Dynamic brain states during encoding and their post-encoding reinstatement predicts episodic memory in children

Reviewer #1 (Public review):

Summary:

Zeng et al. characterized the dynamic brain states that emerged during episodic encoding and the reactivation of these states during the offline rest period in children aged 8-13. In the study, participants encoded scene images during fMRI and later performed a memory recognition test. The authors adopted the BSDS approach and identified four states during encoding, including an "active-encoding" state. The occupancy rate of, and the state transition rates towards, this active-encoding state positively predicted memory accuracy across participants. The authors then decoded the brain states during pre- and post-encoding rests with the model trained on the encoding data to examine state reactivation. They found that the state temporal profile and transition structure shifted from encoding to post-encoding rest. They also showed that the mean lifetime and stability (measured with self-transition probability) of the "default-mode" state during post-encoding rest predict memory performance.

Strengths:

How brain dynamics during encoding and offline rest support long-term memory remains understudied, particularly in children. Thus, this study addresses an important question in the field. The authors implemented an advanced computational framework to identify latent brain states during encoding and carefully characterized their spatiotemporal features. The study also showed evidence for the behavioral relevance of these states, providing valuable insights into the link between state dynamics and successful encoding and consolidation.

Weaknesses:

(1) If applicable, please provide information on the decoding performance of states during pre- and post-encoding rests. The Methods noted that the authors applied a threshold of 0.1 z-scored likelihood, and based on Figure S2, it seems like most TRs were assigned a reinstated state during post-encoding rest. It would be useful to know, for the decodable TRs, how strong the evidence was in favor of one state over others. Further, was decoding performance better during post- vs. pre- encoding rest? This is critical for establishing that these states were indeed "reinstated" during rest. The authors showed individual-specific correlations between encoding and post-encoding state distribution, which is an important validation of the method, but this result alone is not sufficient to suggest that the states during encoding were the ones that occurred during rest. The authors found that the state dynamics vary substantially between encoding and rest, and it would be helpful to clarify whether these differences might be related to decoding performance. I am also curious whether, if the authors apply the BSDS approach to independently identify brain states during rest periods (instead of using the trained model from encoding), they find similar states during rest as those that emerged during encoding?

(2) During post-encoding rest, the intermediate activation state (S1) became the dominant state. Overall, the paper did not focus too much on this state. For example, when examining the relationship between state transitions and memory performance, the authors also did not include this state as a part of the analyses presented in the paper (lines 203-211). Could the author report more information about this state and/or discuss how this state might be relevant to memory formation and consolidation?

(3) Two outcome measures from the BSDS model were the occupancy rate and the mean lifetime. The authors found a significant association with behavior and occupancy rate in some analyses, and mean lifetime in others. The paper would benefit from a stronger theoretical framing explaining how and why these two different measures provide distinct information about the brain dynamics, which will help clarify the interpretation of results when association with behavior was specific to one measure.

(4) For performance on a memory recognition test, d' is a more common metric in the literature as it isolates the memory signal for the old items from response bias. According to Methods (line 451), the authors have computed a different metric as their primary behavioral measure (hits + correction rejections - misses - false alarms). Please provide a rationale for choosing this measure instead. Have the authors considered computing d' as well and examining brain-behavior relationships using d'?

(5) While this study examined brain state dynamics in children, there was no adult sample to compare with. Therefore, it is hard to conclude whether the findings are specific to children (or developing brains). It would be helpful to discuss this point in the paper.

Reviewer #2 (Public review):

This paper investigates the latent dynamic brain states that emerge during memory encoding and predict later memory performance in children (N = 24, ages: 8 -13 years). A novel computational approach (Bayesian Switching Dynamic Systems, BSDS) discovers latent brain states from fMRI data in an unsupervised and parameter-free manner that is agnostic to external stimuli, resulting in 4 states: an active-encoding state, a default-mode state, an inactive state, and an intermediate state. The key finding is that the percentage of time occupied in the active-encoding state (characterized by greater activity in hippocampal, visual, and frontoparietal regions), as well as greater transitions to this state, predicts memory accuracy. Memory accuracy was also predicted by the mean lifetime and transitions to the default-mode state (characterized by greater activity in medial prefrontal cortex and posterior cingulate cortex) during post-encoding rest. Together, the results provide insights into dynamic interactions between brain regions that may be optimal for encoding novel information and consolidating memories for long-term retention.

The approach is interesting and important for our understanding of neural mechanisms of memory during development, as we know less about dynamic interactions between memory systems in development.

Moreover, the novel methodology may be broadly useful beyond the questions addressed in this study. The manuscript is well-written and concise. Nonetheless, there are several areas for improvement:

(1) The study focuses on middle childhood, but there is a lack of engagement in the Introduction or Discussion about what is known about memory development and the brain during this period. Many of the brain regions examined in this study, particularly frontoparietal regions, undergo developmental changes that could influence their involvement in memory encoding and consolidation. The paper would be strengthened by more directly linking the findings to what is already known about episodic memory development and the brain.

(2) A more thorough overview of the BSDS algorithm is needed, since this is likely a novel method for most readers. Although many of the nitty-gritty details can be referenced in prior work, it was unclear from the main text if the BSDS algorithm discovered latent states based on activation patterns, functional connectivity, or both. Figure 1F is not very informative (and is missing labels).

(3) A further confusion about the BSDS algorithm was whether it necessarily had to work on the rest data. Figure 4A suggests that each TR was assigned one of the four states based on the maximum win from the log-likelihood estimation. Without more details about how this algorithm was applied to the rest data, it is difficult to evaluate the claim on page 14 about the spontaneous emergence of the states at rest.

(4) Although the BSDS algorithm was validated in prior simulations and task-based fMRI using sustained block designs in adults, it is unclear whether it is appropriate for the kind of event-related design used in the current study. Figure 1G shows very rapid state changes, which is quantified in the low mean lifetime of the states (between 1-3 TRs on average) in Figure 4C. On the one hand, it is a strength of the algorithm that it is not necessarily tied to external stimuli. On the other hand, it would be helpful to see simulations validating that rapid transitions between states in fMRI data are meaningful and not due to noise.

(5) The Methods section mentions that participants actively imagined themselves within the encoded scenes and were instructed to memorize the images for a later test during the post-encoding rest scan. This detail needs to be included in the main text and incorporated into the interpretation of the findings, as there are likely mechanistic differences between spontaneous memory replay/reinstatement vs. active rehearsal.

(6) Information about the general linear model used to discover the 16 ROIs that showed a subsequent memory effect are missing, such as: covariates in the model (motion, etc.), group analysis approach (parametric or nonparametric), whether and how multiple-comparisons correction was performed, if clusters were overlapping at all or distinct, if the total number of clusters was 16 or if this was only a subset of regions that showed the effect.

Reviewer #3 (Public review):

Summary:

This paper uses a novel method to look at how stable brain states and the transitions between them promote memory formation during encoding and post-encoding rest in children. I think the paper has some weaknesses (detailed below) that mean that the authors fall short of achieving their aims. Although the paper has an interesting methodological approach, the authors need better logic, and are potentially "double dipping" in their results - meaning their logic is circular. I think the method that they are using could be useful to the broader neuroimaging community, although they need to make this argument clearer in the paper.

Strengths:

The paper is interesting in that they use a novel method to look at brain state dynamics and how they might support memory.

Weaknesses:

The paper has several weaknesses:

(1) The authors use children as their study subjects but fail to reconcile why children are used, if the same phenomena are expected to be seen in adults (or only children), and if and how their findings change with age across an age range that ranges from middle childhood into early adolescence. They need to include more consideration for the development of their subject population. The authors should make it clear why and how memory was tested in children and not adults. Are adults and children expected to encode and consolidate in a similar manner to children? Do the findings here also apply to adults? Do the findings here also apply to adults? How was the age range of 8-13-year-old children selected? Why didn't the authors look at change with age? Does memory performance change with age? Do the BSDS dynamics change with age in the authors' sample?

(2) The authors look for brain state dynamics within a preselected set of ROIs that are selected because they display a subsequent memory effect. This is problematic because the state that is most associated with subsequent memory (S3, or State 3) is also the one that shows most activity in these regions (that have already been a priori selected due to displaying a subsequent memory effect). This logic is circular. It would be helpful if they could look at brain state dynamics in a more ROI agnostic whole brain approach so that we can learn something beyond what a subsequent memory analysis tells us. I think the authors are "double dipping" in that they selected regions for further analysis based on a subsequent memory association (remembered > forgotten contrast) and then found states within those regions showing a subsequent memory effect to further analyze for being associated with subsequent memory. Would it be possible instead to do a whole-brain analysis (something a bit more agnostic to findings) using the BSDS framework, and then, from a whole-brain perspective, look for particular brain states associated with subsequent memory? As it stands, it looks like S3 (state 3) has greater overall activation in all brain regions associated with subsequent memory, so it makes sense that this brain state is also most associated with subsequent memory. The BSDS analysis is therefore not adding anything new beyond what the authors find with the simple subsequent memory contrast that they show in Figure 1C. This particularly effects the following findings: (a) active-encoding state occupancy rate correlated positively with memory accuracy, (b) transitions to the active-encoding state were beneficial / Conversely, transitions toward the inactive state (S4) were detrimental, with incoming transitions showing negative correlations with memory accuracy / The active-encoding state serves as a "hub" configuration that facilitates memory formation, while pathways leading to this state enhance performance and transitions away from it impair encoding.

(3) The task used to test memory in children seems strange. Why should children remember arbitrary scenes? How this was chosen for encoding needs to be made clear. There needs to be more description of the memory task and why it was chosen. Why was scene encoding chosen? What does scene encoding have to do with the stated goal of (a) "Understanding how children's brains form lasting memories", (b) "optimizing education" and (c) "identifying learning disabilities"? What was the design of the recognition memory test? How many novel scenes were included in the test, and how were they chosen? How close were the "new" images to previously seen "old" images? Was this varied parametrically (i.e., was the similarity between new and old images assessed and quantified?)

(4) They ultimately found four brain states during encoding. It would be helpful if they could make the logic and foundation for arriving at this number clear.

(5) There is already extant work on whether brain states during post-encoding rest predict memory outcomes. This work needs to be cited and referred to. The present manuscript needs to be better situated within prior work. The authors should look at the work by Alexa Tompary and Lila Davachi. They have already addressed many of the questions that the authors seek to answer. The authors should read their papers (and the papers they cite and that cite them) and then situate their work within the prior literature.

More minor weaknesses:

(1) The authors should back up the claim that "successful episodic memory formation critically depends on the temporal coordination between these systems. Brain regions must coordinate their activity through dynamic functional interactions, rapidly reconfiguring their activity and connectivity patterns in response to changing cognitive demands and stimulus characteristics." Do they have any specific evidence supporting this claim?

(2) These claims seem overstated: "this work has broad implications for understanding memory function in children, for developing educational interventions that enhance memory formation, and enabling early identification of children at risk for learning disabilities." Can the authors add citations that would support these claims, or if not, remove them?