Episodic memory in aspects of brain information transfer by resting-state network topology

  1. Tianyi Yan
  2. Gongshu Wang
  3. Li Wang
  4. Tiantian Liu
  5. Ting Li
  6. Luyao Wang
  7. Dingjie Suo
  8. Shintaro Funahashi
  9. Duanduan Chen
  10. Bin Wang
  11. Jinglong Wu

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    The paper describes a re-analysis of a public fMRI data set which includes measures of resting state connectivity and separate task-based scans of memory encoding and memory retrieval tasks. The paper proposes an analysis method termed "information transfer" that reveals functional interactions between various brain networks during encoding and retrieval as well as differences in these interactions during encoding vs. retrieval. While the methods are potentially interesting, the payoff-or new insight afforded by these methods compared to existing methods-is not fully established.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.”)

Reviewed by

Read the full article See related articles

Discuss this preprint

Start a discussion What are Sciety discussions?

Listed in

Log in to save this article

Abstract

Studies suggest that resting-state functional connectivity conveys cognitive information; also, activity flow mediates cognitive information transfer. However, the exact mechanism of interregional interactions underlying episodic memory remains unclear. We performed a combined analysis of task-evoked activity and resting-state functional connectivity by activity flow mapping to estimate the information transfer mechanism of episodic memory. We found that the cognitive control and attentional networks were the most recruited structures in information transfers during both encoding and retrieval processes; these networks were correlated with task-evoked activation. Differences in information transfer intensity between encoding and retrieval mainly existed in the visual, somatomotor and hippocampal systems. Furthermore, information transfer showed high predictive power for episodic memory ability and mediated relationships between task-evoked activation and memory performance. Additional analysis indicated that structural connectivity had a transportive role in information transfer. Finally, our study presented the information transfer mechanism of episodic memory from multiple neural perspectives.

Article activity feed

  1. eLife

    Evaluation Summary:

    The paper describes a re-analysis of a public fMRI data set which includes measures of resting state connectivity and separate task-based scans of memory encoding and memory retrieval tasks. The paper proposes an analysis method termed "information transfer" that reveals functional interactions between various brain networks during encoding and retrieval as well as differences in these interactions during encoding vs. retrieval. While the methods are potentially interesting, the payoff-or new insight afforded by these methods compared to existing methods-is not fully established.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. The reviewers remained anonymous to the authors.”)

  2. eLife

    Reviewer #1 (Public Review):

    In this study, Yan and colleagues performed a comprehensive analysis of information flow during an episodic memory task. Using the information transfer technique from Michael Cole's group, they showed that there were different patterns of information transfer during encoding and retrieval phases of the episodic memory task. They showed greater correlation between information transfer and task activation during the encoding phase, compared with the retrieval phase. Furthermore, information transfer intensities could be used to predict memory performance, above and beyond using simple functional connectivity or task activations. A mediation analysis showed that task activation indirectly affected memory performance via information transfer. Finally, information transfer was stronger for direct anatomical connections compared with no/weak anatomical connections (as measured by diffusion MRI).

    Overall, I think this is an ambitious study, packed with an impressive amount of analyses. However, there are serious issues that need to be addressed, including the fact that there is a serious lack of methodological details, so it is hard to fully judge the paper. Furthermore, given that this is a memory study, an obvious weakness is the lack of inclusion of hippocampus in the analyses.

    1. From what I can tell, the authors only utilized cortical regions in their analyses. Their HIPP system basically covered parahippocampal and entorhinal cortex, but does not include the hippocampus itself. This seems like a major weakness given this study is focused on episodic memory.

    2. Sentences like "The activity flow mapping procedure unified both biophysical and computational mechanisms into a single information-theoretic framework" are overly strong and causal. How does activity flow incorporate biophysical mechanisms?

    3. I have some conceptual concerns about information transfer. Basically, information transfer (ITE) is defined as the difference between the goodness of matched prediction (MatchAB) and goodness of mismatched prediction (MismatchAB). As such, ITE can be significant even if goodness of matched prediction is poor, as long as goodness of mismatched prediction is even worse.

    4. It is still unclear to me whether the information transfer mapping was applied to the task contrast beta values? There was also no explanation of what task regressors were used? For example, did the authors simply model the each encoding and each retrieval trial as a "block"? And each encoding and retrieval trial had its own beta coefficient? In the analysis, were the betas from control trials utilized? For the retrieval trials, did the authors distinguish between correct and incorrect trials?

    5. I don't understand how Figure 5 was obtained. My understanding is that information transfer intensity is a 360 x 360 matrix (Figure 2). How do we end up with a spatial map? What is being correlated for each region? The correlation is across subjects?

    6. The associations between information transfer intensity and memory performance (Figure 6) are impressive. The strength of prediction (Figure 7) is also quite strong. I would suggest the authors also report the coefficient of determination R^2 (see https://doi.org/10.1016/j.neuroimage.2019.02.057).

    7. The section "Prediction model establishment" is unclear. More specifically, it is unclear whether the prediction results were good because of "circular reasoning" with feature selection performed on the full set of subjects. This seems to be the case based on lines 816 to 819: "First, correlation analysis (Pearson correlation, Spearman correlation or robust regression) between each edge in the information transfer matrices and memory scores was performed across subjects. Second, a threshold was applied to the matrix that only retained edges that were significantly and positively correlated with behavior (p < 0.01)". The feature selection process should only be performed in the training set (n - 1 subjects).

    8. In this dataset, did the bipolar group have worse memory performance than the control group? If not, I don't see the value of Figure 10. If yes, could differences in information transfer intensity explain the performance gap?

    9. The UCLA dataset also contains participants with ADHD and schizophrenia/schizoaffective disorder. Did these participants also have worse memory performance? Why were these two other patient groups not considered?

    10. More details need to be provided about the analysis in Figure 10 - Supplementary Figure 2. For example, were training and prediction performed for each trial? The input is also not clear. The authors wrote "The input was the averaged resting-state BOLD signal in each brain region" So the authors averaged the bold signal within each ROI, but did they use the first time point of the trial? Do they average BOLD signal across all timepoints within each trial? What was the cost function they minimize? What was the step size? Neural networks have many hyperparameters? How were they tuned? Hyperparameters should be tuned on a separate validation set, which is separate from the training and test set. How did the authors split their data into training, validation and test sets? In Figure 10 - Supplementary Figure 2, the authors drew the ROI signals like a time series. So did the authors feed in a time series to the neural network or was it a single input vector?

  3. eLife

    Reviewer #2 (Public Review):

    In this paper, Yan and colleagues examined the relationship between resting-state functional connectivity and memory task-related activity. They used a recently developed method for measuring the flow of information across network nodes. This technique allows one to test the predictive relationship between two regions, weighted by their functional connectivity, and to link these predictions to differences between task conditions. The relation between resting-state functional connectivity and task-related activation differences is an important and timely issue. A few previous studies have investigated this question in the context of memory, although none to my knowledge using this technique. However, the manuscript does not clearly justify what new knowledge is gained through this technique, nor how these findings are specific to memory versus other kinds of cognitive tasks. Part of the issue is simply that the paper provides insufficient explanation of the technique- it takes for granted that the method reveals the 'cognitive information transfer between brain regions' without explaining what exactly this means or why this should be the case. Another issue is that the paper provides insufficient details about how the analyses are implemented. For instance, it is not clear which task conditions are being compared to support the information transfer analysis- but this is absolutely critical for evaluating the methods and interpreting the results. Finally, the results are not well integrated with the memory literature, limiting its impact on the field.

  4. eLife

    Reviewer #3 (Public Review):

    The stated goal of the paper and analyses is to "provide a more intuitive understanding of the brain communication mechanism underlying episodic memory." This question is quite vague and, ultimately, I do not think the paper succeeds in providing a more intuitive understanding of how brain regions communicate during episodic memory. The paper presents some interesting methods, but I do not think the results leave the reader with a clear(er) intuition about how and why the various measures do or do not differ from alternative, existing approaches for measuring correlations/interactions/connectivity between brain regions.

    The paper lacks an overarching question or theoretical framework. Instead, the paper has a stronger methods focus. The methods are sophisticated and potentially interesting, but the payoff of these methods is not established. What is the key new insight? What problem does this new method solve that other, existing methods cannot solve? These concerns are amplified by the fact that many of the results are actually entirely consistent with existing evidence (e.g., see the Discussion).

    The paper is quite dense, with an abundance of different analyses and methods. The result is that none of the individual results are considered in sufficient detail or with a sufficient number of control analyses or comparisons that would help establish the true utility of the current method compared to existing methods.

    The description of the main measure as "information transfer" is misleading. The term implies that the measure is reflecting some kind of information about the stimulus (or memory state) and that there is some directionality (in time) to the transfer. But neither of these is true. The "information" measure is simply related to the accuracy with which a distributed pattern of activity is predicted from one region/voxel to another. But, if two different regions (or voxels) have correlated activity, then it will be mathematically true that one region's activity 'predicts' the other region's activity. I am not saying that this is unimportant or uninteresting, but it is different than showing that some information (e.g., about the stimulus or task state) that is contained within one region is (re) expressed in a different region at a subsequent time point (which would be more consistent with information transfer).

    Ultimately, the regions that were implicated in encoding and retrieval using the information transfer method are precisely the regions that would be expected based on univariate studies from the past 2 or 3 decades. I realize that the approach used here for identifying these regions is much more sophisticated, but is the method ultimately just picking up on univariate effects (in a more sensitive way)? For example, is it the case that a high percentage of successful "information transfer" at retrieval is BECAUSE that region tends to show univariate activation (or deactivation) during retrieval? Indeed, the paper specifically reports positive relationships between task evoked responses and information transfer. The question, then, is whether the measures are ultimately redundant-or to what degree are they redundant? It therefore seems essential to show that there is some fundamental, new insight afforded by the information transfer method compared to task-based univariate measures. Or, for some of the specific relationships between regions that differed during encoding vs. retrieval, the question would be whether the current method has a clear, demonstrable advantage relative to other connectivity approaches like beta-series correlations? The effects in Figure 4, for example, do appear to demonstrate some very robust differences between encoding and retrieval, but without a direct comparison against alternative methods, the potential impact of this approach is not clear.

    Even for places where the information transfer is compared against other methods (e.g., task activation) it is not clear whether this is a fair comparison-are the number of features/predictors matched across these measures?

    The data in Figure 10 are perhaps the most interesting because they highlight a dramatic difference between task-evoked measures and the information transfer measures. But, again, a deeper consideration of these data-and potential explanations-would be helpful.

    The "brain state" analysis (line 422) is not well motivated and it was hard to understand exactly how this analysis was performed (or why). Likewise, the artificial neural network is considered in such superficial detail that it is hard to draw much of an inference from the simulation.

    While there are some clear relationships between information transfer measures and memory performance (Fig 6), it would again be useful to compare these relationships using measures other than information transfer. For example, what about resting state FC? This was done for the analyses in Figure 7 - Supp 2, but not for the analyses in Fig 6. In fact, resting state FC was almost as good a predictor as the information transfer measure. The comparison against resting state FC is a particularly relevant comparison because several of the networks show fairly similar correlations regardless of whether the task was encoding or retrieval (see Fig 6). If correlations persist when using resting state measures, this would undermine any argument about these correlations reflecting the success of "information transfer" that specifically occurred during the memory task.

    The Discussion is not well focused and is overly speculative. There is a lot of speculation about how the different network interact and what, exactly, these interactions reflect. But the current study really does not inform these ideas one way or another. For example, there is discussion of attention, recollection vs. familiarity, etc., but none of these things are tested in any way in the present study.

  5. Version published to 10.1101/2021.02.28.433300 on bioRxiv