Background & Summary

Anesthesia is fundamental to modern medical practice, enabling millions of surgical procedures globally each year1. Since the first public demonstration in 1846, anesthesiology has significantly evolved, enhancing both patient safety and procedural efficacy. Considerable progress has also been made in understanding the neurobiological mechanisms by which anesthetics reversibly suppress consciousness and memory2. Furthermore, anesthetic agents uniquely offer safe, reproducible, and titratable modulation of consciousness, providing a powerful experimental framework for exploring the neural basis of human consciousness and cognition.

Functional magnetic resonance imaging (fMRI) has significantly advanced our understanding of brain function by enabling detailed, non-invasive investigation of neural responses. Early studies demonstrated that anesthetics cause a dose-dependent reduction in brain activation in response to external stimuli3. Subsequent research showed that anesthesia preferentially disrupts activity in higher-order association cortices while preserving activity in primary sensory areas4,5,6,7,8,9,10. These findings informed the concept of “information received but not perceived,” suggesting a disconnection between sensory input and conscious perception under anesthesia11. The introduction of resting-state fMRI and functional connectivity analyses12,13 further enhanced the field by enabling the mapping of intrinsic brain networks, thereby deepening our understanding of the neural mechanisms underlying anesthesia-induced unconsciousness14,15,16.

Although resting-state and sensory-evoked paradigms have been useful, they are limited in their ability to detect covert consciousness. To overcome this limitation, we adopted well-validated mental imagery tasks17 to investigate volitional cognitive processes during graded sedation in healthy human volunteers. Additionally, this controlled and reproducible sedation model allowed us to systematically dissect layers of brain functionality and identify the neural networks underlying conscious processing.

Previous analyses of this dataset have identified instances where participants maintained volitional mental imagery despite behavioral unresponsiveness under sedation18. Subsequent investigations revealed that the anterior insula plays a critical role in regulating transitions between brain network states19 and that there are asymmetric neural dynamics associated with the loss and recovery of consciousness20.

In alignment with open science principles21,22, we are making this dataset publicly accessible via OpenNeuro23. All data conform to Brain Imaging Data Structure (BIDS) standards and have undergone standardized preprocessing with fMRIPrep24,25,26, providing a robust, reproducible, and unique resource for further exploration into the neural mechanisms of anesthesia and consciousness.

Methods

General description

This study, funded by NIH grant R01-GM103894 and approved by the University of Michigan Institutional Review Board, involved collecting fMRI data from 26 healthy volunteers between May 2017 and July 2019. All procedures were conducted following relevant ethical guidelines and regulations. Participants completed mental imagery tasks (tennis, spatial navigation, and hand squeeze) along with a motor response task (actual hand squeeze), under varying levels of propofol sedation. The experimental paradigm employed a pseudo-randomized block design, alternating 15-second intervals of mental imagery tasks with 15-second rest intervals. Behavioral responsiveness was continuously assessed by measuring hand-squeeze force during motor tasks, providing objective markers for loss and recovery of responsiveness.

Participants

Twenty-six healthy adults (aged 19–34 years; 13 women) participated in the study after detailed explanation of the procedures and provision of written informed consent. The consent explicitly covered both participation in the research procedures and the sharing of de-identified data for future research, consistent with NIH Certificate of Confidentiality protections. Each participant was assigned a unique numeric identifier to maintain confidentiality across all data types, including behavioral, physiological, and MRI records. Sex was recorded based on self-report, with the understanding that this may not fully reflect biological sex. Gender identity information was not collected. Data are presented disaggregated by sex to support transparency and enable future exploratory analyses.

Recruitment

Healthy volunteers were recruited via postings on UMClinicalStudies.org and through community notices at local colleges and organizations in Ann Arbor. Interested volunteers underwent an initial telephone screening, which included assessments of medical history, demographic information, handedness, and criteria specific to MRI procedures. Responses were reviewed by the study team, and participants underwent an additional onsite medical evaluation by an attending anesthesiologist to confirm their eligibility. Once eligibility was confirmed, participants were scheduled for a single research session. Individuals with psychiatric or neurological disorders (current or historical) were excluded from participation.

Inclusion and exclusion criteria

Eligible participants were healthy, right-handed adults (ages: 18–40 years) classified as American Society of Anesthesiologists physical status 1, with a body mass index (BMI) below 30, and prior experience in racquet sports (minimum of 30 lifetime sessions). Exclusion criteria included inability to speak English, contraindications to MRI scanning (such as pregnancy, obesity, metallic implants, claustrophobia), anxiety disorders, cardiopulmonary conditions, history of neurological or cardiovascular illnesses, significant head trauma with loss of consciousness, learning disabilities, developmental disorders, severe snoring or sleep apnea, significant sensory or motor impairments, gastroesophageal reflux disease, recent alcohol or substance use (within 24 hours prior to scanning), history of drug use or positive drug screening, tattoos on the head or neck region, egg allergies, structural abnormalities detected on T1-weighted MRI, or experiencing significant discomfort during fMRI scanning.

Anesthetic administration and monitoring

Propofol was selected as the anesthetic agent due to its widespread use and favorable profile for human fMRI studies. Its minimal effects on cerebral hemodynamics, particularly the preservation of cerebral blood flow-metabolism coupling, make it suitable for interpreting fMRI data accurately27. Propofol primarily suppresses neuronal activity by enhancing GABA-A receptor-mediated inhibition, influencing extensive neural circuits throughout the brain28. Safety profiles in healthy individuals have demonstrated no adverse cognitive or arousal-related outcomes even after prolonged exposure to surgical anesthesia without surgery, with cognition typically returning to baseline within three hours post-emergence29.

All participants underwent an 8-hour fasting period before the study. On the day of the experiment, an attending anesthesiologist conducted a thorough pre-operative evaluation and physical examination. Two anesthesiologists were continuously present throughout the experimental session. An intravenous cannula was inserted following subcutaneous injection of local anesthetic (1% lidocaine; 0.3–0.5 ml). Continuous monitoring was conducted for spontaneous respiration, end-tidal CO2, heart rate, pulse oximetry, and electrocardiogram. Non-invasive blood pressure was measured using an MRI-compatible automatic device. Supplemental oxygen (2 L/min via nasal cannula) was provided to all participants.

Propofol administration employed a target-controlled intravenous bolus followed by a continuous rate infusion, guided by the pharmacokinetic Marsh model30 implemented in the STANPUMP software (http://opentci.org/code/stanpump). Dosages (bolus and infusion rates) were incrementally increased every five minutes by increments of 0.4 μg/ml until the desired sedation level was achieved, marked by the loss of behavioral responsiveness (LOR). Initial target effect-site concentrations (ESC) were set at either 0.4 μg/ml (14 participants) or 1.0 μg/ml (12 participants). The final ESC was either 2.4 μg/ml (for 6 participants previously studied18) or set at one increment (0.4 μg/ml) above the ESC at which participants first lost responsiveness (20 participants). For instance, if a participant experienced LOR at an ESC of 2.0 μg/ml, the final target ESC was 2.4 μg/ml. This strategy minimized agitation and associated head movements, thus improving the quality of the fMRI data. The final ESC was maintained for approximately 21.6 minutes on average (±SD = 10.2 minutes), after which propofol infusion was stopped to allow spontaneous recovery. A detailed table listing each participant’s final ESC is available in derivatives/Participant_Info.xlsx within the dataset, allowing users to examine individual variability in anesthetic depth.

Experimental task

Participants performed mental imagery and motor response tasks before, during, and after the incremental propofol sedation. Three mental imagery tasks were used: tennis imagery, navigation imagery, and hand-squeeze imagery, alongside an actual motor response task (hand squeeze). For tennis imagery, participants were instructed to visualize themselves standing stationary on a tennis court, swinging their arm to return a ball to an imagined instructor. In the navigation imagery task, they were asked to mentally visualize walking through the streets of a familiar city or their home, vividly imagining the details they would observe. For the squeeze imagery task, participants imagined squeezing an MRI-compatible rubber ball. This was followed by a motor response task where participants physically squeezed the ball (Fig. 1).

Fig. 1
figure 1

Experimental protocol. Propofol was administered with incrementally increasing target effect-site concentrations (ESCs), rising by 0.4 mg/mL every five minutes. Infusion continued until participants reached loss of behavioral responsiveness (LOR). The final target concentration, set one increment above the LOR point, was maintained for approximately 22 minutes. After this, the infusion was stopped to allow for spontaneous recovery. Mental imagery and motor response tasks were conducted before, during, and after propofol infusion. Participants performed three imagery tasks (tennis, navigation, and hand squeeze) and an actual hand squeeze motor response task. The timing of “action” instructions and motor responses helped define three key periods: PreLOR (before loss of responsiveness), LOR (during loss of responsiveness), and ROR (recovery of responsiveness). Baseline data included two 10-minute resting-state recordings (Rest 1 and Rest 2) and two 15-minute task-based recordings (Imagery 1 and Imagery 4).

Full size image

The tasks were organized using a pseudo-randomized (Latin square) block design. Each experimental cycle consisted of alternating periods: 15 seconds of tennis or navigation imagery, 10 seconds of squeeze imagery followed by a hand squeeze performed within 5 seconds after instruction, and 15 seconds of rest. Each scanning session comprised 180 cycles (60 per condition). Auditory cues (“tennis imagery,” “navigation imagery,” “squeeze imagery,” or “action”) marked the beginning of each task period, with “relax” indicating rest periods.

Verbal instructions were delivered using E-Prime 3.0 software (Psychology Software Tools, Pittsburgh, PA) through an MRI-compatible audiovisual presentation system. Headphone volumes were adjusted individually for participant comfort. Behavioral responses (hand squeeze force) were recorded in mmHg using the BIOPAC MP160 system with AcqKnowledge software (version 5.0). By aligning the timing of “action” instructions with actual motor responses during and after propofol infusion, distinct phases of participant responsiveness were identified: retention of responsiveness (PreLOR), loss of responsiveness (LOR), and recovery of responsiveness (ROR). Specifically, the offset of PreLOR, onset of LOR, offset of LOR, and onset of ROR were defined by the timing of the last successful squeeze, first failure to squeeze, final unsuccessful squeeze attempt, and first successful squeeze response after LOR, respectively.

Imaging data acquisition

Data were collected at the University of Michigan Hospital using a 3 T Philips scanner equipped with a standard 32-channel transmit/receive head coil. Prior to functional imaging, high-resolution anatomical images were obtained using a T1-weighted spoiled gradient recalled echo (SPGR) sequence with the following parameters: 170 sagittal slices, slice thickness = 1.0 mm (no gap), repetition time (TR) = 8.1 s, echo time (TE) = 3.7 ms, flip angle = 8°, field of view (FOV) = 240 mm, and a 256 × 256 image matrix (Table 1).

Table 1 MRI Acquisition Parameters.
Full size table

Whole-brain functional images were acquired using a gradient-echo echo-planar imaging (EPI) sequence with parameters: 28 slices, TR/TE = 800/25 ms via multiband (MB) acquisition (MB factor = 4), slice thickness = 4 mm, in-plane resolution = 3.4 × 3.4 mm, FOV = 220 mm, flip angle = 76°, and a 64 × 64 image matrix. Prior to MRI hardware upgrades, six participants were scanned with slightly modified parameters: 21 slices, TR/TE = 800/25 ms, MB factor = 3, and slice thickness = 6 mm (Table 1).

All participants completed two resting-state fMRI scans, each lasting 10 minutes—one at the beginning (Rest 1) and one at the end (Rest 2) of the session. During these scans, participants were instructed to remain awake, still, and with their eyes closed while receiving verbal cues through earphones.

Task-based fMRI data were acquired across four runs: a 15-minute awake baseline (Imagery 1), a 30-minute scan during propofol sedation (Imagery 2), a 30-minute post-infusion scan (Imagery 3), and a 15-minute recovery baseline (Imagery 4). The duration and structure of all scans were consistent across participants, and stimulus timing files for all runs are available under derivatives/Stimulus_Timing/ in the dataset.

Behavioral responsiveness varied across sedation levels as expected. Participant sub-16 became unresponsive after Imagery 2, while five participants (sub-14, sub-18, sub-22, sub-24, and sub-28) regained responsiveness after Imagery 3. Exact time points of these transitions were not recorded. Participant sub-29 did not complete the study due to excessive motion during recovery, resulting in missing data for Imagery 4 and Rest 2.

Preprocessing

Preprocessing of all fMRI scans was conducted using fMRIPrep (version 23.2.1)24,25,26, which is based on Nipype31,32 (version 1.8.6) and utilizes tools from Nilearn33 (version 0.10.2). Each preprocessing step is described in detail below.

Anatomical data preprocessing

For each participant, a single T1-weighted (T1w) structural image was acquired and processed using standard neuroimaging workflows. Intensity non-uniformity was corrected with N4BiasFieldCorrection34 implemented in ANTs35 (v2.5.0), followed by skull stripping via the antsBrainExtraction.sh routine using the OASIS30ANTs template. Brain tissue was segmented into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) compartments using FSL FAST36 (v6.0.7.7). Cortical surfaces were reconstructed with FreeSurfer37 (recon-all, v7.3.2), and the brain mask was refined through a customized adaptation of the Mindboggle38 method to align ANTs- and FreeSurfer-derived gray matter boundaries. The resulting images were registered to both native anatomical (T1w) and standard MNI space (MNI152NLin2009cAsym, via TemplateFlow39,40, v23.1.0) using nonlinear spatial normalization with ANTs antsRegistration.

The anatomical derivatives from the above preprocessing pipeline can be found under each subject’s anat sub-folder (location: derivatives/fmriprep_output/sub- < subject_ID > /anat/).

Functional data preprocessing

After preprocessing the anatomical data, fMRIPrep performs a series of steps to preprocess each subject’s functional BOLD data. For each of the six BOLD runs acquired per subject (four mental imagery scans and two resting-state scans), the following preprocessing was performed. First, a reference BOLD volume was generated using a custom fMRIPrep methodology, to be used for head motion correction. Head-motion parameters with respect to the BOLD reference (volume-to-reference transformation matrices, as well as six corresponding rotation and translation parameters) were then estimated using FSL’s mcflirt41 before any spatiotemporal filtering. To account for the differences in data acquisition timing across different slices, slice timing correction was also performed using AFNI’s (version 24.0.05) 3dTshift tool42. Afterwards, the BOLD reference was co-registered to the T1w image using FreeSurfer’s mri_coreg, which implements a fast, rigid linear registration configured with six degrees of freedom; this co-registration transform was in turn used to resample the preprocessed BOLD time-series into the designated output spaces (native anatomical T1w space and standard MNI152NLin2009cAsym space).

Several confounding time-series were subsequently extracted from the preprocessed BOLD signal. Although fMRIPrep does not denoise the BOLD scans, such confounders can be regressed from the signal using other software (see the section titled “Denoising” below). First, the framewise displacement (FD), temporal derivative of the root-mean-squared variance over voxels (DVARS), and three region-wise global signals were calculated for each functional run. FD was computed using two formulations (absolute sum of relative motions43, and relative root mean square displacement between affines41). FD and DVARS were calculated for each functional run, using their implementations in Nipype (following the definitions in reference43). Besides these time-series, the three global signals within the CSF, WM, and whole-brain masks were also extracted. Head motion parameters were additionally computed, including six rigid-body motion parameters (three translations and three rotations), and output within the corresponding confounds file. The reported confounding time-series derived from the global signals and head motion estimates further included the temporal derivatives and quadratic terms for each parameter44, to capture a larger degree of variance related to head motion.

Physiological noise correction was implemented using the CompCor45 approach, which identifies components representing non-neuronal signal fluctuations in regions unlikely to contain gray matter. Principal components were derived from preprocessed BOLD time series after high-pass filtering (128 s cutoff). Both anatomical (aCompCor) and temporal (tCompCor) variants were computed. For aCompCor, probabilistic masks of white matter (WM), cerebrospinal fluid (CSF), and their combined regions were generated in anatomical space. To minimize gray matter contamination, voxels with substantial gray matter volume fraction—determined from dilated FreeSurfer-derived GM masks—were excluded. The masks were then transformed into BOLD space and binarized at a 0.99 probability threshold. For tCompCor, components were extracted from the top 2% most variable voxels within the brain mask. For each variant, the number of components retained was determined such that they collectively explained at least 50% of the variance within the respective nuisance mask. Additional regressors were obtained from a narrow peripheral “crown” of voxels surrounding the brain boundary, providing an extra measure of artifact control46.

All aforementioned confounding time-series were saved by fMRIPrep as a TSV file (sub- < subject_ID > _[specifiers]_desc-confounds_timeseries.tsv) that can be used to denoise the BOLD signal time-series. The functional derivatives from the above preprocessing pipelines can be found under each subject’s func sub-folder (location: derivatives/fmriprep_output/sub- < subject_ID > /func/). For a summary of each subject’s anatomical and functional preprocessing steps, accompanied by extensive visual assessments, the interested reader may open the corresponding sub- < subject_ID > .html file found under derivatives/fmriprep_output/.

For more details on fMRIPrep’s preprocessing pipelines, please refer to the associated manuscript24 as well as excellent online documentation: https://fmriprep.org/en/latest/workflows.html. For a detailed description on each one of fMRIPrep’s outputs and corresponding naming structure, please refer to: https://fmriprep.org/en/latest/outputs.html.

Denoising

Post-processing of fMRIPrep outputs

The eXtensible Connectivity Pipeline-DCAN (XCP-D)44,47,48 (version 0.10.6) was used to post-process the outputs of fMRIPrep. XCP-D was built with Nipype31 (version 1.9.2).

Segmentations

The following brain atlases were implemented in this workflow: the Schaefer Supplemented with Subcortical Structures (4S) atlas at 10 different resolutions (156, 256, 356, 456, 556, 656, 756, 856, 956, and 1056 parcels)49,50,51,52,53, the Glasser atlas54, the Gordon atlas55, the Tian subcortical atlas56, and the HCP CIFTI subcortical atlas53. Each atlas was warped to the same space and resolution as the BOLD image being processed.

Anatomical data

Native-space T1w images were transformed to MNI152NLin2009cAsym space at 1 mm3 voxel resolution.

Functional data

For each of the six BOLD runs found per subject (four task-based BOLD runs and two resting-state BOLD runs), four complementary denoising algorithms were performed, based on whether to include or not global signal regression and on whether to band-pass or high-pass filter the BOLD signal time-series.

Considering the ongoing debate in the literature as to whether global signal regression should be performed57, we opted to perform two nuisance regression analyses to analyze our functional data:

  1. (1)

    In the first approach, 36 nuisance regressors were selected from the preprocessing confounds exported by fMRIPrep, according to the ‘36 P’ strategy44,58. These nuisance regressors included six motion parameters (n = 6), time-series corresponding to the first temporal derivatives of the six motion parameters (n = 6), together with their quadratic terms (n = 12). Additional confounders included the mean, temporal derivatives, and quadratic terms for the three signals (global signal, white matter signal, and cerebrospinal fluid signal) (n = 12). The BOLD data were then despiked using AFNI’s 3dDespike. Nuisance regressors were regressed from the BOLD data using Nilearn.

  2. (2)

    The same nuisance confounders were regressed from the BOLD data with the notable exception of the global signal’s mean, temporal derivative, and respective quadratic terms (total number of nuisance regressors left = 32).

For each nuisance regression approach above, we performed two filtering analyses, assessing the effect of frequency range considered within the BOLD signal:

  1. (1)

    In the first approach, the BOLD signal time-series were band-pass filtered using a second-order Butterworth filter, to retain signals between 0.008–0.1 Hz. The same filter was applied to the confounds. The resulting time-series were denoised via linear regression. The denoised BOLD was then smoothed using Nilearn with a Gaussian kernel (FWHM = 6 mm).

  2. (2)

    In the second approach, the BOLD signal time-series were instead high-pass filtered to retain all signals above 0.008 Hz.

The outputs from nuisance regression approach (1) and filtering approach (1) can be found under the folder: derivatives/xcp_d_with_GSR_bandpass_output; the outputs from nuisance regression approach (2) and filtering approach (1) can be found under the folder: derivatives/xcp_d_without_GSR_bandpass_output; the outputs from nuisance regression approach (1) and filtering approach (2) can be found under the folder: derivatives/xcp_d_with_GSR_highpass_output; and the outputs from nuisance regression approach (2) and filtering approach (2) can be found under the folder: derivatives/xcp_d_without_GSR_highpass_output.

Following preprocessing, the amplitude of low-frequency fluctuations (ALFF)59 was calculated from the denoised, mean-centered, and variance-normalized BOLD time series. The signal was transformed into the frequency domain using the Lomb–Scargle periodogram60,61,62,63, and the average square root of power within the 0.01–0.08 Hz range was computed at each voxel to quantify spontaneous low-frequency activity. To preserve the original signal scaling, the resulting ALFF values were rescaled by the standard deviation of the denoised BOLD data. The final ALFF maps were spatially smoothed with a 6 mm FWHM Gaussian kernel implemented in Nilearn. In addition, regional homogeneity (ReHo)64,65 was computed using AFNI’s 3dReHo66 to assess local synchronization of neighboring voxel time series.

Processed functional time-series were extracted from the residual BOLD signal with Nilearn’s NiftiLabelsMasker for each brain region of the examined atlas. For a brain region’s parcellated time-series to be retained, at least 10% of the voxels within that region must have had non-zero, non-NaN values (Fig. 2). Functional connectivity matrices were then derived for each atlas, where each edge was defined as the Pearson’s correlation between any two brain regions unsmoothed functional BOLD time-series.

Fig. 2
figure 2

Brain coverage. The overall cortical and subcortical brain coverage—the percentage of voxels within a brain region including non-zero, non-NaN values—is depicted for each BOLD run, averaged across subjects. Brain regions with warmer colors (e.g., cingulate cortex) indicate regions with high brain coverage across subjects, whereas brain regions with colder colors (e.g., temporal pole) indicate regions with lower overall coverage across subjects. Regional brain coverage for each subject can be found in the tsv file: sub- < subject_ID > _task- < BOLD_run > _space-MNI152NLin2009cAsym_seg- < atlas_of_choice > _stat-coverage_bold.tsv under: derivatives/xcp_d_*_output/sub- < subject_ID > /func/. The figure above was generated based on the Schaefer 456 atlas and xcp_d_with_GSR_bandpass_output denoising pipeline.

Full size image

Many internal operations of XCP-D use AFNI42,67, ANTS68, TemplateFlow (version 24.2.2)39, matplotlib (version 3.10.0)69, Nibabel (version 5.3.2)70, Nilearn (version 0.11.1)33, numpy (version 2.2.1)71, pybids (version 0.18.1)72, and scipy (version 1.15.1)73.

The denoised BOLD signal time-series for each atlas are saved by XCP-D as a TSV file (sub- < subject_ID > _[specifiers]-mean_timeseries.tsv) under each subject’s functional folder (location: derivatives/xcp_d_*_output/sub- < subject_ID > /func/). For a summary of each subject’s denoising outputs accompanied by extensive visual assessments, the interested reader may open the corresponding sub- < subject_ID > .html file that can be found under derivatives/xcp_d_*_output/.

For more details on XCP-D preprocessing pipelines, please refer to the associated manuscript47 as well as excellent online documentation: https://xcp-d.readthedocs.io. For a detailed description on XCP-D’s outputs and corresponding naming structure, please refer to: https://xcp-d.readthedocs.io/en/latest/outputs.html.

Data transmission, de-identification and format conversion

All DICOM files were de-identified following HIPAA regulations and converted to BIDS-compliant NIfTI files. Metadata were extracted and stored in JSON sidecar files. Behavioral and physiological data were organized in tab-separated value (.tsv) files. To ensure de-identification of subjects, the facial structure from each anatomical (T1w) MRI scan was removed using the PyDeface tool (version 2.0.2)74.

Data Records

All data have been made publicly available on OpenfMRI23: https://doi.org/10.18112/openneuro.ds006623.v1.0.0. To ensure a consistent naming convention with other publicly available datasets, we organized the files based on the BIDS format (version 1.8.0).

At the top-level, we are providing each subject’s unprocessed anatomical and functional scans, organized in their respective sub- < subject_ID > folders; all preprocessed and denoised data are organized under a derivatives folder. Participant demographic information and anesthetic dosage parameters are provided in derivatives/Participant_Info.xlsx. Individual propofol infusion profiles and hand-squeeze force traces are available under derivatives/Propofol_Infusion/ and derivatives/Squeeze_Force/, respectively, to facilitate individualized analyses of dose–response and behavioral responsiveness.

For each subject, we are specifically providing the outputs from the FreeSurfer-derived anatomical reconstruction pipeline (folder: derivatives/sourcedata/freesurfer), the MRIQC-derived quality control pipeline (folder: derivatives/mriqc_output), the fMRIPrep-derived preprocessing pipeline (folder: derivatives/fmriprep_output), and the XCP-D-derived denoising pipelines (folders: derivatives/xcp_d_*_output). A schematic tree structure of the top-level directory and its contents is provided, along with the contents of one of the denoising pipelines (Fig. 3).

Fig. 3
figure 3

A schematic tree structure of the top-level directory and its contents. Unprocessed anatomical and functional scans for each subject are stored in their respective sub- < subject_ID > folders. All preprocessed and denoised data are located in the derivatives folder.

Full size image

Specific output files that might be of interest to the reader include:

  • Dataset → derivatives → xcp_d_*_output → sub- < subject ID > → func → sub- < subject ID > _task- < functional scan ID > _space-MNI152NLin2009cAsym_seg- < atlas ID > _stat-mean_timeseries.tsv

    Description: Denoised regional BOLD signal time-series for subject < subject ID > . The dimensions of this file are number of fMRI volumes x number of brain regions. The former is determined by the functional scan analyzed (i.e., functional scan ID in the file name), while the latter is determined by the specific atlas being used (i.e., atlas ID in the file name).

  • Dataset → derivatives → xcp_d_*_output → sub- < subject ID > → func → sub- < subject ID > _task- < functional scan ID > _space-MNI152NLin2009cAsym_res-4_desc-denoised_bold.nii.gz and sub- < subject ID > _task- < functional scan ID > _space-MNI152NLin2009cAsym_res-4_desc-denoisedSmoothed_bold.nii.gz

    Description: Denoised BOLD volumes in standardized space (unsmoothed and smoothed versions).

  • Dataset → derivatives → xcp_d_*_output → sub- < subject ID > → func → sub- < subject ID > _task- < functional scan ID > _motion.tsv

    Description: Text file containing several motion-related parameters, including framewise displacement (in mm) for each fMRI volume, that can be used as potential confounder(s) in subsequent analyses.

  • Dataset → derivatives → fmriprep_output → sub- < subject ID > .html

    Description: Visual quality assessment report, provided as a web link, that allows the user to visually inspect the quality of preprocessing (i.e., fMRIPrep).

  • Dataset → derivatives → xcp_d_*_output → sub- < subject ID > .html

Description: Visual quality assessment report, provided as a web link, that allows the user to visually inspect the quality of denoising (i.e., XCP-D).

Output data files are provided in various formats (Table 2). Volumetric files, such as anatomical and functional images, are provided in NIfTI format, descriptor files are provided as text files in tab-delimited (.tsv) and Javascript object notation (.json) formats, and visual assessments are provided as websites in.html format:

Table 2 Output Data Formats.
Full size table

Technical Validation

Quality control measures

Each subject’s T1-weighted anatomical and fMRI scans underwent rigorous quality assessment using the MRI Quality Control protocol (MRIQC version 25.0.0rc0: https://mriqc.readthedocs.io/en/latest/). Following minimal MRI preprocessing, MRIQC calculates several quality control metrics: for T1-weighted anatomical images, metrics can include contrast-to-noise ratio (CNR), entropy focus criterion (EFC), foreground-background energy ratio (FBER), and signal-to-noise ratio (SNR) (Table 3); for BOLD fMRI data, computed metrics can include foreground-background energy ratio, ghost-to-signal ratio, FD, temporal SNR, SNR, and DVARS (Table 4). All quality control metrics are provided with the complete dataset (location: derivatives/mriqc_output/) to enable users to independently determine appropriate de-noising methods or data exclusions based on their specific analytical requirements.

Table 3 Quality control metrics for anatomical (T1-weighted) MRIs.
Full size table
Table 4 Quality control metrics for functional (BOLD) MRIs.
Full size table

To further assess the quality of our data against a well-established dataset, we computed the aforementioned quality control metrics on the Human Connectome Project dataset (n = 100 unrelated individuals; https://www.humanconnectome.org/study/hcp-young-adult/data-releases) and visualized the results against our dataset (Figs. 4 and 5).

Fig. 4
figure 4

Quality Control of Anatomical Images. Visual comparison of quality control metrics derived from T1-weighted anatomical scans in our dataset and those from the Human Connectome Project (HCP; n = 100 unrelated individuals). All analyses were conducted on de-identified T1-weighted images. Definitions of each metric are provided in Table 3.

Full size image
Fig. 5
figure 5

Quality Control of Functional Images. Visual comparison of key quality control metrics from BOLD functional scans across conditions. For reference, corresponding metrics from the first resting-state scan of 100 unrelated participants in the Human Connectome Project (HCP) are included. Three outlier subjects in the Imagery 3 condition, with framewise displacement values of 0.96 mm, 1.27 mm, and 2.16 mm, are excluded from the framewise displacement subpanel for visual clarity. Definitions of each metric are provided in Table 4.

Full size image

Whole brain general linear model analysis

Four event types corresponding to the three mental imagery tasks (tennis, navigation, and squeeze) and the motor response task (action) were defined for each of the five experimental conditions (Base1, PreLOR, LOR, ROR, and Base2). Each event type was modeled by convolving event onset times with a canonical hemodynamic response function using the BLOCK-model within the 3dDeconvolve function in AFNI. Event durations were set at 15 seconds for tennis and navigation imagery tasks, 10 seconds for squeeze imagery, and 2 seconds for the motor action task. Voxel-wise activation amplitudes (beta values) were estimated across the whole brain after censoring frames exceeding a framewise displacement threshold of 0.8 mm. Additionally, verbal instruction-induced activations were estimated in a separate general linear model. In this model, all verbal instruction events (“tennis imagery,” “navigation imagery,” “squeeze imagery,” “action,” and “relax”) were combined into a single event type with a duration of 2 seconds. Consequently, five distinct activation maps (tennis imagery, navigation imagery, squeeze imagery, motor action, and verbal instruction) were generated for each condition per participant. The above analysis utilized high-pass filtered data (>0.008 Hz), spatially smoothed, and specifically excluded global signal regression from the xcp_d_without_GSR_highpass_output pipeline.

While fMRIPrep employed slightly different preprocessing toolboxes and pipelines compared to our previously published work19, we confirmed that these differences had a negligible impact on the results. We further validated that the current dataset robustly captures distinct patterns of brain activation associated with various mental imagery tasks across different levels of sedation (Fig. 6). Notably, previous analyses of these data identified instances of volitional mental imagery despite behavioral unresponsiveness18, elucidated the anterior insula’s role in modulating conscious access19, and characterized asymmetric neural dynamics associated with transitions in consciousness20, thus supporting the neural inertia hypothesis75.

Fig. 6
figure 6

Task-Induced Brain Activity. Group-level z-maps are shown for five task conditions: tennis imagery (Tennis), navigation imagery (Navigation), squeeze imagery (Squeeze), actual hand squeeze (Action), and verbal instruction (Instruction) during five study phases: Imagery 1, PreLOR, LOR, ROR, and Imagery 4. All z-maps reflect one sample t-tests against zero, cluster-level corrected at α < 0.05. During LOR, task-evoked activity from mental imagery was absent. Instruction-evoked responses were attenuated and localized to the primary auditory cortex, with concurrent deactivation in the bilateral anterior insular cortex.

Full size image

Functional connectivity analysis and validation of Integration-Segregation Difference (ISD)

To further validate the dataset, we computed the integration–segregation difference (ISD), as defined in prior work76. ISD was calculated as the difference between network integration and segregation, quantified by multi-level global efficiency and clustering coefficients, respectively. For these analyses, we used xcp_d_without_GSR_bandpass_output to construct the condition-specific functional connectivity matrices (Fig. 7, top). The first and last five minutes of the LOR condition were omitted from functional connectivity calculation to mitigate transient effects of consciousness state transitions. For segregation calculation, the first mode was subtracted after matrix eigendecomposition to regress the global integrative component. A total of 50 logarithmically spaced values between 0.001 to 1.0 was used as multi-level thresholds. Consistent with our earlier study76, which employed alternative preprocessing pipelines and software to analyze fMRI data, we observed a significant decrease in ISD during the LOR condition compared with conscious states, reflecting reduced network integration relative to segregation during deep sedation (Fig. 7).

Fig. 7
figure 7

Functional Connectivity and Integration–Segregation Difference (ISD) Across Conditions. The cortex was parcellated into 400 regions using the Schaefer 400-parcel atlas and organized according to the Yeo seven-network parcellation, while the subcortex was defined by 56 regions from the Tian subcortical atlas. fMRI time series were extracted from all 456 regions of interest (ROIs). Pairwise Pearson correlation coefficients were Fisher z-transformed to generate functional connectivity matrices for each experimental condition. The ISD metric was calculated as the difference between integration and segregation, where integration was quantified by multi-level global efficiency and segregation by the multi-level clustering coefficient. Asterisks (*) indicate statistically significant differences after FDR correction (p < 0.05); hash marks (#) indicate uncorrected significance (p < 0.05); and n.s. denotes non-significant differences. Abbreviations: VIS = visual, SMN = somatomotor, DAN = dorsal attention, VAN = ventral attention, LIM = limbic, FPN = frontoparietal, DMN = default mode network, SUB = subcortical, LH = left hemisphere, RH = right hemisphere.

Full size image