.3%) PC2 (4.0%) (b) City principal eigenvectors. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (c) Art coherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (d) Art incoherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (e) City coherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 6 4 2 0 2 4 6 (f) City incoherent PC1. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (g) Art coherent PC2. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (h) Art incoherent PC2. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (i) City coherent PC2. AF3AF4 C3C4 CP1CP2 CP5CP6 Cz F3F4 F7F8 FC1FC2 FC5FC6 Fp1Fp2 Fz O1O2 Oz P3P4 P7P8 PO3PO4 Pz T7T8 2 1 0 1 2 (j) City incoherent PC2. Figure B.14: Temporal PCA analysis of the TRPs from the detected cuts. 40

Instant research discovery