How much data is enough data?

We publish today a new study that establishes the minimal recording duration required to capture the typical fluctuations of individual brain activity in the resting state. The new study is published in open access to everyone in the journal Neuroimage.

The study was led by Dr. Alex Wiesman, NIH F32 fellow in the lab, assisted by Jason da Silva Castanheira, graduate trainee from McGill’s Integrated Program in Neuroscience.

Jason had published earlier this year a study that demonstrated that magnetoencephalography (MEG) recordings as short as 30 seconds were sufficient to differentiate between individuals within a large cohorts of participants.

In this new piece, we examined what is the minimum data duration required to capture the fluctuations of slow to fast brain rhythms and background (aperiodic, arhythmic) brain activity, despite their intrinsic variability, in the task free, resting state of the brain.

Stability analysis pipeline. (A) Source-imaged MEG data was epoched into consecutive, non-overlapping segments and transformed into the frequency domain per vertex. Power spectral density values were then averaged within canonical frequency bands and over vertices within Desikan-Killiany atlas regions. For each of 1000 permutations, the temporal order of these epochs was randomized. Data for each cortical location, frequency, and participant were then averaged over time bins of progressively larger numbers of epochs and used to compute ICC with time bins averaged over the same number of different epochs. (B) Graphical workflows. All analysis steps are indicated for the derivation of (1) stability estimates per each combination of region and frequency, (2) stability estimates per each combination of frequency and sample size, and (3) change scores representing the omission of each participant from the stability analysis. The extra inset to the far right indicates the participant subsamples on which each workflow was implemented. AUC: area under the curve. Cam-CAN: Cambridge Centre for Ageing and Neuroscience participant subsample. ICC: intraclass correlation coefficient. MAD: median absolute deviation. OMEGA: Open MEG Archives participant subsample.

To this end, we established the minimal length of MEG data required to yield a robust one-session snapshot of the frequency-spectrum derivatives that are typically used to characterize the complex dynamics of the brain's resting-state.

We studied the stability of common spectral measures of resting-state MEG source time series obtained from large samples of single-session recordings from shared data repositories featuring different recording conditions and instrument technologies (OMEGA: N = 107; Cam-CAN: N = 50).

We discovered that the rhythmic and arrhythmic spectral properties of intrinsic brain activity can be robustly estimated in most cortical regions when derived from relatively short segments of 30-s to 120-s of resting-state data, regardless of instrument technology and resting-state paradigm (e.g., eyes closed vs. eyes open).

Intra-session temporal stability of band-limited power. Intraclass correlation coefficients (y-axes) were obtained from 1000 permutations of epoch order and represent a measure of the stability of spectral power for each frequency band of interest as a function of data duration (x-axes, in seconds; top: OMEGA, N = 107; bottom: Cam-CAN, N = 50). Coloured lines represent the median across regions, dotted lines indicate ± one median absolute deviation across regions, and coloured shaded intervals represent the range of stability values across all modelled cortical regions of the Desikan-Killiany atlas. Horizontal shaded intervals in each plot represent typical thresholds used to define moderate (ICC > 0.50), good (ICC > 0.75), and excellent (ICC > 0.90) reliability. Note that the maximum data duration from OMEGA was shorter than from Cam-CAN.

Brain maps of intra-session stability of parameterized periodic features. Parcellated surface maps per each spectral frequency (denoted by Greek letters to the top-left of each set) indicate the length of data (colour bar; in seconds) required to achieve accepted thresholds for moderate (ICC > 0.50; right), good (ICC > 0.75; middle), and excellent (ICC > 0.90; left) reliability in each participant sample (top: OMEGA, N = 107; bottom: Cam-CAN, N = 50) for each modelled cortical region of the Desikan-Killiany atlas. Warmer colors indicate worse temporal stability, and regions left grey did not achieve the indicated level of reliability within the maximum length of data available for the respective participant sample (OMEGA: 120 s; Cam-CAN: 210 s).

Brain maps of intra-session stability of parameterized aperiodic features. Parcellated surface maps per each aperiodic feature (left: slope; right: offset) indicate the length of data (colour bar; in seconds) required to achieve accepted thresholds for moderate (ICC > 0.50; bottom rows), good (ICC > 0.75; middle rows), and excellent (ICC > 0.90; top rows) reliability in each participant sample (top: OMEGA, N = 107; bottom: Cam-CAN, N = 50) for each modelled cortical region of the Desikan-Killiany atlas. Warmer colors indicate worse temporal stability, and regions left grey did not achieve the indicated level of reliability within the maximum length of data available for the respective participant sample (OMEGA: 120 s; Cam-CAN: 210 s).

Using an adapted leave-one-out approach and Bayesian analysis, we also provide evidence that the stability of spectral features over time is unaffected by age, sex, handedness, and general cognitive functions.

Bayesian model evidence for null effects of participant sample characteristics on the stability of spectral power estimates. Dots indicate Bayesian model evidence (y-axis; in BF01) and associated model errors (denoted by relative dot size) for null effects of common participant sample characteristics (x-axis) on temporal stability for each spectral frequency (denoted by dot colour). Horizontal shaded intervals represent accepted thresholds for weak evidence for H1 (BF01 = 0.3 – 1), weak evidence for H0 (BF01 = 1 – 3), and moderate evidence for H0 (BF01 ≥ 3). Note that these models were only obtained for the Cam-CAN sample, due to its even distribution of age and availability of ACE-R cognitive scores.

In summary, short recording sessions of two minutes or even 30 seconds (depending on targeted brain rhythms or brain regions) are sufficient to yield robust estimates of frequency-defined brain activity during resting-state. We hope our study helps guide future empirical designs in the field, particularly when recording times need to be minimized, e.g., to maximize the comfort of patient or special populations, while ensuring that spectral brain signal derivatives are robust and reliable.


Data used in the preparation of this work were obtained from the Cam-CAN repository (available at http://www.mrc-cbu.cam.ac.uk/datasets/camcan/; Shafto et al., 2014; Taylor et al., 2017) and the OMEGA repository (available at https://www.mcgill.ca/bic/resources/omega; Niso et al., 2016). Code for MEG preprocessing and the stability analysis is available at https://github.com/aiwiesman/rsMEG_StabilityAnalysis. Source mapping powered by Brainstorm.

Previous
Previous

Boosting (auditory) memory with (visual) stimulations

Next
Next

We’re “Over the rainbow” with this new paper