The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) data repository: Structural and functional MRI, MEG, and cognitive data from a cross-sectional adult lifespan sample

1The Cambridge Centre for Ageing and Neuroscience (Cam-CAN) Stage 2 cohort study is a large-scale (approx. N = 700), multi-modal (MRI, MEG, and behavioural), cross-sectional, population-based adult lifespan (18–87 years old) investigation of the neural underpinnings of successful cognitive ageing. The project is an interdisciplinary collaboration involving researchers with expertise in cognitive psychology, cognitive neuroscience, psychiatry, engineering, and public health. The full Cam-CAN study consists of three stages, described briefly below (for full protocol, see ). The focus of the present paper is the Stage 2 dataset, which includes raw and preprocessed MRI, MEG, and cognitive-behavioural data, along with demographic data and other cross-referenced measures collected from the Stage 2 cohort in Stage 1. Collection of Stage 3 data has recently completed, and these data will be added to the database in due course.

2A key focus of the Cam-CAN project is integrative analysis across domains of cognition and measures of neural structure, function, and connectivity, with the goal of understanding how neurocognitive systems adapt in order to overcome age-related changes. For example, combined T2*-weighted resting-state fMRI data with resting-state MEG data to show that ageing affects the vascular response independently of neural activity; showed how age-related differences in functional connectivity change with cognitive state; and combined fractional-anisotropy measures from the diffusion-weighted (DWI) MRI data with volumetric measures from the T1-weighted MRI data to show that white matter (WM) and grey matter (GM) in frontal cortex make independent contributions to age-related declines in fluid intelligence and multitasking.

3Several features of the Cam-CAN Stage 2 dataset together make it unique. First, the sample is derived from a larger, population-based sample (approximate N = 3000) recruited from the general population via Primary Care Trust lists, which can be related to national data, and which allows quantification of bias in the Stage 2 sample of people willing and able to undergo neuroimaging. Second, the distribution of ages was selected to be roughly uniform, allowing sufficient statistical power to test for differences within as well as across age groups. Third, Stage 2 of the study involved collecting a broad range of behavioural measures from 14 experiments spanning five main cognitive domains (attention, emotion, action, language, and memory), which can also be related to considerable amounts of demographic, health and lifestyle data obtained in Stage 1. Finally, Stage 2 includes a wide range of neuroimaging measures: high-resolution (1 mm3) T1- and T2-weighted images, diffusion-weighted images (DWI), magnetisation-transfer (MT) images, and BOLD EPI images during rest, a sensorimotor task and movie-watching, as well as MEG data during rest and the same sensorimotor task. The depth of this neurocognitive phenotyping is currently unparalleled, and provides a testing ground for the development of new multimodal analysis methods.

4In the following section, we briefly describe the three Cam-CAN data-collection stages. Acquisition and analysis of the MRI and MEG data in Stage 2 are described more fully in Section 2.

5In Stage 11, 2681 participants were interviewed in their homes to acquire demographic information; measures of cognitive, mental and physical health; and lifestyle information. Tests of vision, hearing, balance, and speeded response times were administered, and participants completed detailed self-report questionnaires about their physical activity and life experiences. Neuropsychological tests included general cognitive assessments (MMSE,; ACE-R,, ), tests of memory (logical memory from WMS-III UK, ), and verbal intelligence (Spot the Word, ).

6Measures taken in Stage 1 additionally served to screen participants for participation in Stage 2: To continue, participants were required to be willing to continue, be cognitively healthy (MMSE > 24), to meet hearing, vision, and English language ability criteria necessary for completing experimental tasks, and to be free of MRI or MEG contraindications and neurological or serious psychiatric conditions.

7In Stage 2, participants (target N = 700: 50 men, 50 women from each age decade) were recruited to attend testing sessions at the Medical Research Council (UK) Cognition and Brain Sciences Unit (MRC-CBSU) in Cambridge, UK. Owing to recruitment problems for the youngest decade (18–27), only 56 (27 men) were tested from this decade. In this stage, structural and functional MRI scans, MEG recordings, and cognitive task data were collected over three separate sessions. Structural and functional MRI scans collected in Stage 2 are listed in Table 1, Table 2; MEG recordings are listed in Table 3; and cognitive behavioural tasks are listed in Table 4. Physiological data (height, weight, and blood pressure) were also collected, and a saliva sample was taken for future genetic analysis. As the data from Stage 2 are the focus of the repository described in this paper, these data and derived measures are described in more detail in Section 2.

8In Stage 3, a subset of participants (target N = 280: 20 men, 20 women from each decade) were recruited to attend further MRI and MEG sessions within 3 years of their assessment in Stage 2. Over three sessions, structural MRI and physiological measures were collected, along with fMRI and MEG data on a variety of cognitive tasks. Structural MRI scans (all participants) included a repeat T1-weighted structural image, as well as T2-weighted FLAIR, and arterial spin labelling (ASL). Functional MRI tasks (target N = 140 each) investigated emotion regulation, emotional memory, fluid intelligence, picture naming, response selection and inhibition, sentence comprehension, and visual short-term memory; repeat resting-state data and field maps for distortion correction were also collected. Height, weight, and blood pressure, which were measured in Stage 2, were re-measured at the Stage 3 MRI session. MEG tasks investigated incidental memory, oddball processing, picture naming, response selection and inhibition, sentence comprehension, and word recognition, as well as a resting state. Hearing, vision, and cognitive status (MMSE), which were also measured in Stage 1, were re-evaluated in the MEG session.

9Using semi-automated Matlab and Linux shell scripts, raw data were pulled from various sources (testing laptops, MRI and MEG data servers) into a central location. Once there, further automated scripts identified new raw data and submitted them to the appropriate processing scripts. Behavioural data were analysed by custom Matlab scripts; MRI and MEG data were processed using Automatic Analysis (aa 4.2; ) pipelines and modules which called relevant functions from neuroimaging analysis software and toolboxes (SPM12, Wellcome Department of Imaging Neuroscience, London, UK; FSL,; Freesurfer, Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Boston, MA, USA; in-house code).

10The remit of the database is limited to the Cam-CAN initial study cohort project. Therefore, upon completion of data collection and preprocessing analyses, the repository will effectively be static (i.e., no further raw data will be added). However, as existing pipelines are improved, new data releases are made, and as new pipelines are developed, new processed data will become available. Further, any future studies that use the same cohort and adopt a similar data-sharing policy will likely be incorporated into the database (e.g., there are plans to re-test the Stage 2 cohort on a subset of the same cognitive and neuroimaging measures in the future, to provide longitudinal data).

11The Stage 2 repository contains MRI, MEG, and behavioural data from 656 participants aged 18–87 years old. All data are labelled with unique project IDs. Data were quality-control checked by semi-automated scripts monitored by the Cam-CAN methods team. All analysis scripts (including aa modules and recipes) are stored in the repository and can therefore be viewed by any user. Matlab scripts are also available to query the repository and compile data cross-referenced by participant identifiers.

12All MRI datasets were collected at a single site (MRC-CBSU) using a 3 T Siemens TIM Trio scanner with a 32-channel head coil. Participants were scanned in a single 1-hour session. Before scanning, physiological measurements were taken, and two behavioural experiments were run. In the scanner, memory foam cushions were used for comfort and to minimise head movement. Instructions and visual stimuli for functional tasks were back-projected onto a screen viewed through a mirror mounted on the head coil; auditory stimuli were presented via MR-compatible etymotics headphones; and manual responses were made with the right hand using a custom-built MR-compatible button-box. Cardiac data were recorded using photoplethysmograph/pulse-oximeter on the left index finger, sampled at 50 Hz.

13MRI data collected were structural (T1, T2, DWI, and MTI) and functional (resting-state, movie-watching, and sensorimotor task); details are provided in Table 1, Table 2. For the resting state scan, participants rested with their eyes closed for 8 min and 40 s. In the movie-watching task, participants watched and listened to an excerpt of a compelling but unfamiliar film: Alfred Hitchcock's “Bang! You're Dead”, a black-and-white television drama. The film was edited from its original running time of 30 min down to 8 min while maintaining the essential plot. Finally, in the sensorimotor task, also 8 min and 40 s long, participants detected the presentation of two circular checkerboards visually presented simultaneously to the left and right of a central fixation cross (34 ms duration) and a binaural tone (300, 600, or 1200 Hz; equal numbers of trials pseudorandomly ordered; 300 ms duration), presented either simultaneously (120 trials + 1 initial practice trial), or separately (8 trials; 4 visual only, 4 auditory only; included to discourage strategic responding to one modality). These task trials were combined with null trials of the same length, and all trials were pseudorandomly ordered using a 255-length m-sequence (m = 2 and minimal SOA of 2 s; ), resulting in effective stimulus onset asynchronies (SOAs) ranging from 2 to 26 s. Participants responded by pressing a button with their right index finger whenever they saw or heard any stimuli.

14All MRI data are available in standard NIfTI-1.1 format using single file (.nii) storage (3D for structural and 4D for functional); derived measures (e.g., ROI data) are available in ASCII text or Matlab data formats. All MRI analysis was conducted in SPM12, automated and parallelised by aa, except where noted.

15We describe the main processing pipelines developed and applied to date, but emphasize that other pipelines are possible and may be superior. MRI data were processed in separate streams (see Fig. 1): (i) a voxel-based morphometry (VBM) stream for structural analysis; (ii) a DWI stream with separate branches for diffusion tensor imaging (DTI) with nonlinear fitting and diffusion kurtosis imaging (DKI); (iii) an MTR stream; and (iv) an fMRI stream with separate branches for resting state, movie-watching, and sensorimotor tasks. Each stream was independent of the others, with two exceptions: Image types from each participant were coregistered to that participant's T1-weighted image via a rigid-body (6-df) linear transformation (blue dashed lines in Fig. 1), and normalisation parameters from the DARTEL procedure in the VBM stream were applied in the normalisation stages of the other streams (red dashed lines in Fig. 1), as described below. This procedure ensures a voxel-to-voxel correspondence for all metrics derived from each modality (e.g., GM volume, FA, MTR, BOLD timeseries, etc.). Each processing stream was applied to each individual participant's data independently, with the exception of the DARTEL group template stage in the VBM stream and later stages of fMRI processing (group ICA, inter-subject correlations, and 2nd-level analysis of task-related activity). Structural images have been de-faced in order to protect participant anonymity.

16In the VBM stream, the T1 image was initially coregistered to the MNI template, and the T2 image was then coregistered to the T1 image using a rigid-body (6-df) linear transformation. The coregistered T1 and T2 images were used in a multi-channel segmentation (SPM12 Segment, based on “New Segment” in SPM8; ) routine in order to extract probabilistic maps of 6 tissue classes: GM, WM, cerebrospinal fluid (CSF), bone, soft tissue, and residual noise. The native-space GM and WM images for all participants who passed quality-control checks (N = 651 in the current release003) were then submitted to diffeomorphic registration (DARTEL; ) to create group template images. The group template was then normalised to the MNI template via an affine transformation, and combined normalisation parameters (native to group template and group template to MNI template) were applied to each individual participant's GM and WM images. From this stage, the structural analysis stream followed two branches: For VBM analysis, individual normalised GM and WM images were smoothed (8 mm FWHM Gaussian kernel); for region of interest (ROI) analysis, applied masks from anatomically labelled template images (Harvard-Oxford atlas,,; in separate sub-branches) to extract mean regional GM and WM signal values for each participant.

17In the DWI stream, data were first skull-stripped using the BET utility in FSL. From there, two parallel branches were implemented: One for nonlinear estimation of the typical second-order diffusion tensor (and associated derived metrics, such as FA, mean diffusivity (MD), radial and axial diffusivity (RD, AD), etc.), and another for estimation of higher-order moments like kurtosis (and associated derived metrics, such as mean, radial, and axial kurtosis (MK, RK, AK); both DTI and DKI pipelines use in-house code).

18In the fMRI stream, the multiple echoes collected in the movie-watching task were first combined. Then, data from each functional run (resting state, movie-watching, and sensorimotor task) were unwarped (using field-map images) to compensate for magnetic field inhomogeneities, realigned to correct for motion, and slice-time corrected. After the EPI data were co-registered to the T1 image, the normalisation parameters from the VBM stream were then applied to warp functional images into MNI space. From this stage, three branches split off: For ROI analysis, mean regional time-courses were extracted using the template method described in VBM above; for task-related fMRI analyses, normalised images were smoothed (12 mm Gaussian kernel), and a general linear model (GLM) was applied with regressors defined by stimulus onsets (simultaneous and unimodal events separately) and 6 motion parameters. Further fMRI analysis pipelines are in development, for example, i) to assess functional connectivity using independent components analysis (ICA), ii) remove residual motion artifacts using wavelet de-spiking and regression of WM/CSF signals and higher-order expansions of the movement parameters, iii) calculate resting-state fluctuation amplitudes (RSFA) in order to scale task activations, and iv) investigate voxel-wise inter-subject correlations evoked by the movie-watching task.

19The final number of participants whose MRI data were processed to completion, i.e., have ROI data for T1, T2, DWI, MT and T2* (for rest, movie and sensorimotor task) at end of paths in Fig. 1 (excluding those with artifacts, missing or incomplete data, etc), is 614 (numbers per seven decades from 18 to 87, respectively: 51, 101, 93, 91, 96, 93, 89).

20All MEG datasets were collected at a single site (MRC-CBSU) using a 306-channel VectorView MEG system (Elekta Neuromag, Helsinki), consisting of 102 magnetometers and 204 orthogonal planar gradiometers, located in a light magnetically shielded room (MSR). Data were sampled at 1 kHz with a highpass filter of 0.03 Hz. Recordings were taken in the seated position. Head position within the MEG helmet was estimated continuously using four Head-Position Indicator (HPI) coils to allow for offline correction of head motion. Two pairs of bipolar electrodes were used to record vertical and horizontal electrooculogram (VEOG, HEOG) signals to monitor blinks and eye-movements, and one pair of bipolar electrodes records the electrocardiogram (ECG) signal to monitor pulse-related artefacts. Instructions and visual stimuli were projected onto a screen through an aperture in the front wall of the MSR; auditory stimuli were presented via etymotic tubes; responses were made via a custom-built button box with fibre optic leads.

21MEG data were collected during resting state, a sensorimotor task, and an audio-visual (passive) task (see Table 3). During the resting state recording, participants sat still with their eyes closed for at least 8 min and 40 s, to match the fMRI resting state scan. During the sensorimotor task recording, participants performed the same task as in the fMRI version. The audio-visual (passive) task used the same stimuli as the sensorimotor task, but with visual and auditory stimuli presented in isolation rather than simultaneously (in order to facilitate separation of the MEG responses evoked by each modality), and participants were not required to respond. In this task, 120 trials of unimodal stimuli (60 visual bilateral checkerboards presented simultaneously, 60 auditory tones at one of three equiprobable frequencies; see definition of stimuli in MRI section above) were presented at a rate of approximately 1 per second.

22Raw and maxfiltered MEG data are available in Neuromag's FIF format; subsequently preprocessed data are available in SPM12 format, with output from some intermediate stages (e.g., ICA) in Matlab format. All MEG analyses were implemented using aa.

23The MEG preprocessing pipeline is illustrated in Fig. 2. For each run, temporal signal space separation (tSSS,; MaxFilter 2.2, Elekta Neuromag Oy, Helsinki, Finland) was applied to continuous MEG data to remove noise from external sources and from HPI coils (correlation threshold 0.98, 10-sec sliding window), for continuous head-motion correction (in 200-ms time windows), and to virtually transform data to a common head position (‘-trans default’ option with origin adjusted to the optimal device origin, [0,+13,− 6]). MaxFilter was also used to remove mains-frequency noise (50-Hz notch filter) and to automatically detect and virtually reconstruct any noisy channels. Data were then imported to SPM12 format. ICA was used to identify physiological artefacts from blinks, eye-movements, and pulse, i.e. those ICs that had the highest correlation with the VEOG, HEOG and ECG channels respectively. These 3 ICs were then projected out of the data.

24For the sensorimotor and passive audio-visual tasks, evoked analysis of event-related fields (ERFs) and time–frequency analysis of induced signals were performed, time-locked to stimuli and manual responses. Further pipelines are being developed for the MEG data, including source estimation of task-evoked signals (using a head model defined using each participant's structural MRI), and connectivity analysis of sensor- and source-level data (e.g., using multivariate autoregressive (MVAR) modelling or correlation of amplitude envelopes).

25The final number of participants whose MEG data from all three sessions (rest, sensorimotor task, and passive audio-visual task) were successfully processed to the stage of artefact rejection in Fig. 2 is 623 (numbers per seven decades from 18 to 87, respectively: 47, 102, 100, 97, 94, 93, 90). The final number of participants with complete MRI and complete MEG data is 572 (numbers per seven decades: 44, 96, 90, 86, 89, 84, 83).

26Behavioural data were collected on 14 cognitive tasks selected to assess five core cognitive domains: executive function, emotional processing, motor and action function, language processing, and memory. Most experiments were computerised tasks run on a laptop computer using E-Prime v1 or v2, Visual Basic, the Cogent toolbox for Matlab; several were paper-and-pencil tasks, with data input via custom Matlab scripts; the remainder required specialist equipment or were naturalistic table-top tasks. A brief description of each task is given in Table 3; for full descriptions, see. Demographic information is also available from Stage 1. Behavioural and demographic data are available in ASCII (.txt) format.

27Height was measured with a portable stadiometer with a sliding head plate, a base plate and a connecting rod marked with a measuring scale. Weight was measured with a portable battery operated electronic weighing scales. Blood pressure measures were measured with a Digital Blood Pressure Monitor (A&D Medical UA-774); three measurements were taken in order to ensure reliability. These physiological measures are available in the behavioural data section of the repository. A saliva sample was also collected for future genetic analyses.

28Currently, the repository is currently stored on a Linux file system and can be accessed via secure shell (SSH) connection or secure file-transfer (SFTP) protocol. Instructions for accessing the data are available here: http://www.mrc-cbu.cam.ac.uk/datasets/camcan/.

29Future studies based on the same or similar samples will be incorporated into the database.

30The Cam-CAN Stage 2 data repository is a large, multi-modal, cross-sectional adult lifespan dataset designed to facilitate characterisation of age-related changes in cognition and brain structure and function, and to enable analysis aimed at discovering neurocognitive mechanisms that support healthy cognitive ageing. The database contains raw and preprocessed structural MRI, functional MRI (active tasks and resting state), and MEG data (active tasks and resting state), derived scores from cognitive behavioural experiments, and demographic and neuropsychological data. This repository of multi-modal neuroimaging and cognitive behavioural data thus provides an unprecedented depth of neurocognitive phenotyping, enabling integrative analyses of age-related changes in brain structure, brain function, and cognition.