Gamer in the scanner: Event-related analysis of fMRI activity during retro videogame play guided by automated annotations of game content.

Gamer in the scanner: Event-related analysis of fMRI activity during retro videogame play guided by automated annotations of game content - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Imaging Neurosci (Camb) . 2026 Jun 1;4:IMAG.a.1256. doi: 10.1162/IMAG.a.1256 Gamer in the scanner: Event-related analysis of fMRI activity during retro videogame play guided by automated annotations of game content Yann Harel 1, 2, * , Basile Pinsard 2 , Julie Boyle 2 , Valentina Borghesani 3 , Maximilien Le Clei 2 , Paul-Henri Mignot 4 , Marie St-Laurent 2 , Karim Jerbi 1, 5, 6 , Lune Bellec 1, 2, 5 Author information Article notes Copyright and License information 1 Department of Psychology, University of Montréal, Montréal, Canada 2 Centre de Recherche de l’Institut Universitaire de Gériatrie de Montréal (CRIUGM), Montréal, Canada 3 Swiss National Centre of Competence in Research, University of Geneva, Geneva, Switzerland 4 IMT Atlantique, Brest, France 5 Mila – Quebec AI Institute, Montréal, Canada 6 UNIQUE – Union Neurosciences et Intelligence Artificielle – Québec, Québec, Canada ✉ *Corresponding Author: Yann Harel ( yann.harel@umontreal.ca , yharel109@gmail.com ) Received 2025 Jun 11; Revised 2026 Mar 20; Accepted 2026 Apr 24; Collection date 2026. © 2026 The Authors. Published under a Creative Commons Attribution 4.0 International (CC BY 4.0) license. This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International (CC BY 4.0) license, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. For a full description of the license, please visit https://creativecommons.org/licenses/by/4.0/ . PMC Copyright notice PMCID: PMC13228873  PMID: 42238756 Abstract In recent years, videogames have gathered interest in cognitive neuroscience for their potential to study cognition in dynamical and naturalistic contexts. Yet, the complexity of game environments often challenges traditional modeling approaches, and current annotation methods—typically manual or based on modified games—remain labor-intensive and limited in scope. Here, we introduce a flexible and scalable framework using the gym-retro Python library to emulate a classic action-platformer, Shinobi III: Return of the Ninja Master , and automatically annotate gameplay events directly from the game’s memory states. This setup enables the identification of both player actions (e.g., jumping, hitting) and feedback events (e.g., killing an enemy, being hit), without modifying the game. Four individuals played the videogame for a combined total of 32 h (>7 h each) while undergoing functional magnetic resonance imaging (fMRI). Resulting activation maps revealed distributed engagement of visual, motor, executive, and limbic systems, consistent with the cognitive demands of gameplay. Within-participant reproducibility of brain responses across sessions was robust across event types (r ≈ .25–.55), with some consistency observed even for rarer events like HealthLoss. Between-participant correlations were notably lower, reflecting participant-specific neural signatures. Multivoxel pattern analysis showed that brain responses to different in-game events were highly discriminable, with classification accuracy typically around or above 90%, though occasionally dropping to ~40% for less frequent events. These findings demonstrate that automated emulator-based annotations enable robust, interpretable, and scalable mapping of naturalistic cognitive processes using commercial videogames. Keywords: naturalistic neuroimaging, video games, automated annotation, fMRI, cognitive neuroscience 1. Introduction A growing trend is to study cognition in naturalistic settings, employing rich, multifaceted paradigms ( Herholz et al., 2022 ; Nastase et al., 2020 ) such as passive movie watching ( Eickhoff et al., 2020 ; Jääskeläinen et al., 2021 ; Vanderwal et al., 2019 ), free reading ( Hsu et al., 2019 ; Wehbe et al., 2014 ), simulated driving ( Choi et al., 2017 ; Schweizer et al., 2013 ) and videogame playing. Such naturalistic stimuli, however, present a complex temporal structure, more challenging to model than classical controlled laboratory tasks ( Huettel, 2012 ; Warbrick et al., 2009 ). Notably, cognitive events often overlap, nest within one another, and may lack a clear onset and duration. To capture the neural underpinnings of complex behavior in naturalistic conditions using a general linear model (GLM, Poline & Brett, 2012 ) in functional magnetic resonance imaging (fMRI), one thus needs to model the intricate temporal structure of stimuli with precise and detailed annotations. Since the foundational work of Hasson et al. (2004) , movie watching has emerged as a popular paradigm for naturalistic fMRI experiments. Many forms of movie content annotations have been proposed in the literature, ranging from human raters who identify features of interest manually, such as scenes and characters, to the deployment of machine learning automated tools ( Wallenius, 2010 ). In particular, the Neuroscout platform enables automated movie annotations that capture elements like the presence of faces and buildings, the frequency of words spoken, and other audio-visual features ( de la Vega et al., 2022 ). Movie watching tasks are characterized by a unidirectional flow where the stimulus remains constant irrespective of the participant’s response. In stark contrast, videogames introduce a bidirectional interplay ( Matusz et al., 2019 ) where the environment actively responds and adapts based on the participant’s actions and decisions ( Potter et al., 2014 ). This makes videogames a powerful tool to probe cognitive domains that passive naturalistic tasks cannot access, such as psychomotor learning, decision-making, and problem solving. However, the dynamic nature of videogames heightens the need for annotations that can accurately capture these nuanced interactions, and our work proposes a novel high-throughput framework to generate videogame event annotations and corresponding brain activation maps. Several studies analyzed fMRI data acquired during complex videogame play. Many of these have relied on human raters to annotate game sequences. Weber and colleagues ( Mathiak & Weber, 2006 ; Weber et al., 2006 ) have employed frame-by-frame handmade annotations to identify relevant variables during First Person Shooter (FPS) gameplay, such as passive/dead states, preparation/search phases, potentially dangerous encounters, and moments when the player character is under attack. Wang et al. (2018) recruited 70 players to play League of Legends in a 3T MRI, and had six experimented raters annotate 15-s segments assessing the quality of in-game maneuvers as either good, average, or poor plays. They then used general linear models (GLMs) and seed-based functional connectivity to uncover neural correlates of gameplay proficiency. Zhang et al. (2021) enlisted two human raters to annotate game events during replays of Counter-Strike: Source such as kills, deaths, instances of safe or unsafe exploration, engagements, use of flashbang grenades, and both distant and close engagements. While human videogame annotations can achieve granularity and depth, this approach is prohibitively labor-intensive and cannot easily be scaled to large datasets. Another approach to capturing in-game events is to incorporate modifications directly into game software to log events and annotate fMRI signal. For example, Lim et al. (2019) studied incidental learning using a modified game platform. Within this environment, players encountered distinctive “aliens”, each associated with a particular sound. These occurrences were marked as events within the game’s presentation software, ensuring that they could be readily tracked and analyzed. Similarly, Kätsyri et al. (2013a , 2013b ) used a modified version of the open-source game BZFlag ( Schoeneman & Riker, BZFlag ) to log each success or failure. These annotations were then used to compare active and vicarious play, and to study brain responses to human or artificial opponents using GLMs. Such modified game software approaches, while also labor-intensive and restricted to open-source or custom-made games, offer the advantage of tailoring the gaming experience to the specific needs and goals of the research study, ensuring that pertinent events are captured accurately and consistently. While previous research has explored neural patterns evoked during videogame play, most work employed a tedious and incomplete annotation process, resorting to manual annotations or modified game softwares. Our research introduces a new flexible approach that facilitates the annotation of videogame tasks performed in the scanner by integrating console emulators directly into the neuroimaging setup. This approach is made easily accessible via OpenAI’s gym-retro ( Nichol et al., 2018 ), a Python library designed to facilitate the development of artificial agents and accelerate machine learning experiments by offering a unified interface with the most common console emulators. The library thus provides access to a vast catalog of retro videogames (i.e., videogames released on older consoles), and offers versatility in the selection and control of game parameters. For the Courtois project on neuronal modelling (CNeuroMod; Boyle et al., 2020 ), a large-scale deep phenotyping dataset comprising a diverse set of tasks that includes several naturalistic paradigms, we developed a neuroimaging acquisition setup tailored to record brain activity during retro videogame sessions played on a console emulator. Participants each played over 7 h of Shinobi III: Return of the Ninja Master ( Shinobi III: RotNM ; Sega, 1993 ), a classic action platformer, under naturalistic play conditions while undergoing fMRI. We annotated play content with high timing precision using the game’s memory states, which captured diverse in-game events such as kills, health losses and player actions (button presses). We aimed to demonstrate that this high-throughput annotation framework can be used to generate appropriate events for GLM analyses, and more specifically demonstrate that (1) there are enough repetitions of different event types and, (2) different event types have sufficiently small correlations that it is possible to derive reliable and distinct maps of brain networks using a GLM and these annotations. Specifically, we produced individual brain activation maps linked to different types of automatically annotated events and assessed their quality by comparing these against statistical maps obtained from functional localizer tasks acquired in the same participants. With this approach, we evaluated the consistency of these activations between and within participants using spatial correlations measures. We also assessed the distinctiveness of these spatial neural networks with multivoxel pattern analysis (MVPA) classification. Our results indicate that annotations derived automatically from game emulators can greatly facilitate the identification of meaningful brain activity patterns that reflect various game events. 2. Methods 2.1. CNeuroMod participants The CNeuroMod project compiled a series of datasets involving six participants (sub-01 to sub-06) who engaged in various tasks over 6 years ( Boyle et al., 2020 ). Informed written consent was obtained from all participants, and the experimental protocol was approved by the “ Comité d’éthique de la recherche vieillissement-neuroimagerie ” of the CIUSSS du Centre-Sud-de-l’île-de-Montréal (CIUSSS-CSMTL). Participants were right-handed English speakers with normal or corrected-to-normal hearing and vision, aged 31 to 47 years at the start of data collection. 2.2. Datasets This study utilizes two CNeuroMod datasets: shinobi and hcptrt . Due to the scope and duration of the project, not all participants completed all tasks. For this study, we focused on the subset of participants who completed both the shinobi and hcptrt protocols. A more detailed description of these datasets can be found in the CNeuroMod technical documentation ( CNeuroMod, 2025a ). The shinobi dataset explores behavioural and brain correlates of videogame play and includes four participants (sub-01, sub-02, sub-04, and sub-06; two females, two males) who played Shinobi III: Return of the Ninja Master (RotNM; Sega, 1993 ) inside an MRI scanner. While not a central focus of this study, prior videogame experience may have influenced the results. Among the participants, sub-01 had prior experience with Shinobi III: RotNM , and three participants (sub-01, sub-02, and sub-04) reported regular videogame play. One participant (sub-06) reported no videogame experience. The hcptrt dataset includes a range of tasks designed by the Human Connectome Project ( Barch et al., 2013 ), repeated multiple times per participant to generate brain maps of various systems. For this study, we used hcptrt data from the same four participants who completed the shinobi acquisitions. The dataset encompasses seven distinct tasks, each administered up to 15 times per participant, resulting in approximately 10 h of fMRI data per participant ( Rastegarnia et al., 2023 ). Here, hcptrt serves as a benchmark for assessing the consistency of brain maps derived from established functional localizer paradigms and evaluating whether brain maps generated from videogame data align with those from conventional localizers. 2.3. Behavioral acquisitions 2.3.1. Videogame interface The videogame was played on a console emulator using OpenAI’s gym-retro library ( Nichol et al., 2018 ), a Python-based platform supporting emulators for over 10 retro consoles and thousands of games. Built on gym ( Brockman et al., 2016 ), a library designed for reinforcement learning, gym-retro integrates console emulators via the Libretro API ( Libretro, n.d. ). This setup allows both artificial and human agents to play retro games from saved states and provides access to the game’s rapid-access memory (RAM) through a Python API. In this experiment, Shinobi III: Return of the Ninja Master (RotNM) was played and recorded at a 60 Hz framerate. Because the game is fully deterministic (identical inputs produce identical results), only player inputs (button presses) were recorded, enabling the reconstruction of exact replays of gameplay sequences. These replays facilitated a posteriori RAM mining, mapping emulator memory addresses and retrieving value changes in identified variables. 2.3.2. Videogame task Shinobi III: RotNM ( Sega, 1993 ) is a classic action platformer released on the Sega Genesis console. The platformers genre was selected by the CNeuroMod because gym-retro was developed for a major reinforcement competition centered on this genre ( OpenAI, 2018 ). This game was further selected for its compatibility with gym-retro and its commercial success, ensuring engaging gameplay. As in other platformers, players navigate a linear two-dimensional environment filled with enemies and obstacles to reach a destination. For this study, three levels or portions of levels (1-0, 4-1, and 5-0; abbreviated as 1, 4, and 5) were selected for their gameplay similarity, despite varying difficulty. In all selected levels, players began on the far left and progressed to the far right using seven primary actions: move left, move right, crouch (move down), hang (move up), jump, attack, and power up. By combining buttons or executing timed sequences, players could perform additional actions such as dashing (forward, forward), wall-jumping (JUMP + LEFT/RIGHT against a wall), low-kicking (DOWN + HIT near an enemy), and jump-kicking (JUMP + DOWN + HIT near an enemy). Levels 1 and 5 concluded with a boss fight against a stronger enemy. Each attempt at completing a level, referred to as a “repetition,” began with 16 health points and 50 projectiles. Players needed to complete the level, including the boss fight, without dying. Bonuses scattered throughout the levels allowed players to restore health or replenish projectiles. A selection of game frames is shown in Figure 1 . Fig. 1. Open in a new tab Examples of game frames from the three selected levels. Shown are the first frame of each level (“Start,” left), an arbitrarily selected mid-level frame (“Mid-stage,” middle), and a final frame (“End,” right) depicting either a boss fight (levels 1 and 5) or the level’s exit (level 4). 2.4. Neuroimaging acquisitions 2.4.1. MRI sequences MRI data were collected using a 3T Siemens PrismaFit scanner with a 64-channel head-neck coil at the functional neuroimaging unit (UNF) of the Centre de Recherche de l’Institut Universitaire de Gériatrie de Montréal (CRIUGM) in Montréal. Additional acquisition details are available in the CNeuroMod technical documentation ( CNeuroMod, 2025c ). Anatomical data were acquired during separate dedicated sessions using a T1-weighted MPRAGE 3D sagittal sequence (duration 6:38 min, TR = 2.4 s, TE = 2.2 ms, flip angle = 8 deg, voxel size = 0.8 mm isotropic, R = 2 acceleration). These scans demonstrated high structural stability of individual brain anatomy throughout the CNeuroMod project ( Boudreau et al., 2025 ). Functional data were acquired using an accelerated simultaneous multi-slice gradient echo-planar imaging sequence from the Human Connectome Project (slice acceleration factor = 4, TR = 1.49 s, TE = 37 ms, flip angle = 52°, 2 mm isotropic spatial resolution, 60 slices, 96 x 96 acquisition matrix; Setsompop et al., 2012 ). 2.4.2. MRI acquisition set-up Visual stimuli were projected using an Epson Powerlite L615U onto a blank screen behind the MRI bore, viewed through a head coil-mounted mirror. Audio was delivered via S15 Sensimetrics headphone inserts. To minimize motion artifacts and improve fMRI signal quality, the CNeuroMod project employed individualized polystyrene head cases (CaseForge; Lynch et al., 2021 ). Physiological data were recorded concurrently with neuroimaging but are not within the scope of this study. 2.4.3. MRI-compatible videogame interface During neuroimaging sessions, the game was presented using a Python software stack ( CNeuroMod, 2025e ) that employed PsychoPy ( Peirce et al., 2019 ) to synchronize task onset with scanner trigger pulses, initiate the gym-retro library with predefined save states for specific levels, and record player inputs (button presses). The software also produced real-time outputs (sound and video frames) for participants and detected conditions for ending level repetitions (player death or level completion). Participants used a custom MRI-compatible videogame controller, designed with 3D-printed plastic and fiber optics, connected via USB to the stimulation computer. Developed by our team, this controller allowed high performance gameplay while effectively minimizing acquisition artefacts caused by scanner interference. To promote reproducible science, the design plans were publicly shared ( Harel et al., 2023 ). The controller replicated the layout of a SuperNES™ controller ( Inoue & Ashida, 1993 ), featuring two sets of four buttons accessible via the thumbs. For this task, the left buttons controlled character motion (RIGHT, LEFT, UP, DOWN), while the right buttons controlled actions (JUMP, HIT), mirroring the original console’s button layout. 2.4.4. MRI preprocessing Anatomical data were preprocessed using the sMRIPrep pipeline, with T1w and T2w scans from the first two anatomical sessions of all participants as input. These scans were coregistered and averaged to enhance structural stability. Functional MRI data were preprocessed using fMRIPrep 20.2 LTS ( Esteban et al., 2019 ), built on Nipype 1.6.1. Confounds time series for subsequent analyses were derived from the fMRIPrep pipeline output. We employed the “simple” denoising strategy, including high-pass filtering, regression of average signal within white matter and ventricles, and motion parameters, following the recommendations of Wang et al. (2024) for datasets with low motion levels. This approach preserved temporal continuity in the denoised time series. Detailed preprocessing steps are available on the CNeuroMod website ( CNeuroMod, 2025b ). For subsequent analysis, the confounds time series were added to the design matrix using the following Nilearn command: confounds = nilearn.interfaces.fmriprep.load_confounds( fmri_fname, strategy=('motion', 'high_pass', 'wm_csf'), motion='full', wm_csf='basic' ) 2.4.5. The shinobi dataset 2.4.5.1. At-home training All participants were given extensive time to practice during the weeks or months preceding neuroimaging. By the start of the experiments, all participants were familiar with the game, as indicated by clear progress on performance metrics. Corresponding behavioral practice data are publicly available as the shinobi_training dataset ( CNeuroMod, 2025f ). A table describing breaking down the amount of training time per participant and per level is presented in Appendix Table A1 , and a comparison of the performance between at-home training and in-scanner sessions is presented in Appendix Figure A1 . 2.4.5.2. In-scanner play sessions Participants played Shinobi III: RotNM to the best of their ability while undergoing MRI scanning. The shinobi dataset also served as a pilot for the gym-retro interface in preparation of future videogames acquisitions via our platform. During each fMRI run, participants sequentially played the three selected levels (1, 4, and 5) for a minimum of 10 consecutive minutes. Runs were manually stopped at the end of the ongoing repetition after 10 min, regardless of the level reached. If a participant died or cleared a level, they immediately proceeded to the next one. Each fMRI session included 4–5 runs, with total playtime approaching 7 h for sub-06 and 8 h for all other participants (see Table 1 ; a more comprehensive table of descriptive metrics is provided in Supplementary Materials as SM1). Each participant completed 12 sessions, except sub-04, who completed 11. The behavioural and neuroimaging data are publicly available as the shinobi dataset ( CNeuroMod, 2025d ). Session 009 of sub-06 was ignored in subsequent analysis because of a software issue that prevented the recording of behavioral data. Table 1. Descriptive table of the shinobi dataset. sub-01 sub-02 sub-04 sub-06 Total level-1 N 87 77 67 46 277 cleared 82 74 66 45 267 duration (h:m:s) 03:51:53 03:31:10 03:17:55 03:29:24 14:10:22 level-4 N 61 57 56 42 216 cleared 37 26 25 23 111 duration (h:m:s) 01:54:00 02:09:48 02:06:00 02:44:22 08:54:11 level-5 N 53 55 53 12 173 cleared 53 47 47 11 158 duration (h:m:s) 02:05:45 02:37:35 02:39:48 00:45:31 08:08:38 Total sessions 12 12 11 11 46 runs 53 55 55 47 210 N 201 189 176 100 666 cleared 172 147 138 79 536 duration (h:m:s) 07:51:38 08:18:33 08:03:42 06:59:18 31:13:11 Open in a new tab 2.4.6. Quality control We extracted the following QC measures from fMRIprep ( Esteban et al., 2019 ) and the MRIQC BIDS app ( Provins et al., 2023 ): Framewise displacement (FD) summary measures: the mean and the max value of each run. Max value will show the worst instantaneous motion in each run. Motion extent or excursion, as the max difference between max and min values of each of the 6 motion parameters. Temporal signal to noise ratio (TSNR) averaged in the brain mask, as extracted by MRIQC. Crowdsourced MRIQC measures were pulled from the MRIQC API, filtering for approximately matching acquisition parameters in terms of magnetic field strength, TR, voxel size, and multiband acceleration factor, and with minimum duration of 100 TRs. A figure showing the run-median FD values is presented in SM3 Supplementary Figure S1 . 2.5. Annotation of task events 2.5.1. Memory (RAM) mining The videogame console’s memory stores temporary variables that encode the current game state. The gym-retro library includes an integration software for researchers to explore and map memory addresses in order to extract meaningful game variables. This software allows users to play the game, to log performance and progress metrics and to track value changes in the memory. Replay files are exported in a gym-retro -specific format called bk2. Replay bk2 files encoded all button presses and could be used with gym-retro to regenerate gameplay videos and produce a complete mapping of memory states for each game step (see Fig. 2 ). The process of mapping game variables to memory addresses is a feat of reverse engineering. The gym-retro library provides a graphical interface that allows the reading, lookup and modification of RAM values as the game is running within the emulator. Through monitoring memory changes as the game unfolds while deliberately manipulating the content of target variables (e.g., dying on purpose, moving left or right etc.), we deduced the location of variables encoding the score , health , lives and player horizontal position . Fig. 2. Open in a new tab Exploration of variables in the memory during replay. Capture of a selected game frame (910; left), and time course (in frames; right) for normalized variables score (blue), health (yellow), lives (green) and player horizontal position (“X_player”, red) for one level repetition. At the selected frame (black vertical line), the player has been hit by an enemy and the health variable (yellow) shows a concomitant decrement. 2.5.2. Game-related events We first produced player action events that marked the timing of button presses and releases. While we recorded all possible actions, current analyses ignore the POWERUP , START , SELECT and move UP actions that were either too sparse or didn’t meaningfully impact the game. Instead, analyses considered the (move) RIGHT , LEFT and DOWN actions, as well as the JUMP and HIT actions. We then generated two types of game annotation events from memory variables, designed to parallel rewarding and punishing events in reinforcement learning. This second type of events does not map one-to-one to button-presses, and constitutes the key innovation of our memory mining approach, expanding event-based analysis to game outcomes without relying on manual annotations. We thus annotated the following events: Kill corresponds to moments when the player kills an enemy, and carries positive affective valence. Kill events are detected with a sudden score increase of 200, 300, 400 or 500 points, depending on enemy type. HealthLoss corresponds to moments when the player loses health when hit by an enemy or trap, which carries a negative affective valence. This event is indicated by a negative change of the health variable. 2.5.3. Sensory and motor features Psychophysical properties of the audiovisual stimulus were computed from the game replay files and added to the set of modeled annotations. Luminance was computed as the mean grayscale pixel intensity of each video frame. Optical flow was estimated between consecutive frames using the Farneback dense optical flow algorithm (OpenCV; pyr_scale = 0.5, levels = 3, winsize = 15), and summarized as the mean flow magnitude per frame. Audio envelope was computed as the root-mean-square (RMS) of the audio signal within each video frame window. These three features, originally sampled at the video frame rate (60 Hz), were downsampled to match the fMRI acquisition rate by averaging within TR-length bins (1.49 s). Button press count was constructed by counting the total number of controller button presses (all actions, including those not modeled as individual event regressors) occurring within each TR, to capture the overall density of motor actions. All four continuous regressors were then convolved with the SPM canonical HRF before being entered into the GLM as task regressors. Like game-event annotations, these low-level features were fitted one at a time in separate single-regressor GLMs. 2.5.4. Data structure for events Task-related events were structured according to guidelines defined by the BIDS (Brain Imaging Data Structure) standard ( Gorgolewski et al., 2016 ). For each fMRI run, an event file was generated in .tsv format, named *_desc-annotated_events.tsv. Rows represented different types of events, while columns described features such as occurrence type, onset time, and duration. Specifically, repetition events marked the timing of each game level (repetition) within an fMRI run. These events included additional columns for the level played and the path to a corresponding bk2 file. Annotation events included Kill and HealthLoss, which were given an arbitrary duration of 0 s. Action events included all the button presses, with durations calculated as the interval between button press and release. 2.6. Statistical analyses 2.6.1. Regressors A general linear model (GLM) ( Poline & Brett, 2012 ) was used to model the effects of seven game-related events on the BOLD signal after full preprocessing (see Fig. 3A for an overview). Seven game events were considered: RIGHT, LEFT, DOWN, HIT and JUMP player actions, along with Kill and HealthLoss annotations. All analyses were performed using Nilearn ( Abraham et al., 2014 ). Fig. 3. Open in a new tab Graphical summary of the analysis pipeline applied to MRI and behavioral data. (A) Describes the preprocessing steps. (B) Shows an example of a GLM design matrix built by combining the preprocessed data. (C) Describes the analysis pipeline used to assess the quality of brain activation maps derived from game events. Confounds exported by fMRIPrep were included as additional columns to the design matrix. To further account for potential motion and task-unrelated processes during resting periods between gameplay repetitions, the frames corresponding with these periods were censored using an additional set of one-hot regressors to the design matrix. 2.6.2. First-level (session) general linear model We tested the main effect of each annotation regressor separately, by fitting a series of single-regressor GLMs. For each model, the annotation time series was convolved with the SPM canonical HRF, and fitted to the smoothed BOLD data (FWHM = 5 mm). fMRIPrep confound regressors (6 motion parameters, white matter, CSF, and global signal) along with high-pass filtering regressors were included in all models. See Figure 3B for an example of a full design matrix for one run. Session-level (first-level) models combined runs from the same session using a fixed-effects approach with an AR(1) noise model. F-tests were used to assess the main effect of each regressor, yielding z-score maps and effect estimates (beta values). The resulting z-maps were corrected for multiple comparisons using voxel-wise FDR (q < 0.05), with an additional minimum cluster extent of 10 voxels. 2.6.3. Second-level (participant) general linear model Second-level models were computed for each participant by fitting an intercept-only design (equivalent to a one-sample F-test) over all of their uncorrected session-level z-score maps, using nilearn’s SecondLevelModel with no additional smoothing. This resulted in one participant-level map of standardized activations per annotation. The participant-level z-score maps were corrected for multiple comparisons using voxel-wise FDR (q < 0.05) with a minimum cluster extent of 10 voxels. The corrected maps were then clustered (voxel threshold z(F) > 3, minimum cluster size > 5 voxels) using atlasreader 0.3.2 ( Notter et al., 2019 ) for AAL projection, in order to generate a summary of the AAL regions where significant clusters were found (see SM3 Supplementary Table S1 ). 2.6.4. Brain activation maps derived from HCP test-retest dataset To assess the quality of brain activation maps derived from game events, we compared these with activation maps obtained from tasks targeting diverse cognitive domains. Each task engaged a distinct subset of cognitive processes, structured through experimental conditions designed to facilitate precise identification of task-related neural activity. Brain activation maps for each task were generated using a GLM at the run level, modeling the main effect of single experimental conditions. These tasks and contrasts are described in detail elsewhere ( Barch et al., 2013 ; Pinho et al., 2018 ; Rastegarnia et al., 2023 ) and summarized below: Gambling: participants guessed whether a hidden number between 1 and 9 was above or below 5 to earn a symbolic monetary reward, simulating reward-based decision-making. Stimuli were biased toward either reward or punishment, with trials presented in blocks. Participants were unaware of each block’s nature (reward or punishment). Effects: reward, punishment. Motor: participants performed specific movements (left/right finger tapping, left/right toe squeeze, tongue movement) in response to visual cues to identify motor control regions. Effects: left_hand, right_hand, left_foot, right_foot, tongue, cue. Language Processing: this task assessed language and numerical reasoning. Participants either listened to an auditory story or solved math problems (addition and subtraction of varying lengths). Each trial was followed by a forced-choice question with two alternatives. Effects: story, math. Social Cognition: participants watched video clips of objects (squares, circles, triangles) that either interacted in a socially meaningful way (mental condition) or moved randomly (random condition). They were tasked with judging social intent. Effects: mental, random. Relational Processing: this task assessed relational reasoning. In the relational condition, participants determined whether a pair of objects differed along the same dimension (shape or texture) as a reference pair. In the match condition, they judged whether an object matched one of two options based on a specified dimension. Effects: relational, match. Emotion Processing: participants identified which of two faces (angry or fearful) matched a third face, or which of two shapes matched a reference shape. They were not explicitly instructed that faces conveyed specific emotions. Effects: face, shape. 2.6.5. Beta maps correlations To assess the inter-session consistency of beta maps derived from game-based annotations, we computed Pearson correlations between voxel-wise beta values for each pair of session-level brain maps associated with the same annotation. We opted for Pearson correlation between whole-brain beta maps as a multivariate measure of spatial pattern similarity, rather than voxel-wise reliability metrics such as ICC. ICC-based approaches decompose variance into within- and between-participant components, requiring a sufficient sample size to robustly estimate between-participant variance, a condition not met in our four-participant dataset. Moreover, voxel-wise reliability metrics tend to underestimate the consistency of distributed activation patterns, even for well-established paradigms ( Elliott et al., 2020 ). In addition, we calculated inter-annotation correlations by measuring the relationship between brain maps from different annotations. This analysis was conducted separately for within-participant correlations (pairs of maps from the same participant) and between-participant correlations (pairs of maps from different participants). As a reference for reliability estimates, we performed the same analysis on run-level beta maps from the hcptrt dataset, which includes diverse event annotations representing motor and cognitive processes. Here, they serve as a benchmark for typical fMRI activation reliability rather than a direct comparison of experimental design principles. The shinobi and hcptrt datasets indeed differ fundamentally in design. Shinobi activation maps are derived from entire fMRI sessions, as ecological tasks such as videogames naturally allow for extended individual sampling. This prolonged sampling is necessary for certain videogame annotations, such as HealthLoss , which occur infrequently in some participants. In contrast, hcptrt activation maps are generated from single fMRI runs, with highly controlled designs specifically optimized for that short duration. 2.6.6. MVPA classification To determine whether each videogame contrast captured unique spatial patterns of brain activity, we conducted an MVPA-based brain decoding experiment using Nilearn ( Abraham et al., 2014 ). For each participant, a linear support vector machine (LinearSVC; scikit-learn, version 1.5.2) was trained on z-scored F maps derived from all session-wise activation maps from shinobi (11 regressors, 12–14 sessions per participant), and all run-level activation maps from hcptrt (16 regressors, 14–15 runs per participant). In total, the MVPA classified N = 27 contrast labels from 345–395 brain maps, independently for each participant. The classifier used default parameters: L2 penalty, squared hinge loss, one-vs-rest multiclass strategy, and regularization strength C = 1.0. The model was evaluated using a leave-one-session-out cross-validation procedure, training on all sessions (for shinobi ) and all runs (for hcptrt ) except one, and testing on the remaining session or run. As part of the cross-validation, feature selection was performed at the voxel level using an ANOVA-based screening method, where voxels were ranked independently, and the top 20% were retained to train the SVM. Model performance was assessed using a confusion matrix for all 27 classes, with statistical significance determined by comparison against a dummy classifier. 3. Results 3.1. Game events are frequent and long enough for fMRI analysis The most common actions were HIT (mean = 1,166, range: 491–2,000 occurrences per session), moving RIGHT (mean = 1,164, range: 604–1,580), and JUMP (mean = 1,041, range: 480–1,595), as these were essential for progressing through the game. Kill events were also relatively frequent (mean = 400, range: 234–589), whereas HealthLoss events were considerably rarer (mean = 51, range: 9–113) ( Fig. 4A ). Fig. 4. Open in a new tab Descriptive statistics of game events. (A) The events count per session. All events except HealthLoss occurred more than 200 times per session, making them appropriate for GLM analysis. (B) The distributions of button press durations. Movement-related button presses ( RIGHT , LEFT , DOWN ) were sustained for longer durations than JUMP and HIT button presses, for which duration didn’t meaningfully impact gameplay. Keypress durations varied across actions ( Fig. 4B ). The RIGHT button was generally held down for longer durations (mean = 1.27 s, IQR: 0.12–1.73 s), while LEFT was used in shorter bursts, primarily for position adjustments and boss fights (mean = 0.67 s, IQR: 0.22–0.88 s). DOWN keypresses had intermediate durations (mean = 0.82 s, IQR: 0.38–1.03 s). HIT , being a discrete action unaffected by keypress duration, had consistently short occurrences (mean = 0.24 s, IQR: 0.13–0.25 s). JUMP duration influenced game behavior, as longer presses resulted in higher jumps, but the mean duration remained relatively short (mean = 0.37 s, IQR: 0.20–0.48 s). Overall, the frequency (~250–1,200 per session, except for HealthLoss ) and duration (typically several hundred milliseconds) of game events were well-suited for event-related fMRI analyses. 3.2. GLM regressors had low to moderate correlations for most game events In-game actions and events were interdependent. For instance, in Shinobi III: RotNM , killing an enemy required executing the HIT action. Therefore, a correlation between the HIT and Kill annotations was expected. More complex dependencies also emerged. For instance, a complex action is the fly kick, which is executed by first pressing JUMP , followed by simultaneous pressing DOWN + HIT . To quantify these co-occurrences, regressors were generated for each annotation by convolving a boxcar function with the SPM HRF. Four low-level regressors were additionally included to capture stimulus-driven variance in the BOLD signal unrelated to discrete game events. Pearson correlations were then computed for each pair of regressors on every run and averaged across runs per participant to produce participant-specific correlation matrices ( Fig. 5 ). Fig. 5. Open in a new tab Average Pearson correlation between event regressors. Pearson correlations were computed for each pair of regressors on every run and then averaged per participant to produce correlation matrices. Dependence between game events ranges from very high to non-existent. As anticipated, HIT and Kill were moderately correlated (0.18 < r < 0.59), with Kills consistently preceded by HIT . DOWN also correlated with both HIT (0.40 < r < 0.81) and Kill (0.48 < r < 0.66). Many pairs of annotations, such as JUMP and HIT for sub-04, showed low levels of correlations (r < 0.11 across all participants). Correlations between game annotations and low-level regressors were generally low to moderate. The strongest coupling emerged between JUMP and luminance (0.43 < r < 0.57), consistent with the abrupt visual changes accompanying vertical movement. RIGHT also correlated moderately with audio envelope (0.35 < r < 0.56) and optical flow (0.36 < r < 0.50), reflecting the sustained audiovisual dynamics of horizontal movement. Audio envelope showed moderate correlations with most action-related annotations (mean r ranging 0.08–0.49), while HIT and Kill showed low-to-moderate correlations with all four low-level regressors (r < 0.43). HealthLoss was largely decoupled from all low-level features (r < 0.16 in all cases). Among the four low-level regressors themselves, moderate inter-correlations were observed (0.27 < r < 0.58), the strongest being luminance – audio envelope (0.49 < r < 0.58) and audio envelope – button press count (0.39 < r < 0.52), reflecting their shared sensitivity to overall gameplay intensity. Overall, these results indicate that while fMRI regressors derived from game events exhibited some degree of correlation, they captured distinct aspects of gameplay. Given these collinearities, we analysed main effects for each annotation in separate regression models, an approach consistent with the recommendations of Mumford et al. (2015) . Our goal was to assess the reliability of brain–behavior associations in videogame play rather than isolate the contributions of individual event types. 3.3. Brain activations related to game events reflect sensorimotor system involvement Z-scored statistical maps from second-level models are presented for each participant in relation to the Kill annotation ( Fig. 6A ). A table summary of activated regions using the AAL template in SM3 Supplementary Table S1 , and between-participant conjunction maps for all participants is provided in SM3 Supplementary Figure S2 . All the generated maps are made available in Supplementary Materials (SM2) . The majority of annotations elicited strong bilateral activation in regions associated with visual (superior and middle occipital gyri, cuneus) and sensorimotor (precentral and postcentral gyri, superior parietal gyrus) processing, along with multimodal areas (middle temporal gyrus, precuneus). Fig. 6. Open in a new tab Activation maps obtained with Kill (i.e., “killing an enemy”) events. (A) Each panel shows the participant-level maps (left side) and 4 session-level maps (right side) with the highest significant voxel count in the whole brain for each participant. The values shown are z-scores derived from the F-tests, thresholded at z > 3. (B) The conjunction of participant-level z-scores of the Kill and Audio envelope conditions. Voxels that were activated (3<|z|<6) in the Kill condition only are shown in blue (lighter to darker), and voxels activated in the Audio envelope condition only are shown in red. Voxels that were activated above threshold (|z|>3) in both conditions are shown in purple, with the average of the two z-scores shown from lighter to darker color. (C) Same as (B) but with the reward condition of hcptrt instead. Beyond these shared activations, some event-specific patterns emerged. HIT and JUMP actions were linked to motor regions, including the cerebellum and supplementary motor area, which were not as prominently engaged in the Kill and HealthLoss conditions. While Kill and HealthLoss annotations primarily activated visual and sensorimotor networks, they also engaged executive (e.g., superior frontal gyrus) and limbic (e.g., superior and middle temporal cortex) systems. Kill events, in particular, elicited activations in ventral (fusiform gyrus) and occipital (calcarine sulcus) visual cortices. Notably, in sub-01, clusters of activation were observed in the left frontal pole and anterior cingulate cortex in response to HealthLoss events. To assess the specificity of the Kill -related activation pattern with respect to both low-level stimulus properties and reward processing, Figure 6B and C display three-color surface overlays for each participant. The Kill vs. Audio envelope comparison ( Fig. 6B ) reveals that the large majority of cortical activation was shared with the Audio envelope (purple), spanning occipital, temporal, and parietal cortices, suggesting that much of the broad visual and multimodal response to Kill events was confounded with the low-level sensory signal. Kill -unique activations (blue) were more broad and extended to sensorimotor and frontal regions. By contrast, in the Kill vs. reward comparison ( Fig. 6C ), reward -unique activations (red) spanned nearly the entire cortex across all participants, vastly exceeding the spatial extent of Kill -unique activations (blue) which were more localized but included more parietal cortices. Overall, these findings demonstrate that event-related analyses leveraging automated gameplay annotations yield interpretable activations within visual, sensorimotor, and executive networks. 3.4. Within- and between-participant variability of brain maps compared to classical functional localizer tasks We assessed the robustness of first-level GLM beta maps generated for the same videogame annotations across individuals and recording sessions. To provide a meaningful reference, we also analyzed first-level beta maps derived from more conventional annotations from a battery of localizer tasks in the hcptrt dataset. Pearson correlations were computed between pairs of session-wise and run-wise beta maps for the same annotation, considering both videogame ( shinobi ) and hcptrt events ( Fig. 7A, B ). A dataset of the correlations are available in Supplementary Materials (SM4) . Fig. 7. Open in a new tab Test-retest consistency of brain activation patterns. The figure presents pairwise Pearson correlations (r) between beta maps derived from single fMRI sessions or runs. Points represent the average correlation between pairs of beta maps for each annotation, computed separately for (A) within-participants and (B) between-participants. Error bars indicate one standard deviation above and below the mean. Annotations are colour-coded according to cognitive domains, as indicated in the legend. The videogame annotations are shown in orange. (C–F) show within-participant correlations obtained from session-level maps built using a varying number of runs. They overall show diminishing returns, with the addition of the 5th run improving consistency less than the addition of a 2nd run. Distributions of within-participant correlations for the 28 annotations analyzed varied widely, ranging from r = .666 (SD = .119) for the hcptrt relational task to r = .0140 (SD = .188) for inter-annotation comparisons, which served as a lower-bound baseline. Among the 7 game-event annotations, three ranked within the top 10 highest average correlation values: Kill (r = .500; SD = .196), HIT (r = .495; SD = .186), and DOWN (r = .470; SD = .192). The remaining four game-events annotations achieved moderate positive correlations: RIGHT (r = .389; SD = .173), JUMP (r = .372; SD = .175), LEFT (r = .251; SD = .182), and HealthLoss (r = .215; SD = .186). The four low-level regressors also showed moderate-to-strong within-participant consistency: Button press ranked third overall (r = .509; SD = .175), followed by Audio envelope (r = .472; SD = .176), Luminance (r = .438; SD = .188), and Optical flow (r = .361; SD = .179). Inter-annotation comparisons yielded low correlations (r = .140; SD = .188), confirming the distinctiveness of annotation-specific neural patterns. A detailed breakdown of within-participant correlations across all pairs of maps is available in SM3 Supplementary Figure S3 . Between-participant correlations were generally lower than within-participant correlations, spanning from r = .385 (SD = .094) for the relational task to r = .067 (SD = .057) for LEFT . Among the game-event annotations, Kill (r = .196; SD = .095), HIT (r = .190; SD = .096), and DOWN (r = .186; SD = .096) showed the highest between-participant consistency, followed by JUMP (r = .148; SD = .065) and RIGHT (r = .137; SD = .075). HealthLoss (r = .068; SD = .056) and LEFT (r = .067; SD = .057) showed weak or negligible between-participant correlations. The low-level regressors yielded between-participant correlations broadly comparable to those of the game-event annotations: Button press (r = .218; SD = .096), Audio envelope (r = .209; SD = .087), Luminance (r = .200; SD = .095), and Optical flow (r = .108; SD = .068). Inter-annotation comparisons resulted in near-zero correlation (r = .074; SD = .097). Figure 7C–F show within-participant correlations for each of the 11 annotations as a function of the number of runs aggregated into each session-level map (from 1 to 5 runs), separately for each participant. Across almost all annotations and participants, within-participant correlations increased monotonically with the number of runs, with the largest gain occurring between 1 and 2 runs. Subsequent additions yielded progressively smaller improvements, consistent with a pattern of diminishing returns: the benefit of a fifth run was substantially smaller than that of a second (and even decreased consistency for sub-01 and for 2 annotations of sub-06). This pattern was consistent across participants and across both game-event and low-level feature annotations, though the asymptotic level varied: annotations with higher overall reliability (e.g., Kill , HIT ) reached higher plateaus than those with lower reliability (e.g., HealthLoss , LEFT ). These results suggest that a minimum of 2–3 runs per session is sufficient to obtain stable within-participant brain maps for videogame annotations, with limited additional benefit beyond that. These results indicate that within-participant consistency of brain maps derived from videogame events is comparable to that of established functional localizer tasks, though they are not necessarily the highest overall. The lower between-participant correlations highlight the personalized nature of neural representations elicited by videogame interactions. The overall findings suggest that videogame-derived brain maps exhibit reliability at the individual level while preserving event-specific distinctiveness. 3.5. Game events activation maps are discriminable We used a linear support vector machine (SVM) classifier to test whether fMRI activation maps could support reliable classification between different game events. The classification was conducted separately for each participant using a leave-one-session-out cross-validation approach. Specifically, an SVM multiclass model was trained on session-wise Z maps from Shinobi gameplay and tested on run-wise activation maps from the HCP test–retest dataset. Classification performance was evaluated based on the most frequently predicted class across folds for each annotation and participant ( Fig. 8 ). The trained classifiers and per-fold confusion matrices are available in Supplementary Materials (SM5) . Fig. 8. Open in a new tab Confusion matrices of MVPA classifications. A linear SVM was trained to classify Z maps from session-wise game annotations and run-wise HCP test–retest annotations, and the models were evaluated with a Leave-One-Group-Out cross-validation. For all game annotations, the classifier correctly identified the true classes above chance level, with occasional confusions between game-related and motor-related maps. Overall, activation maps exhibited high discriminability across annotation types (all p < 0.001). Kill and HIT were classified with the highest accuracy among game events (mean accuracy = 0.98 across participants), followed by RIGHT (0.96) and JUMP (0.92). DOWN (0.85), LEFT (0.78), and HealthLoss (0.71) showed comparatively lower accuracy, with the largest drops observed in sub-06 ( HealthLoss : 0.36, LEFT : 0.46) and sub-02 ( DOWN and JUMP : 0.75). The four low-level regressors were also reliably classified, with audio envelope achieving the highest accuracy (0.96), followed by button press count (0.89), optical flow (0.87), and luminance (0.87). In most cases, misclassifications of game events involved other game events. Other classification errors on game events were predictions of the motor task, suggesting partial representational overlap with the motor domain. These results confirm that automatically annotated game events elicit distinctive, task-specific activation patterns that can be robustly decoded from fMRI data. 4. Discussion Four participants played three levels of Shinobi III: Return of the Ninja Master (1993) during fMRI scanning. Gameplay was implemented using a console emulator integrated with the gym-retro library ( Nichol et al., 2018 ) and controlled via an MRI-compatible gamepad ( Harel et al., 2023 ). BOLD activity was modeled using automated annotations derived directly from the game’s RAM variables. As a proof-of-concept, we generated interpretable, consistent, and distinct brain activation maps linked to these automatically extracted game events and player actions, demonstrating the feasibility and scalability of this novel annotation approach. 4.1. Automated content annotation during naturalistic videogame play By mining the game’s internal memory states, we automatically annotated both player actions and game events. Events such as defeating an enemy ( Kill ) and losing health ( HealthLoss ) were recorded alongside player actions, including character movement, jumping ( JUMP ), and hitting ( HIT ). The Kill and HealthLoss events represent positively and negatively valenced occurrences, respectively, analogous to manually annotated events in previous studies ( Cole et al., 2012 ; Kätsyri et al., 2013a , 2013b ; Zhang et al., 2021 ). Unlike these earlier efforts, our annotations were generated automatically from an unmodified commercial videogame, enabling seamless scalability to a substantial dataset (32 h of cumulative gameplay), from which we extracted thousands of event instances per participant. This automatic annotation strategy represents a key departure from previous studies that relied on custom-built games or manual annotation pipelines. For instance, Cole et al. (2012) analyzed brain responses during gameplay in Re-Mission , a serious game designed for cancer patients, collecting 6–7 h of fMRI data across 57 participants. Kätsyri et al. (2013a , 2013b ) manually annotated around 15–20 events per condition from 11 participants playing BZ-Flag . In contrast, our method operates on an unmodified commercial game, scales seamlessly across 32 h of data, and yields thousands of precisely timed (frame-level) event annotations per participant, which is orders of magnitude higher than earlier work. Unlike manual annotations, which often involve downsampling and introduce timing imprecision, our automated approach retains high temporal fidelity, making it suitable not only for fMRI but also for high-temporal-resolution modalities like MEG, EEG, or iEEG. That scaling factor enabled our analysis to yield interpretable and sensitive maps even for rare events like HealthLoss . This suggests that automated annotation not only supports scalability, but can retain the sensitivity of more labor-intensive methods. Nonetheless, imbalances in event prevalence should be handled with care when making contrasts, especially given the heightened salience and stronger neural responses associated with rare, unexpected events ( Bueti et al., 2010 ; Matthews & Gheorghiu, 2016 ). 4.2. Player actions and game-events induce widespread activations including visual and motor cortices Our GLM analyses revealed a distributed pattern of visual and sensorimotor activations, most consistently involving occipital regions (e.g., cuneus, lingual gyrus) and parietal cortices (e.g., precentral and postcentral gyri). This finding is consistent with the visuomotor demands of playing an action game like Shinobi III: Return of the Ninja Master , and mirrors cortical recruitment patterns previously documented in the literature (e.g., Li et al., 2016 ). Videogame players are known to exhibit enhanced visuomotor skills compared to non-players, along with structural and functional differences in relevant brain regions ( Choi et al., 2020 ). Moreover, videogame training has been shown to induce changes in brain structure and function within the cognitive domains engaged by gameplay ( Brilliant et al., 2019 ). Player actions such as HIT and JUMP were associated with activation in the supplementary motor area and cerebellum. These regions are central to the human motor system and play crucial roles in motor preparation, voluntary control, visuomotor adaptation, and temporal prediction ( Nachev et al., 2008 ; Popa & Ebner, 2019 ; Tanaka et al., 2021 ; Tzvi et al., 2022 ). In particular, the observed activation of the supplementary motor area during player-initiated actions supports prior evidence that this region is engaged when motor commands are self-generated rather than externally driven ( Nachev et al., 2008 ; Orban et al., 2015 ). The HIT , JUMP , and Kill events also elicited activation in the superior frontal gyrus, a region implicated in the generation of complex motor sequences ( Martino et al., 2011 ). Zhang et al. (2021) reported that kill events in a first-person shooter elicited distributed activation patterns, including the supplementary motor area and frontal eye fields—regions associated with inhibitory control and oculomotor processing ( Diarra et al., 2019 ). Kätsyri et al. (2013a) found that success (kills) and failure (deaths) events modulated activity in ventromedial and orbitomedial regions of the prefrontal cortex, areas implicated in reward processing ( Platt & Padoa-Schioppa, 2009 ) potentially reflecting affective and attentional responses to task feedback. In our dataset, one participant exhibited anterior cingulate activation in response to health loss, potentially reflecting the processing of negative emotional valence ( Rolls, 2019 ). To directly probe this overlap, we compared Kill -related activations against a canonical reward condition from the hcptrt gambling task ( Fig. 6C ). reward yielded substantially more widespread cortical activations than Kill , likely reflecting the considerably higher statistical power afforded by the hcptrt design’s rather than a fundamentally broader neural response. Kill -unique activations were spatially more focal, concentrated in sensorimotor and posterior visual regions. Combined with the Kill vs. audio envelope comparison ( Fig. 6B ), which showed that much of the Kill response is shared with low-level auditory features, these findings suggest that Kill -related activations represent a more specific signal embedded within a broader landscape of stimulus-driven and general task-related responses. Most annotations also consistently engaged the precuneus and middle temporal gyrus, two regions not typically highlighted in primary sensorimotor or visual processing. The precuneus is a functionally heterogeneous structure whose subdivisions span sensorimotor, cognitive/associative, and visual networks ( Margulies et al., 2009 ; Zhang et al., 2023 ), while the middle temporal gyrus is a recognized hub for audiovisual and action-feedback integration ( Scheliga et al., 2022 ; van Kemenade et al., 2019 ). Their shared recruitment across both player actions and game outcomes likely reflects the continuous multimodal integration demands of gameplay, and illustrates how a single naturalistic task can simultaneously engage the full processing chain from early sensory encoding to higher-order integration. Collectively, these results converge with existing literature, showing that brain activity patterns identified via automated annotation correspond to previously reported neural correlates of gameplay. This provides strong evidence that automated annotations can validly and effectively model relevant neural processes. 4.3. Event-related videogame activations show test-retest consistency, and are participant-specific Our analyses revealed that brain activation patterns derived from automatically annotated game events were not only consistent across sessions, but also carried clear participant-specific signatures. This level of test-retest reliability, demonstrated using entire sessions as independent folds, is made possible by the unprecedented scale of the CNeuroMod dataset. Such large individual datasets allow us to examine both the reproducibility and individuality of activation patterns in a way that is rarely feasible in fMRI studies. While metrics such as ICC would provide a more formal reliability framework, they require a sufficient sample size to robustly estimate between-participant variance. Our design, with few participants but many sessions per participant, is better suited to multivariate spatial correlation approaches that assess the consistency of whole-brain patterns within and between individuals. We observed that while certain annotations (e.g., HIT , Kill ) led to highly consistent activation patterns across sessions, the strongest signal came from intra-individual similarity: each participant showed a reliable “neural fingerprint” across time. This participant specificity suggests that the neural encoding of complex game events is both stable and individualised. Low-level regressors ( button press count , audio envelope , luminance ) also exhibited within-participant consistency comparable to or exceeding that of several hcptrt localizer tasks, suggesting that part of the reliability observed for game-event annotations may be driven by their correlation with stable low-level stimulus features. Notably, this individualization likely reflects more than just biological variability. Because videogames are interactive and open-ended, players shape their own experience through strategy, pacing, and response to in-game challenges, effectively constructing personalized versions of the task. These player-specific gameplay trajectories can drive idiosyncratic cognitive demands, contributing to the observed neural distinctiveness. In this context, videogame-based paradigms offer a powerful lens for investigating how specific brains engage with dynamic, ecologically valid environments, with promising applications in the study of individual differences and clinical populations. Between-participant conjunction maps ( SM3 Supplementary Fig. S2 ) nonetheless reveal that several activation patterns, particularly in visual and sensorimotor cortices, were consistently observed across all four participants for most annotations. This suggests shared neural substrates despite the overall dominance of individual-specific signatures. However, our design with only four participants does not allow for a formal statistical assessment of group-level effects beyond these descriptive observations. An additional analysis examining how within-participant map consistency scales with the number of runs per session revealed a pattern of diminishing returns: the largest gain in reliability occurred when moving from a single run to two runs, with each subsequent run contributing progressively less. It also shows that even with a few runs reliable brain maps can be derived, depending on the type of annotation. This finding has practical implications for experimental design: the frequency and regularity of the events of interest should be carefully considered when determining the volume of data required to achieve stable, reliable brain maps for a given annotation. 4.4. Event-related videogame activations are annotation-specific The MVPA classification of session-wise activation maps revealed that neural patterns linked to distinct game annotations were highly discriminable. This is especially striking given that all annotations originated from a single, continuous task rather than from separate experimental conditions. Comparable classification accuracy has previously been achieved across very different tasks (e.g., motor vs language) in large datasets like the Human Connectome Project ( Porter et al., 2023 ; Rastegarnia et al., 2023 ). Here, we show that similarly distinct neural signatures can be recovered from different aspects of a single task, further highlighting the richness of cognitive states evoked during gameplay. This discriminability is noteworthy given the substantial temporal co-occurrence of several game events (e.g., DOWN , HIT , and Kill ; r up to 0.81). Although collinearity was handled by fitting each annotation in a separate GLM, the resulting whole-brain patterns remained highly distinguishable, consistent with the known sensitivity of multivariate decoding to distributed, condition-specific signals that univariate approaches cannot isolate ( Haxby et al., 2001 ; Jimura & Poldrack, 2012 ). The four low-level regressors were also reliably classified, with accuracies comparable to several game-event annotations, confirming that these continuous sensory and motor features engage spatially distinct brain patterns rather than capturing undifferentiated nuisance variance. The lower classification accuracy for LEFT and HealthLoss , particularly in sub-06, likely reflects insufficient data: both events were infrequent, and the reliability scaling analysis ( Fig. 7C–F ) showed that annotations with fewer occurrences reach lower reliability and do not plateau in our dataset. For HealthLoss specifically, additional variability may arise from the context-dependent nature of outcome processing ( Fouragnan et al., 2018 ), though disentangling data-volume and encoding-variability contributions would require denser event sampling. These results position automated videogame annotation as a viable approach for cognitive state decoding within a single continuous naturalistic task, without the artificiality of blocked designs. The robust within-participant patterns observed here, combined with evidence that naturalistic paradigms preserve or enhance individual neural signatures ( Finn, 2020 ; Vanderwal et al., 2017 ), suggest promising applications in single-participant brain mapping and the study of individual differences. 4.5. Limitations Our approach relies on reverse engineering memory addresses to identify the location of relevant game variables. Although this mapping can be exhaustive in principle, it is often incomplete in practice, spanning only a limited subset of variables. Depending on the game software’s architecture and physical constraints of the support (size, bit-depth), certain variables can be encoded in unintuitive ways and be tedious to recover. This process however only has to be performed once, and partial or complete variables mapping can sometimes be available online through resources from the modding community, limiting the amount of work required to access experimentally meaningful variables. Memory-based annotations provide a temporal resolution equivalent to the games framerate, which was 60 Hz in our case. Although this temporal resolution is far beyond that of fMRI, other brain measurement techniques such as MEG and EEG offer millisecond resolution. Using our annotations framework in these cases requires precise logging of individual frames’ timing, which enables matching a variable change to the exact moment at which the corresponding frame is displayed. Given the continuous nature of gaming stimuli however, events commonly spans multiple frames, and stimulus onset as perceived by the player (e.g., feedback on a given action, appearance of an enemy on screen) is not always perfectly aligned with the frame corresponding to the variable change and can vary depending on animations and inherent delays. The GLM approach used in the current set of analyses carries inherent limitations that constrain its application to more complex event structures. Because GLMs assume low collinearity among regressors, we had to analyse each annotation separately, an approach that may not scale well as annotation granularity increases. In practice, some collinearity is unavoidable: in dynamic tasks like gameplay, events unfold in a causal sequence and are not temporally isolated. GLMs impose a linear relationship between regressors and BOLD signals, which may not adequately capture the nonlinear dynamics of brain responses during naturalistic, context-dependent events. Prior studies have demonstrated such nonlinearities in complex motor tasks ( Alahmadi et al., 2016 ) and in contexts where event meaning varies ( Orban et al., 2015 ; Tsivilis et al., 2003 ). Exploring nonlinear modeling, such as brain encoding with AI models ( Ahmadi et al., 2024 ; Cross et al., 2021 ; Kemtur et al., 2023 ), may offer deeper insights into how diverse cognitive processes interact during gameplay ( Gonzalez-Castillo et al., 2012 ). Additionally, model-free approaches such as finite impulse response (FIR) models could be used to characterize the shape of event-locked BOLD responses without assuming a canonical hemodynamic profile, providing a complementary validation of the activations identified through our GLM analyses. 4.6. Relevance and future directions The present analyses focused on low-level, discrete game events that map basic perceptual and motor components of gameplay. However, videogame play notoriously engages higher-order processes, including learning, reasoning, planning, and affective evaluation, that our current annotations do not capture ( Schrodt et al., 2017 ). Extending this framework to study learning, for instance, would require segmenting platformer levels into smaller gameplay units based on recurring design elements (e.g., platform layout, enemy types; Smith et al., 2008 ), chunking level repetitions, and tracking progress along specific performance metrics; we explore this direction in a companion benchmark proposal ( Harel et al., 2025 ). Affective processing represents another natural extension: the Kill and HealthLoss annotations already serve as proxies for positively and negatively valenced outcomes, and physiological data (cardiac, respiratory) were recorded concurrently with fMRI in the present dataset. Other datasets acquired within the CNeuroMod project further include phenomenological questionnaires alongside physiological recordings, and will be the subject of future work. Automated annotations from memory can support these efforts at scale, and when combined with the gym-retro library, which enables programmatic control over the emulator’s state, a game can be “taskified” (i.e., configured into a well-controlled experimental task; Allen et al., 2024 ) to target specific aspects of higher-order cognition. A notable strength of videogame paradigms is their capacity to generate large volumes of repeated events under sustained participant engagement. In the present dataset, a single annotation such as Kill was observed approximately 400 times per session and over 4,000 times per participant across the full experiment. For comparison, large-scale visual neuroscience datasets typically present on the order of thousands of unique images with only a handful of repetitions each: the Natural Scenes Dataset presents ~10,000 images shown 3 times each (~30,000 total trials; Allen et al., 2022 ), CNeuroMod-THINGS presents ~4,300 images shown 3 times (~13,000 trials; St-Laurent et al., 2026 ), and THINGS-data presents ~1,854 images shown once ( Hebart et al., 2023 ). A single videogame event type thus yields a comparable number of observations to the entire stimulus set of these datasets, but with each instance occurring in a unique gameplay context involving different enemies, spatial configurations, and action sequences. This contextual variability, combined with intrinsic participant motivation, makes videogames well suited for studying how the brain generalises representations across naturalistic conditions. Moreover, because gameplay simultaneously engages perception, decision-making, motor planning, and affective evaluation as a coupled dynamical system, it offers a window into cognitive processes that are typically studied in isolation through separate experimental tasks. Importantly, videogame research can draw on a robust ecosystem of design practices, community-developed tools, and analytic frameworks, which makes implementation both flexible and scalable. Videogame paradigms are also inherently engaging and accessible, lending themselves well to use in populations for whom conventional cognitive tasks may be unsuitable such as children ( Vanderwal et al., 2019 ) or clinical populations ( Chitale et al., 2022 ; Peñuelas-Calvo et al., 2022 ). Moreover, games may uncover compensatory mechanisms in individuals with cognitive challenges, offering a promising avenue for therapeutic applications ( Han et al., 2009 ). Because they are embedded in context-sensitive, continuous tasks, videogames also allow for the detection of subtle behavioural and neural shifts that might be missed in more constrained paradigms. Our approach, which enables scalable, automated annotation across thousands of games, opens the door to broader, more inclusive applications of these tools. Videogames are also well suited to studying states of intense task engagement. Psychological flow, characterized by deep absorption, loss of self-consciousness, and a sense of effortless control, is notoriously difficult to induce and observe under laboratory conditions ( Abuhamdeh, 2020 ), yet arises readily during videogame play ( Khoshnoud et al., 2020 ; 2022 ). Unlike most experimental tasks, videogames allow researchers to recruit individuals with genuine expertise from the general population, ensuring that the skill levels required for flow to emerge are ecologically represented. Moreover, because videogames are inherently engaging, participants can sustain performance over extended acquisition sessions, increasing the likelihood that rare, spontaneous phenomena such as flow will occur in sufficient quantities to be analysed. Combined with physiological recordings and automated annotations, the fine-grained performance metrics available through game analytics can complement subjective reports to help identify and characterise these episodes, constituting a promising avenue for future studies. These paradigms also hold significant promise for bridging human and artificial neural systems ( Bellec & Boyle, 2019 ). Artificial neural networks (ANNs) trained on multimodal gameplay data can be used to predict neural activity, as shown by Cross et al. (2021) , who demonstrated correspondence between ANN and human brain responses in sensorimotor regions. Yet, the interpretability of ANN models remains a challenge due to the complexity of their internal representations. Automated annotation strategies like ours provide a pathway toward greater interpretability by linking brain responses to specific, meaningful in-game events (e.g., “killing an enemy” or “losing health”). For instance, if an ANN layer becomes selectively responsive to Kill events, we can begin to isolate the visual or cognitive components underlying this prediction. This synergy between data-driven modeling and semantically grounded annotations enhances the explanatory power of brain–ANN alignment. In addition, videogame tasks can support advanced machine learning approaches such as imitation learning ( Hussein et al., 2017 ), where models learn to replicate human gameplay. These models can be constrained or validated using brain activity, enabling the development of brain-aligned ANNs that mirror individual players’ neural patterns ( Freteault et al., 2025 ; Kemtur et al., 2023 ). Our approach, with its combination of naturalistic stimuli and rich, structured annotations, offers a promising testbed for training and evaluating such models. Finally, the abundance of annotated data generated by this approach facilitates exploratory analyses and hypothesis generation, helping to uncover the neural mechanisms underlying real-world behaviour in a scalable, interpretable way. 4.7. Conclusion Our results show that automated annotations derived from unmodified retro videogame play can reliably produce interpretable brain activation maps. These maps engage biologically plausible networks that are specific to distinct in-game events, demonstrating that dynamic, naturalistic stimuli like videogames can be leveraged to model reproducible and functionally meaningful neural responses. Although this study focused on Shinobi III: Return of the Ninja Master, the methodology generalizes readily to the entire gym-retro library, spanning over 1,000 games across 10 different consoles, and can be extended to many other platforms. We hope that this work will encourage broader adoption of commercial videogames in cognitive neuroscience and contribute to a richer understanding of brain function under naturalistic, ecologically valid conditions. Supplementary Material Supplementary Material IMAG.a.1256_supp.zip (921.1MB, zip) Acknowledgments The Courtois project on neural modeling was made possible by a generous donation from the Courtois foundation, administered by the Fondation Institut Gériatrie Montréal at CIUSSS du Centre-Sud-de-l’île-de-Montréal and University of Montreal. The Courtois NeuroMod team is based at “Centre de Recherche de l’Institut Universitaire de Gériatrie de Montréal”, with several other institutions involved. See the CNeuroMod documentation for an up-to-date list of contributors ( https://docs.cneuromod.ca ). L.B. is a senior fellow (“chercheuse boursière senior”) of the “Fonds de recherche du Québec – Santé” and a Discovery Grant from the Natural Sciences and Engineering Research Council of Canada. K.J. is supported by funding from the Canada Research Chairs program and a Discovery Grant from the Natural Sciences and Engineering Research Council of Canada. The authors acknowledge the use of generative AI in refining the phrasing and flow of the manuscript, improving clarity and coherence throughout the text. All suggestions provided by the language model were reviewed and validated by the authors. The authors would like to thank Himanshu Agarwal for kindly sharing his code that served as a basis for the RAM variables visualization tool. The authors would like to thank the Courtois-Neuromod scanning crew for acquiring the data. Appendices Appendix Table A1. Descriptive table of the shinobi_training dataset. sub-01 sub-02 sub-04 sub-06 Total level-1 N 164 263 57 34 518 cleared 101 211 40 17 369 duration (h:m:s) 07:31:27 12:56:27 03:56:13 02:26:04 26:50:11 level-4 N 108 288 98 54 548 cleared 39 54 37 5 135 duration (h:m:s) 03:01:36 10:39:10 05:33:00 02:33:39 21:47:25 level-5 N 116 51 57 36 260 cleared 86 16 26 9 137 duration (h:m:s) 05:00:33 02:58:27 03:13:23 01:54:04 13:06:27 Total sessions 53 110 24 15 202 N 388 602 212 124 1326 cleared 226 281 103 31 641 duration (h:m:s) 15:33:36 26:34:04 12:42:37 06:53:46 61:44:04 Open in a new tab A.1. Performance Comparison of Training vs Scanning Phases Training sessions were ordered chronologically by timestamp and scan sessions by acquisition order; each sequence was split at its midpoint into a first and second half. Three consecutive half-pairs were compared within training, from training to scanning, and within scanning, using two-sided Welch’s t -tests with Benjamini-Hochberg FDR correction applied across all comparisons per participant. This was done separately for each level and metric (final score, proportion cleared, health loss, and screen-clearing rate). A.2. Skill Acquisition During Training Transfers to the Scanner Participants trained at home for 59–84 weeks prior to scanning (16–110 sessions depending on participant), followed by 11–12 MRI scan sessions. All four participants improved substantially across the training phase, with significant gains in proportion of screens cleared and final score at levels 1 and 4 in all participants (all q < .05; see Appendix Fig. A1 ). Level-5 improvements were less consistent, likely due to ceiling effects and fewer training sessions. Skills transferred robustly to the scanner at level 1, with all participants entering the scan phase well above their end-of-training performance. Transfer at level 4 was more variable: two participants showed slightly lower performance at scan entry than at training end, possibly reflecting a brief adjustment to the scanner environment. Performance was largely stable across the scanning period, with 32 of 40 within-scan comparisons non-significant. A small number of continued improvements were observed in a few participants at levels 1 and 4, consistent with the more regular weekly scan schedule providing additional structured practice. Appendix Fig. A1. Open in a new tab Skill acquisition across training and scanning phases. For each level and participant, training and scan sessions were each split at their midpoint (in order of repetitions) into a first and second half, and the three consecutive half-pairs were compared: within training, from training to scanning (transfer), and within scanning. Significance was assessed with two-sided Welch’s t-tests, FDR-corrected using the Benjamini-Hochberg procedure: * q < .05; ** q < .01; *** q < .001. Data and Code Availability All datasets used in this study are part of the CNeuroMod dataset ( https://www.cneuromod.ca/ ). The code developed for this research is publicly available on GitHub ( https://github.com/courtois-neuromod/shinobi_fmri ) and archived on Zenodo for long-term preservation ( https://doi.org/10.5281/zenodo.15628556 ). Processed datasets used to generate the tables and figures in this manuscript are provided as Supplementary Material accompanying the paper. Author Contributions Conceptualization: YH, BP, JB, and LB. Methodology: YH, BP, JB, and LB. Software: BP, PHM, and MLC. Formal analysis: YH. Investigation: YH. Resources: BP, PHM, and VB. Data curation: BP. Writing—original draft: YH. Writing—review & editing: All authors. Visualization: YH. Supervision: KJ, LB. Project administration: JB. Funding acquisition: LB. Declaration of Competing Interest The authors hereby declare that they have no competing interests, financial or otherwise, in the participant matter or materials discussed in this manuscript. All research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. Supplementary Materials Supplementary material for this article is available with the online version here: https://doi.org/10.1162/IMAG.a.1256#supplementary-data . References Abraham, A., Pedregosa, F., Eickenberg, M., Gervais, P., Mueller, A., Kossaifi, J., Gramfort, A., Thirion, B., & Varoquaux, G. (2014). Machine learning for neuroimaging with scikit-learn. Frontiers in Neuroinformatics, 8, 14. 10.3389/fninf.2014.00014 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Abuhamdeh, S. (2020). Investigating the “flow” experience: Key conceptual and operational issues. Frontiers in Psychology, 11, 158. 10.3389/fpsyg.2020.00158 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ahmadi, S., Paugam, F., Glatard, T., & Bellec, P. L. (2024). Training compute-optimal vision transformers for brain encoding. arXiv (No. arXiv:2410.19810). 10.48550/arXiv.2410.19810 [ DOI ] [ Google Scholar ] Alahmadi, A. A. S., Samson, R. S., Gasston, D., Pardini, M., Friston, K. J., D’Angelo, E., Toosy, A. T., & Wheeler-Kingshott, C. A. M. (2016). Complex motor task associated with non-linear BOLD responses in cerebro-cortical areas and cerebellum. Brain Structure and Function, 221(5), 2443–2458. 10.1007/s00429-015-1048-1 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Allen, E. J., St-Yves, G., Wu, Y., Breedlove, J. L., Prince, J. S., Dowdle, L. T., Nau, M., Caron, B., Pestilli, F., Charest, I., Hutchinson, J. B., Naselaris, T., & Kay, K. (2022). A massive 7T fMRI dataset to bridge cognitive neuroscience and artificial intelligence. Nature Neuroscience, 25(1), 116–126. 10.1038/s41593-021-00962-x [ DOI ] [ PubMed ] [ Google Scholar ] Allen, K., Brändle, F., Botvinick, M., Fan, J. E., Gershman, S. J., Gopnik, A., Griffiths, T. L., Hartshorne, J. K., Hauser, T. U., Ho, M. K., de Leeuw, J. R., Ma, W. J., Murayama, K., Nelson, J. D., van Opheusden, B., Pouncy, T., Rafner, J., Rahwan, I., Rutledge, R. B., … Schulz, E. (2024). Using games to understand the mind. Nature Human Behaviour, 8(6), 1035–1043. 10.1038/s41562-024-01878-9 [ DOI ] [ PubMed ] [ Google Scholar ] Barch, D. M., Burgess, G. C., Harms, M. P., Petersen, S. E., Schlaggar, B. L., Corbetta, M., Glasser, M. F., Curtiss, S., Dixit, S., Feldt, C., Nolan, D., Bryant, E., Hartley, T., Footer, O., Bjork, J. M., Poldrack, R., Smith, S., Johansen-Berg, H., Snyder, A. Z., … WU-Minn HCP Consortium. (2013). Function in the human connectome: Task-fMRI and individual differences in behavior. NeuroImage, 80, 169–189. 10.1016/j.neuroimage.2013.05.033 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bellec, P., & Boyle, J. A. (2019). Bridging the gap between perception and action: the case for neuroimaging, AI and video games. PsyArXiv. https://osf.io/preprints/psyarxiv/3epws_v1 Boudreau, M., Karakuzu, A., Boré, A., Pinsard, B., Zelenkovski, K., Alonso-Ortiz, E., Boyle, J., Bellec, L., & Cohen-Adad, J. (2025). Longitudinal reproducibility of brain and spinal cord quantitative MRI biomarkers. Imaging Neuroscience, 3, imag_a_00409. 10.1162/imag_a_00409 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Boyle, J. A., Pinsard, B., Boukhdhir, A., Belleville, S., Brambatti, S., Chen, J., Cohen-Adad, J., Cyr, A., Fuente Rainville, P., & Bellec, P. (2020). The Courtois project on neuronal modelling-first data release. In 26th Annual Meeting of the Organization for Human Brain Mapping. https://docs.cneuromod.ca/en/latest/_downloads/a76a8a6439d66e14744a2c9d47937138/abstract_ohbm_cneuromod_release_2020.pdf Brilliant, D., Nouchi, R., & Kawashima, R. (2019). Does video gaming have impacts on the brain: Evidence from a systematic review. Brain Sciences, 9(10), Article 10. 10.3390/brainsci9100251 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., & Zaremba, W. (2016, June 5). OpenAI Gym. arXiv.Org . https://arxiv.org/abs/1606.01540v1 Bueti, D., Bahrami, B., Walsh, V., & Rees, G. (2010). Encoding of temporal probabilities in the human brain. Journal of Neuroscience, 30(12), 4343–4352. 10.1523/JNEUROSCI.2254-09.2010 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chitale, V., Baghaei, N., Playne, D., Liang, H. N., Zhao, Y., Erensoy, A., & Ahmad, Y. (2022). The use of videogames and virtual reality for the assessment of anxiety and depression: A scoping review. Games for Health Journal, 11(6), 341–354. 10.1089/g4h.2021.0227 [ DOI ] [ PubMed ] [ Google Scholar ] Choi, E., Shin, S.-H., Ryu, J.-K., Jung, K.-I., Kim, S.-Y., & Park, M.-H. (2020). Commercial video games and cognitive functions: Video game genres and modulating factors of cognitive enhancement. Behavioral and Brain Functions, 16(1), 2. 10.1186/s12993-020-0165-z [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Choi, M.-H., Kim, H.-S., Yoon, H.-J., Lee, J.-C., Baek, J.-H., Choi, J.-S., Tack, G.-R., Min, B.-C., Lim, D.-W., & Chung, S.-C. (2017). Increase in brain activation due to sub-tasks during driving: fMRI study using new MR-compatible driving simulator. Journal of Physiological Anthropology, 36(1), 11. 10.1186/s40101-017-0128-8 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] CNeuroMod. (2025a). Courtois Neuromod Datasets Documentation. https://docs.cneuromod.ca/en/latest/DATASETS.html CNeuroMod. (2025b). Courtois Neuromod Derivatives Documentation. https://docs.cneuromod.ca/en/latest/DERIVATIVES.html CNeuroMod. (2025c). Courtois Neuromod MRI Acquisitions Documentation. https://docs.cneuromod.ca/en/latest/MRI.html CNeuroMod. (2025d). Scanning phase of the Shinobi Dataset. https://github.com/courtois-neuromod/shinobi CNeuroMod. (2025e). Task Codebase of the Courtois Neuromod Project. https://github.com/courtois-neuromod/task_stimuli CNeuroMod. (2025f). Training phase of the Shinobi Dataset. https://github.com/courtois-neuromod/shinobi_training Cole, S. W., Yoo, D. J., & Knutson, B. (2012). Interactivity and reward-related neural activation during a serious videogame. PLoS One, 7(3), e33909. 10.1371/journal.pone.0033909 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Cross, L., Cockburn, J., Yue, Y., & O’Doherty, J. P. (2021). Using deep reinforcement learning to reveal how the brain encodes abstract state-space representations in high-dimensional environments. Neuron, 109(4), 724.e7–738.e7. 10.1016/j.neuron.2020.11.021 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] de la Vega, A., Rocca, R., Blair, R. W., Markiewicz, C. J., Mentch, J., Kent, J. D., Herholz, P., Ghosh, S. S., Poldrack, R. A., & Yarkoni, T. (2022). Neuroscout, a unified platform for generalizable and reproducible fMRI research. eLife, 11, e79277. 10.7554/eLife.79277 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Diarra, M., Zendel, B. R., Benady-Chorney, J., Blanchette, C.-A., Lepore, F., Peretz, I., Belleville, S., & West, G. L. (2019). Playing Super Mario increases oculomotor inhibition and frontal eye field grey matter in older adults. Experimental Brain Research, 237(3), 723–733. 10.1007/s00221-018-5453-6 [ DOI ] [ PubMed ] [ Google Scholar ] Elliott, M. L., Knodt, A. R., Ireland, D., Morris, M. L., Poulton, R., Ramrakha, S., Sison, M. L., Moffitt, T. E., Caspi, A., & Hariri, A. R. (2020). What is the test-retest reliability of common task-functional MRI measures? New empirical evidence and a meta-analysis. Psychological Science, 31(7), 792–806. 10.1177/0956797620916786 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Eickhoff, S. B., Milham, M., & Vanderwal, T. (2020). Towards clinical applications of movie fMRI. NeuroImage, 217, 116860. 10.1016/j.neuroimage.2020.116860 [ DOI ] [ PubMed ] [ Google Scholar ] Esteban, O., Markiewicz, C., Blair, R. W., Moodie, C., Isik, A. I., Erramuzpe Aliaga, A., Kent, J., Goncalves, M., DuPre, E., Snyder, M., Oya, H., Ghosh, S., Wright, J., Durnez, J., Poldrack, R., & Gorgolewski, K. J. (2019). fMRIPrep: A robust preprocessing pipeline for functional MRI. Nature Methods, 16(1), 111–116. 10.1038/s41592-018-0235-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Finn, E. S. (2020). Idiosynchrony: From shared responses to individual differences during naturalistic neuroimaging. NeuroImage, 216, 116828. 10.1016/j.neuroimage.2020.116828 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Fouragnan, E., Retzler, C., & Bhatt, M. A. (2018). Separate neural representations of prediction error valence and surprise: Evidence from an fMRI meta-analysis. Human Brain Mapping, 39(7), 2887–2906. 10.1002/hbm.24047 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Freteault, M., Le Clei, M., Tetrel, L., Bellec, L., & Farrugia, N. (2025). Alignment of auditory artificial networks with massive individual fMRI brain data leads to generalisable improvements in brain encoding and downstream tasks. Imaging Neuroscience, 3 , imag_a_00525. 10.1162/imag_a_00525 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gonzalez-Castillo, J., Saad, Z. S., Handwerker, D. A., Inati, S. J., Brenowitz, N., & Bandettini, P. A. (2012). Whole-brain, time-locked activation with simple tasks revealed using massive averaging and model-free analysis. Proceedings of the National Academy of Sciences of the United States of America, 109(14), 5487–5492. 10.1073/pnas.1121049109 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gorgolewski, K. J., Auer, T., Calhoun, V. D., Craddock, R. C., Das, S., Duff, E. P., Flandin, G., Ghosh, S. S., Glatard, T., Halchenko, Y. O., Handwerker, D. A., Hanke, M., Keator, D., Li, X., Michael, Z., Maumet, C., Nichols, B. N., Nichols, T. E., Pellman, J., … Poldrack, R. A. (2016). The brain imaging data structure, a format for organizing and describing outputs of neuroimaging experiments. Scientific Data, 3(1), Article 1. 10.1038/sdata.2016.44 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Han, D. H., Lee, Y. S., Na, C., Ahn, J. Y., Chung, U. S., Daniels, M. A., Haws, C. A., & Renshaw, P. F. (2009). The effect of methylphenidate on Internet video game play in children with attention-deficit/hyperactivity disorder. Comprehensive Psychiatry, 50(3), 251–256. 10.1016/j.comppsych.2008.08.011 [ DOI ] [ PubMed ] [ Google Scholar ] Harel, Y., Cyr, A., Boyle, J., Pinsard, B., Bernard, J., Fourcade, M.-F., Aggarwal, H., Ponce, A. F., Thirion, B., Jerbi, K., & Bellec, P. (2023). Open design of a reproducible videogame controller for MRI and MEG. PLoS One, 18(11), e0290158. 10.1371/journal.pone.0290158 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Harel, Y., Bellec, L. P., Paugam, F., Delhaye, H., & Durand, A. (2025). Human-AI alignment of learning trajectories in video games: A continual RL benchmark proposal. In Reinforcement Learning and Video Games Workshop @ RLC 2025. https://openreview.net/pdf?id=YAVB439L9X Hasson, U., Nir, Y., Levy, I., Fuhrmann, G., & Malach, R. (2004). Intersubject synchronization of cortical activity during natural vision. Science, 303(5664), 1634–1640. 10.1126/science.1089506 [ DOI ] [ PubMed ] [ Google Scholar ] Haxby, J. V., Gobbini, M. I., Furey, M. L., Ishai, A., Schouten, J. L., & Pietrini, P. (2001). Distributed and overlapping representations of faces and objects in ventral temporal cortex. Science, 293(5539), 2425–2430. 10.1126/science.1063736 [ DOI ] [ PubMed ] [ Google Scholar ] Hebart, M. N., Contier, O., Teichmann, L., Rockter, A. H., Zheng, C. Y., Kidder, A., Corriveau, A., Vaziri-Pashkam, M., & Baker, C. I. (2023). THINGS-data, a multimodal collection of large-scale datasets for investigating object representations in human brain and behavior. eLife, 12, e82580. 10.7554/eLife.82580 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Herholz, P., Fortier, E., Toneva, M., Farrugia, N., Wehbe, L., & Borghesani, V. (2022). A roadmap to reverse engineering real-world generalization by combining naturalistic paradigms, deep sampling, and predictive computational models. arXiv (No. arXiv:2108.10231). http://arxiv.org/abs/2108.10231 Hsu, C.-T., Clariana, R., Schloss, B., & Li, P. (2019). Neurocognitive signatures of naturalistic reading of scientific texts: A fixation-related fMRI study. Scientific Reports, 9(1), Article 1. 10.1038/s41598-019-47176-7 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Huettel, S. A. (2012). Event-related fMRI in cognition. NeuroImage, 62(2), 1152–1156. 10.1016/j.neuroimage.2011.08.113 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hussein, A., Gaber, M. M., Elyan, E., & Jayne, C. (2017). Imitation learning: A survey of learning methods. ACM Computing Surveys (CSUR), 50(2), 1–35. https://dl.acm.org/doi/10.1145/3054912 [ Google Scholar ] Inoue, Y., & Ashida, K. (1993). Controller for a game machine (United States Patent No. US5207426A). https://patents.google.com/patent/US5207426A/en Jääskeläinen, I. P., Sams, M., Glerean, E., & Ahveninen, J. (2021). Movies and narratives as naturalistic stimuli in neuroimaging. NeuroImage, 224, 117445. 10.1016/j.neuroimage.2020.117445 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Jimura, K., & Poldrack, R. A. (2012). Analyses of regional-average activation and multivoxel pattern information tell complementary stories. Neuropsychologia, 50(4), 544–552. 10.1016/j.neuropsychologia.2011.11.007 [ DOI ] [ PubMed ] [ Google Scholar ] Kätsyri, J., Hari, R., Ravaja, N., & Nummenmaa, L. (2013a). Just watching the game ain’t enough: Striatal fMRI reward responses to successes and failures in a video game during active and vicarious playing. Frontiers in Human Neuroscience, 7. https://www.frontiersin.org/articles/10.3389/fnhum.2013.00278 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kätsyri, J., Hari, R., Ravaja, N., & Nummenmaa, L. (2013b). The opponent matters: Elevated fMRI reward responses to winning against a human versus a computer opponent during interactive video game playing. Cerebral Cortex, 23(12), 2829–2839. 10.1093/cercor/bhs259 [ DOI ] [ PubMed ] [ Google Scholar ] Kemtur, A., Paugam, F., Pinsard, B., Harel, Y., Sainath, P., Clei, M. L., Boyle, J., Jerbi, K., & Bellec, P. (2023). Behavioral imitation with Artificial Neural networks leads to personalized models of brain dynamics during videogame play. bioRxiv, 2023.10.28.564546. 10.1101/2023.10.28.564546 [ DOI ] [ PMC free article ] [ PubMed ] Khoshnoud, S., Igarzábal, F. A., & Wittmann, M. (2020). Peripheral-physiological and neural correlates of the flow experience while playing video games: A comprehensive review. PeerJ, 8, e10520. 10.7717/peerj.10520 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Khoshnoud, S., Shamsi, M., Nazari, M. A., & Makeig, S. (2022). Brain and mind state markers during flow and its disruption. Frontiers in Psychology, 13, 1010261. 10.3389/fpsyg.2022.1010261 [ DOI ] [ Google Scholar ] Li, L., Chen, R., & Chen, J. (2016). Playing action video games improves visuomotor control. Psychological Science, 27(8), 1092–1108. 10.1177/0956797616650300 [ DOI ] [ PubMed ] [ Google Scholar ] Libretro. (n.d.). Libretro API. Retrieved January 23, 2025, from https://www.libretro.com/index.php/api/ Lim, S. J., Fiez, J. A., & Holt, L. L. (2019). Role of the striatum in incidental learning of sound categories. Proceedings of the National Academy of Sciences, 116(10), 4671–4680. 10.1073/pnas.1811992116 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Lynch, C. J., Voss, H. U., Silver, B. M., & Power, J. D. (2021). On measuring head motion and effects of head molds during fMRI. NeuroImage, 225, 117494. 10.1016/j.neuroimage.2020.117494 [ DOI ] [ PubMed ] [ Google Scholar ] Martino, J., Gabarrós, A., Deus, J., Juncadella, M., Acebes, J. J., Torres, A., & Pujol, J. (2011). Intrasurgical mapping of complex motor function in the superior frontal gyrus. Neuroscience, 179, 131–142. 10.1016/j.neuroscience.2011.01.047 [ DOI ] [ PubMed ] [ Google Scholar ] Mathiak, K., & Weber, R. (2006). Toward brain correlates of natural behavior: fMRI during violent video games. Human Brain Mapping, 27(12), 948–956. 10.1002/hbm.20234 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Matthews, W. J., & Gheorghiu, A. I. (2016). Repetition, expectation, and the perception of time. Current Opinion in Behavioral Sciences, 8, 110–116. 10.1016/j.cobeha.2016.02.019 [ DOI ] [ Google Scholar ] Margulies, D. S., Vincent, J. L., Kelly, C., Lohmann, G., Uddin, L. Q., Biswal, B. B., Villringer, A., Castellanos, F. X., Milham, M. P., & Petrides, M. (2009). Precuneus shares intrinsic functional architecture in humans and monkeys. Proceedings of the National Academy of Sciences of the United States of America, 106(47), 20069–20074. 10.1073/pnas.0905314106 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Matusz, P. J., Dikker, S., Huth, A. G., & Perrodin, C. (2019). Are we ready for real-world neuroscience? Journal of Cognitive Neuroscience, 31(3), 327–338. 10.1162/jocn_e_01276 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Mumford, J. A., Poline, J.-B., & Poldrack, R. A. (2015). Orthogonalization of regressors in fMRI models. PLoS One, 10(4), e0126255. 10.1371/journal.pone.0126255 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nachev, P., Kennard, C., & Husain, M. (2008). Functional role of the supplementary and pre-supplementary motor areas. Nature Reviews Neuroscience, 9(11), Article 11. 10.1038/nrn2478 [ DOI ] [ PubMed ] [ Google Scholar ] Nastase, S. A., Goldstein, A., & Hasson, U. (2020). Keep it real: Rethinking the primacy of experimental control in cognitive neuroscience. NeuroImage, 222, 117254. 10.1016/j.neuroimage.2020.117254 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nichol, A., Pfau, V., Hesse, C., Klimov, O., & Schulman, J. (2018). Gotta learn fast: A new benchmark for generalization in RL. arXiv (No. arXiv:1804.03720). 10.48550/arXiv.1804.03720 [ DOI ] Notter, M. P., Gale, D., Herholz, P., Markello, R., Notter-Bielser, M.-L., & Whitaker, K. (2019). AtlasReader: A Python package to generate coordinate tables, region labels, and informative figures from statistical MRI images. Journal of Open Source Software, 4(34), 1257. 10.21105/joss.01257 [ DOI ] [ Google Scholar ] OpenAI. (2018). OpenAI Retro Contest. https://openai.com/index/retro-contest/ Orban, P., Doyon, J., Petrides, M., Mennes, M., Hoge, R., & Bellec, P. (2015). The richness of task-evoked hemodynamic responses defines a pseudohierarchy of functionally meaningful brain networks. Cerebral Cortex, 25(9), 2658–2669. 10.1093/cercor/bhu064 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Peirce, J., Gray, J. R., Simpson, S., MacAskill, M., Höchenberger, R., Sogo, H., Kastman, E., & Lindeløv, J. K. (2019). PsychoPy2: Experiments in behavior made easy. Behavior Research Methods, 51(1), 195–203. 10.3758/s13428-018-01193-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Peñuelas-Calvo, I., Jiang-Lin, L. K., Girela-Serrano, B., Delgado-Gomez, D., Navarro-Jimenez, R., Baca-Garcia, E., & Porras-Segovia, A. (2022). Video games for the assessment and treatment of attention-deficit/hyperactivity disorder: A systematic review. European Child & Adolescent Psychiatry, 31(1), 5–20. 10.1007/s00787-020-01557-w [ DOI ] [ PubMed ] [ Google Scholar ] Pinho, A. L., Amadon, A., Ruest, T., Fabre, M., Dohmatob, E., Denghien, I., Ginisty, C., Becuwe-Desmidt, S., Roger, S., Laurier, L., Joly-Testault, V., Médiouni-Cloarec, G., Doublé, C., Martins, B., Pinel, P., Eger, E., Varoquaux, G., Pallier, C., Dehaene, S., … Thirion, B. (2018). Individual brain charting, a high-resolution fMRI dataset for cognitive mapping. Scientific Data, 5(1), 180105. 10.1038/sdata.2018.105 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Platt, M., & Padoa-Schioppa, C. (2009). Chapter 29—Neuronal representations of value. In Glimcher P. W., Camerer C. F., Fehr E., & Poldrack R. A. (Eds.), Neuroeconomics (pp. 441–462). Academic Press. 10.1016/B978-0-12-374176-9.00029-4 [ DOI ] [ Google Scholar ] Poline, J.-B., & Brett, M. (2012). The general linear model and fMRI: Does love last forever? NeuroImage, 62(2), 871–880. 10.1016/j.neuroimage.2012.01.133 [ DOI ] [ PubMed ] [ Google Scholar ] Popa, L. S., & Ebner, T. J. (2019). Cerebellum, predictions and errors. Frontiers in Cellular Neuroscience, 12, 524. https://www.frontiersin.org/articles/10.3389/fncel.2018.00524 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Porter, A., Nielsen, A., Dorn, M., Dworetsky, A., Edmonds, D., & Gratton, C. (2023). Masked features of task states found in individual brain networks. Cerebral Cortex (New York, N.Y.: 1991), 33(6), 2879–2900. 10.1093/cercor/bhac247 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Potter, S. M., El Hady, A., & Fetz, E. E. (2014). Closed-loop neuroscience and neuroengineering. Frontiers in Neural Circuits, 8, 115. https://www.frontiersin.org/articles/10.3389/fncir.2014.00115 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Provins, C., MacNicol, E., Seeley, S. H., Hagmann, P., & Esteban, O. (2023). Quality control in functional MRI studies with MRIQC and fMRIPrep. Frontiers in Neuroimaging, 1, 1073734. 10.3389/fnimg.2022.1073734 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Rastegarnia, S., St-Laurent, M., DuPre, E., Pinsard, B., & Bellec, P. (2023). Brain decoding of the Human Connectome Project tasks in a dense individual fMRI dataset. NeuroImage, 283, 120395. 10.1016/j.neuroimage.2023.120395 [ DOI ] [ PubMed ] [ Google Scholar ] Rolls, E. T. (2019). The cingulate cortex and limbic systems for emotion, action, and memory. Brain Structure and Function, 224(9), 3001–3018. 10.1007/s00429-019-01945-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Scheliga, S., Kellermann, T., Lampert, A., Rolke, R., Spehr, M., & Habel, U. (2022). Neural correlates of multisensory integration in the human brain: An ALE meta-analysis. Reviews in the Neurosciences, 34(2), 223–245. 10.1515/revneuro-2022-0065 [ DOI ] [ PubMed ] [ Google Scholar ] Schoeneman, C., & Riker, T.. BZFlag (Battle Zone capture Flag). http://www.bzflag.org Schrodt, F., Kneissler, J., Ehrenfeld, S., & Butz, M. V. (2017). Mario becomes cognitive. Topics in Cognitive Science, 9(2), 343–373. 10.1111/tops.12252 [ DOI ] [ PubMed ] [ Google Scholar ] Schweizer, T., Kan, K., Hung, Y., Tam, F., Naglie, G., & Graham, S. (2013). Brain activity during driving with distraction: An immersive fMRI study. Frontiers in Human Neuroscience, 7, 53. https://www.frontiersin.org/articles/10.3389/fnhum.2013.00053 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Sega. (1993). Shinobi III: Return of the Ninja Master [Computer software]. Sega. https://en.wikipedia.org/wiki/Shinobi_III:_Return_of_the_Ninja_Master [ Google Scholar ] Setsompop, K., Gagoski, B. A., Polimeni, J. R., Witzel, T., Wedeen, V. J., & Wald, L. L. (2012). Blipped-controlled aliasing in parallel imaging for simultaneous multislice echo planar imaging with reduced g-factor penalty. Magnetic Resonance in Medicine, 67(5), 1210–1224. 10.1002/mrm.23097 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Smith, G., Cha, M., & Whitehead, J. (2008). A framework for analysis of 2D platformer levels. In Proceedings of the 2008 ACM SIGGRAPH Symposium on Video Games (pp. 75–80). ACM. 10.1145/1401843.1401858 [ DOI ] [ Google Scholar ] St-Laurent, M., Pinsard, B., Contier, O., DuPre, E., Seeliger, K., Borghesani, V., Boyle, J. A., Bellec, L., & Hebart, M. N. (2026). CNeuroMod-THINGS, a densely-sampled fMRI dataset for visual neuroscience. Scientific Data, 13, 141. 10.1038/s41597-026-06591-y [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tanaka, M., Kunimatsu, J., Suzuki, T. W., Kameda, M., Ohmae, S., Uematsu, A., & Takeya, R. (2021). Roles of the cerebellum in motor preparation and prediction of timing. Neuroscience, 462, 220–234. 10.1016/j.neuroscience.2020.04.039 [ DOI ] [ PubMed ] [ Google Scholar ] Tsivilis, D., Otten, L. J., & Rugg, M. D. (2003). Repetition effects elicited by objects and their contexts: An fMRI study. Human Brain Mapping, 19(3), 145–154. 10.1002/hbm.10116 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tzvi, E., Loens, S., & Donchin, O. (2022). Mini-review: The role of the cerebellum in visuomotor adaptation. The Cerebellum, 21(2), 306–313. 10.1007/s12311-021-01281-4 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Vanderwal, T., Eilbott, J., & Castellanos, F. X. (2019). Movies in the magnet: Naturalistic paradigms in developmental functional neuroimaging. Developmental Cognitive Neuroscience, 36, 100600. 10.1016/j.dcn.2018.10.004 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Vanderwal, T., Eilbott, J., Finn, E. S., Craddock, R. C., Turnbull, A., & Castellanos, F. X. (2017). Individual differences in functional connectivity during naturalistic viewing conditions. NeuroImage, 157, 521–530. 10.1016/j.neuroimage.2017.06.027 [ DOI ] [ PubMed ] [ Google Scholar ] van Kemenade, B. M., Arikan, B. E., Kircher, T., & Straube, B. (2019). Distinct roles for the cerebellum, angular gyrus, and middle temporal gyrus in action–feedback monitoring. Cerebral Cortex, 29(4), 1520–1531. 10.1093/cercor/bhy042 [ DOI ] [ PubMed ] [ Google Scholar ] Wallenius, K. (2010). Video annotation for studying the brain in naturalistic settings. Aalto University. http://lib.tkk.fi/Dipl/2010/urn100252.pdf [ Google Scholar ] Wang, H.-T., Meisler, S. L., Sharmarke, H., Clarke, N., Gensollen, N., Markiewicz, C. J., Paugam, F., Thirion, B., & Bellec, P. (2024). Continuous evaluation of denoising strategies in resting-state fMRI connectivity using fMRIPrep and Nilearn. PLoS Computational Biology, 20(3), e1011942. 10.1371/journal.pcbi.1011942 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Wang, M., Dong, G., Wang, L., Zheng, H., & Potenza, M. N. (2018). Brain responses during strategic online gaming of varying proficiencies: Implications for better gaming. Brain and Behavior, 8(8), e01076. 10.1002/brb3.1076 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Warbrick, T., Mobascher, A., Brinkmeyer, J., Musso, F., Richter, N., Stoecker, T., Fink, G. R., Shah, N. J., & Winterer, G. (2009). Single-trial P3 amplitude and latency informed event-related fMRI models yield different BOLD response patterns to a target detection task. NeuroImage, 47(4), 1532–1544. 10.1016/j.neuroimage.2009.05.082 [ DOI ] [ PubMed ] [ Google Scholar ] Weber, R., Ritterfeld, U., & Mathiak, K. (2006). Does playing violent video games induce aggression? Empirical evidence of a functional magnetic resonance imaging study. Media Psychology, 8(1), 39–60. 10.1207/S1532785XMEP0801_4 [ DOI ] [ Google Scholar ] Wehbe, L., Murphy, B., Talukdar, P., Fyshe, A., Ramdas, A., & Mitchell, T. (2014). Simultaneously uncovering the patterns of brain regions involved in different story reading subprocesses. PLoS One, 9(11), e112575. 10.1371/journal.pone.0112575 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zhang, T., Gao, J. S., Çukur, T., & Gallant, J. L. (2021). Voxel-based state space modeling recovers task-related cognitive states in naturalistic fMRI experiments. Frontiers in Neuroscience, 14, 565976. 10.3389/fnins.2020.565976 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Zhang, Y., Luo, Q., Gong, T., Li, L., Li, Y., Hou, F., He, Y., & Xu, F. (2023). The hierarchical organization of the precuneus captured by functional gradients. Brain Structure and Function, 228(7), 1993–2007. 10.1007/s00429-023-02672-5 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Material IMAG.a.1256_supp.zip (921.1MB, zip) Data Availability Statement All datasets used in this study are part of the CNeuroMod dataset ( https://www.cneuromod.ca/ ). The code developed for this research is publicly available on GitHub ( https://github.com/courtois-neuromod/shinobi_fmri ) and archived on Zenodo for long-term preservation ( https://doi.org/10.5281/zenodo.15628556 ). Processed datasets used to generate the tables and figures in this manuscript are provided as Supplementary Material accompanying the paper. Articles from Imaging Neuroscience are provided here courtesy of MIT Press ACTIONS View on publisher site PDF (8.0 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Analysis of psychomotor latency, spatial reasoning, and metabolic health markers in adolescent gamers: A comparative study of the impact of the confounding triad of sleep, diet, and screen modality.#476913
• ACTing Minds: An ACT-based Videogame for Mental Health#852685
• Efficacy of a Microalgae Extract PhaeoSOL Combined With Natural Stimulant on Cognitive Function and Gaming Performance of Video Gamers#849307
• Minds Systems: Functional Monitoring and Contingent Intervention in High-Performance Sports#29862
• Framing the Gamer's Dilemma#699045