Functional brain networks are dominated by stable group and individual factors, not cognitive or daily variation

Neuron 98(2):439–452.e5, doi 10.1016/j.neuron.2018.03.035; NIH author manuscript, PMC5912345 (NIHMS954374). Read for research direction R1, 4 October 2026. Provenance: papers/amadeus/gratton2018_stable_networks.provenance.json.

What was read

• Read in full: the PMC author manuscript (NCBI efetch XML converted to text): summary, introduction, results, discussion, STAR Methods, key resources table, all figure captions, highlights and references.
• Not read: the supplement (Supplementary Figures 1–7: matched-data and ten-session control analyses, per-subject and per-task plots, F statistics), referenced in the text but not in the XML. The figures were not inspected as images. Gordon et al. 2017 (Neuron), the paper that describes the dataset, was not read.

Design

• Data: the public Midnight Scan Club (MSC) dataset (OpenNeuro ds000224). Ten adults (5 women, aged 24–34), each scanned in 12 sessions on separate days, every session starting at midnight. All 12 were completed within 7 weeks. One participant was excluded for motion and self-reported sleep, leaving 9. IRB approved, written consent.
• Per session (10 functional sessions): 30 minutes of eyes-open rest, then four tasks: motor (7.8 min), semantic and visual-coherence (14.2 min together), and incidental-encoding memory (three runs, faces, scenes and words). 3T Trio, TR 2.2 s, 4 mm voxels. An in-scanner eye camera monitored wakefulness.
• Data kept per person (frames with framewise displacement > 0.2 mm censored): rest 219 minutes on average (73%), motor 35, semantic 41, coherence 40, memory 103.
• Analysis:
  • Networks: connectivity among 333 group-defined cortical parcels, built separately per person, task and split-half session group (sessions 1–5 against 6–10).
  • Similarity between networks: multidimensional scaling and PCA, then the added similarity for "same person", "same task", "same person and task" and "same person and session".
  • Single connections: a per-edge mixed-effects ANOVA giving the share of variance (ω²) for person, task, session and their interactions.
  • Two contrasts: subtracting the resting network from each task network, and comparing with task activation maps.

Main results (verified)

• Identity dominates.

  • In multidimensional scaling, dimensions 1–6, where points cluster by person, carry 48.8% of the variance across networks. Dimensions 7–12, where clustering by task appears, carry 19.0%. Clustering by session was not visible.
  • All networks share a common structure (mean similarity z(r) = 0.56). Same person adds 0.52, same person and task adds 0.26 more, same task across people adds 0.04, and same person and session adds 0.05.
  • In relative terms, group and individual effects are each about 35–40% of the total; cross-person task and person-by-session effects about 5% each.

• Single connections are explained mostly by who it is. The model of person, task, session and interactions explains R² = 0.89 of the variance per edge on average:

  • person ω² 0.50;
  • task 0.05;
  • task × person 0.09;
  • session 0.003;
  • session × person 0.02 (0.05 when the ten sessions are analysed separately).

  Adding a motion metric did not change the variance explained.

• Where: individual effects are strongest in control systems (frontoparietal, dorsal attention); group and cross-person task effects are strongest in sensorimotor systems. Session effects fall in low-signal, insular, visual and motor regions, "suggestive of variations in fMRI signal-to-noise and participant drowsiness".

• Task effects are mostly individual: about 5% common across people against about 20% specific to the person, even for simple tasks done equally well by everyone.

• Subtracting rest removes most of the stable part. Group similarity fell 86% and individual similarity 73% (Results; the Discussion gives 85% and 69%). Person-specific task and session effects survive and gain in relative size.

• Activations are the state signal. Task activation maps are less stable across people and much more task-dependent than connectivity. The authors conclude that "functional network measures are better suited for individual-level identification than tracking ongoing cognitive processes, which are better quantified via evoked measurements."

• Stated scope:

  • The sessions span only a few weeks, so the paper says nothing about longer timescales; it cites Poldrack 2015 for stability over a year.
  • It excludes effects of extended training.
  • It covers slow BOLD correlations only: "signals at higher frequencies (e.g., as might be measured with EEG or MEG) … may be more amenable to tracking state-dependent variability."
  • Within a scan, the group's earlier work found "little evidence for time-varying changes in functional networks at rest, once artifacts and state changes (drowsiness, task) are accounted for."

Limits

• Nine people, all young adults, scanned at midnight within seven weeks. Midnight scanning controls circadian variation but invites drowsiness, which the authors flag as a source of session variance.
• Group parcels (333) rather than individual parcels for most analyses. Individual parcels shifted similarity only slightly towards the group (Supp. Fig. 7, not read).
• "Task state" here means five simple laboratory tasks plus rest. Naturalistic thought, mood or memory recall were not manipulated, so the paper bounds what task changes in connectivity, not what thought content changes.
• Session effects were estimated as variance consistent across all runs of a session, which by construction is smaller than run-to-run variability. The authors note this.
• A small internal inconsistency in the rest-subtraction percentages (86/73 against 85/69).

What it means for Amadeus

• A month of resting-state fMRI connectivity mostly re-measures a fingerprint (verified, with the paper's scope). Connectivity is about half person identity, about a tenth task and person-by-task, and almost nothing from day to day. More sessions sharpen the person's stable map; they do not track what the person is thinking. For "thinking, daydreaming, imagining" (brief 6.7) the readout has to be activity patterns tied to labelled moments: evoked responses and decoders of the Tang and Horikawa kind. Connectivity does not do it.
• Dense sampling is still the right first step (inference). It buys the person-specific functional map that person-specific decoders and encoding models are built on. In MSC that took about 3.6 hours of usable rest (mean 219 min) plus about 3.6 hours of usable task data over 10 midnight sessions.
• Feasibility template (verified): 12 sessions on separate days within 7 weeks, around 90 minutes each, in 10 adults under an IRB. The paper does not say how they were recruited. That is a realistic scanner schedule for "a month of one person". By our arithmetic from the stated run lengths and retained frames, that is about 10.8 h of functional data acquired and about 7.3 h kept after motion censoring: near Tang's 7.5-hour plateau and short of Horikawa's 17 hours (existing summaries 04 and 05).
• Choice of modality (inference from the authors' own caveat): if the goal is moment-to-moment state over the month, the slow-BOLD-connectivity null does not carry over to faster signals. EEG, MEG or intracranial recordings are the candidates for state, and they are what a continuous stream would use.