Intrinsic fluctuations in global connectivity reflect transitions between states of high and low prediction error

Reviewer #1 (Public review):

Summary:

This study investigates whether prediction error extends beyond lower-order sensory-motor processes to include higher-order cognitive functions. Evidence is drawn from both task-based and resting-state fMRI, with the addition of resting-state EEG-fMRI to examine power spectral correlates. The results partially support the existence of dissociable connectivity patterns: stronger ventral-dorsal connectivity is associated with high prediction error, while posterior-anterior connectivity is linked to low prediction error. Furthermore, spontaneous switching between these connectivity patterns was observed at rest and correlated with subtle intersubject behavioral variability.

Strengths:

Studying prediction error from the lens of network connectivity provides new insights into predictive coding frameworks. The combination of various independent datasets to tackle the question adds strength, including two well-powered fMRI task datasets, resting-state fMRI interpreted in relation to behavioral measures, as well as EEG-fMRI.

Weaknesses:

Major:

(1) Lack of multiple comparisons correction for edge-wise contrast:

The analysis of connectivity differences across three levels of prediction error was conducted separately for approximately 22,000 edges (derived from 210 regions), yet no correction for multiple comparisons appears to have been applied. Then, modularity was applied to the top 5% of these edges. I do not believe that this approach is viable without correction. It does not help that a completely separate approach using SVMs was FDR-corrected for 210 regions.

(2) Lack of spatial information in EEG:

The EEG data were not source-localized, and no connectivity analysis was performed. Instead, power fluctuations were averaged across a predefined set of electrodes based on a single prior study (reference 27), as well as across a broader set of electrodes. While the study correlates these EEG power fluctuations with fMRI network connectivity over time, such temporal correlations do not establish that the EEG oscillations originate from the corresponding network regions. For instance, the observed fronto-central theta power increases could plausibly originate from the dorsal anterior cingulate cortex (dACC), as consistently reported in the literature, rather than from a distributed network. The spatially agnostic nature of the EEG-fMRI correlation approach used here does not support interpretations tied to specific dorsal-ventral or anterior-posterior networks. Nonetheless, such interpretations are made throughout the manuscript, which overextends the conclusions that can be drawn from the data.

Reviewer #2 (Public review):

Summary:

This paper investigates putative networks associated with prediction errors in task-based and resting-state fMRI. It attempts to test the idea that prediction errors minimisation includes abstract cognitive functions, referred to as the global prediction error hypothesis, by establishing a parallel between networks found in task-based fMRI where prediction errors are elicited in a controlled manner and those networks that emerge during "resting state".

Strengths:

Clearly, a lot of work and data went into this paper, including 2 task-based fMRI experiments and the resting state data for the same participants, as well as a third EEG-fMRI dataset. Overall, well written with a couple of exceptions on clarity, as per below, and the methodology appears overall sound, with a couple of exceptions listed below that require further justification. It does a good job of acknowledging its own weakness.

Weaknesses:

(1) The paper does a good job of acknowledging its greatest weakness, the fact that it relies heavily on reverse inference, but cannot quite resolve it. As the authors put it, "finding the same networks during a prediction error task and during rest does not mean that the networks' engagement during rest reflects prediction error processing". Again, the authors acknowledge the speculative nature of their claims in the discussion, but given that this is the key claim and essence of the paper, it is hard to see how the evidence is compelling to support that claim.

(2) Given how uncontrolled cognition is during "resting-state" experiments, the parallel made with prediction errors elicited during a task designed for that effect is a little difficult to make. How often are people really surprised when their brains are "at rest", likely replaying a previously experienced event or planning future actions under their control? It seems to be more likely a very low prediction error scenario, if at all surprising.

(3) The quantitative comparison between networks under task and rest was done on a small subset of the ROIs rather than on the full network - why? Noting how small the correlation between task and rest is (r=0.021) and that's only for part of the networks, the evidence is a little tenuous. Running the analysis for the full networks could strengthen the argument.

(4) Looking at the results in Figure 2C, the four-quadrant description of the networks labelled for low and high PE appears a little simplistic. The authors state that this four-quadrant description omits some ROIs as motivated by prior knowledge. This would benefit from a more comprehensive justification. Which ROIs are excluded, and what is the evidence for exclusion?

(5) The EEG-fMRI analysis claiming 3-6Hz fluctuations for PE is hard to reconcile with the fact that fMRI captures activity that is a lot slower, while some PEs are as fast as 150 ms. The discussion acknowledges this but doesn't seem to resolve it - would benefit from a more comprehensive argument.

Reviewer #3 (Public review):

Bogdan et al. present an intriguing and timely investigation into the intrinsic dynamics of prediction error (PE)-related brain states. The manuscript is grounded in an intuitive and compelling theoretical idea: that the brain alternates between high and low PE states even at rest, potentially reflecting an intrinsic drive toward predictive minimization. The authors employ a creative analytic framework combining different prediction tasks and imaging modalities. They shared open code, which will be valuable for future work.

However, the current manuscript would benefit from further clarification and empirical grounding, especially with regard to its theoretical framing (that PE-like state fluctuations are intrinsic and help us minimize PE), interpretation of results, and broader functional significance. Below, I outline a few major comments and suggestions that I think would strengthen the contribution.

(1) Consistency in Theoretical Framing

The title, abstract, and introduction suggest inconsistent theoretical goals of the study.

The title suggests that the goal is to test whether there are intrinsic fluctuations in high and low PE states at rest. The abstract and introduction suggest that the goal is to test whether the brain intrinsically minimizes PE and whether this minimization recruits global brain networks. My comments here are that a) these are fundamentally different claims, and b) both are challenging to falsify. For one, task-like recurrence of PE states during resting might reflect the wiring and geometry of the functional organization of the brain emerging from neurobiological constraints or developmental processes (e.g., experience), but showing that mirroring exists because of the need to minimize PE requires establishing a robust relationship with behavior or showing a causal effect (e.g., that interrupting intrinsic PE state fluctuations affects prediction).

The global PE hypothesis-"PE minimization is a principle that broadly coordinates brain functions of all sorts, including abstract cognitive functions"-is more suitable for discussion rather than the main claim in the abstract, introduction, and all throughout the paper.

Given the above, I recommend that the authors clarify and align their core theoretical goals across the title, abstract, introduction, and results. If the focus is on identifying fluctuations that resemble task-defined PE states at rest, the language should reflect that more narrowly, and save broader claims about global PE minimization for the discussion. This hypothesis also needs to be contextualized within prior work. I'd like to see if there is similar evidence in the literature using animal models.

(2) Interpretation of PE-Related Fluctuations at Rest and Its Functional Relevance

It would strengthen the paper to clarify what is meant by "intrinsic" state fluctuations. Intrinsic might mean task-independent, trait-like, or spontaneously generated. Which do the authors mean here? Is the key prediction that these fluctuations will persist in the absence of a prediction task?

Regardless of the intrinsic argument, I find it challenging to interpret the results as evidence of PE fluctuations at rest. What the authors show directly is that the degree to which a subset of regions within a PE network discriminates high vs. low PE during task correlates with the magnitude of separation between high and low PE states during rest. While this is an interesting relationship, it does not establish that the resting-state brain spontaneously alternates between high and low PE states, nor that it does so in a functionally meaningful way that is related to behavior. How can we rule out brain dynamics of other processes, such as arousal, that also rise and fall with PE? I understand the authors' intention to address the reverse inference concern by testing whether "a participant's unique connectivity response to PE in the reward-processing task should match their specific patterns of resting-state fluctuation". However, I'm not fully convinced that this analysis establishes the functional role of the identified modules to PE because of the following:

Theoretically, relating the activities of the identified modules directly to behavior would demonstrate a stronger functional role.

a) Across participants: Do individuals who exhibit stronger or more distinct PE-related fluctuations at rest also perform better on tasks that require prediction or inference? This could be assessed using the HCP prediction task, though if individual variability is limited (e.g., due to ceiling effects), I would suggest exploring a dataset with a prediction task that has greater behavioral variance.

Or even more broadly, does this variability in resting state PE state fluctuations predict general cognitive abilities like WM and attention (which the HCP dataset also provides)? I appreciate the inclusion of the win-loss control, and I can see the intention to address specificity. This would test whether PE state fluctuations reflect something about general cognition, but also above and beyond these attentional or WM processes that we know are fluctuating.

b) Within participants: Do momentary increases in PE-network expression during tasks relate to better or faster prediction? In other words, is there evidence that stronger expression of PE-related states is associated with better behavioral outcomes?

(3) Apriori Hypothesis for EEG Frequency Analysis

It's unclear how to interpret the finding that fMRI fluctuations in the defined modules correlate with frontal Delta/Theta power, specifically in the 3-6 Hz range. However, in the EEG literature, this frequency band is most commonly associated with low arousal, drowsiness, and mind wandering in resting, awake adults, not uniquely with prediction error processing. An a priori hypothesis is lacking here: what specific frequency band would we expect to track spontaneous PE signals at rest, and why? Without this, it is difficult to separate a PE-based interpretation from more general arousal or vigilance fluctuations.

(4) Significance Assessment

The significance of the correlation above and all other correlation analyses should be assessed through a permutation test rather than a single parametric t-test against zero. There are a few reasons: a) EEG and fMRI time series are autocorrelated, violating the independence assumption of parametric tests;
b) Standard t-tests can underestimate the true null distribution's variance, because EEG-fMRI correlations often involve shared slow drifts or noise sources, which can yield spurious correlations and inflating false positives unless tested against an appropriate null.

Building a null distribution that preserves the slow drifts, for example, would help us understand how likely it is for the two time series to be correlated when the slow drifts are still present, and how much better the current correlation is, compared to this more conservative null. You can perform this by phase randomizing one of the two time courses N times (e.g., N=1000), which maintains the autocorrelation structure while breaking any true co-occurrence in patterns between the two time series, and compute a non-parametric p-value. I suggest using this approach in all correlation analyses between two time series.

(5) Analysis choices

If I'm understanding correctly, the algorithm used to identify modules does so by assigning nodes to communities, but it does not itself restrict what edges can be formed from these modules. This makes me wonder whether the decision to focus only on connections between adjacent modules, rather than considering the full connectivity, was an analytic choice by the authors. If so, could you clarify the rationale? In particular, what justifies assuming that the gradient of PE states should be captured by edges formed only between nearby modules (as shown in Figure 2E and Figure 4), rather than by the full connectivity matrix? If this restriction is instead a by-product of the algorithm, please explain why this outcome is appropriate for detecting a global signature of PE states in both task and rest.

When assessing the correspondence across task-fMRI and rs-fMRI in section 2.2.2, why was the pattern during task calculated from selecting a pair of bilateral ROIs (resulting in a group of eight ROIs), and the resting state pattern calculated from posterior-anterior/ventral-dorsal fluctuation modules? Doesn't it make more sense to align the two measures? For example, calculating task effects on these same modules during task and rest?