dFCExpert: Learning Dynamic Functional Connectivity Patterns with Modularity and State Experts

1. While dFCExpert combines GNN, MoE, and soft clustering techniques that are well-established in prior research, could the authors clarify if there are any novel modifications or adaptations made to the GNN, MoE, or clustering components specifically tailored to brain connectivity data? How do these adaptations contribute to the overall innovation of the model?

2. The concatenation strategy used to combine different expert modules may increase the complexity of hyperparameter tuning and pose a risk of optimization instability. How do the authors plan to evaluate the hyperparameter sensitivity, optimization stability, and cross-dataset/task generalizability of their approach to address these concerns?

3. Although the authors describe the method clearly, could they provide a justification for certain key decisions, such as the rationale for selecting GIN from multiple GNN variants and the choice of the soft prototype clustering method? Would presenting comparative experiments strengthen the paper's claims?

4. The experimental validation seems limited despite using two authoritative large-scale datasets. Could the authors elaborate on their plans to conduct extensive experiments across different brain disorder datasets or diverse neuroimaging datasets to comprehensively assess the method's generalization capability? How do they intend to demonstrate that the model can perform well across a variety of brain states, possibly by evaluating on commonly used datasets such as ABIDE I, ABIDE II, ADNI, and ADHD? Given that the portion of HCP data used is only a small part, do they recognize the insufficiency of ABCD data to support their research?

5. The authors state that the criterion for selecting downstream tasks is to "explore fundamental brain-biology and brain-cognition associations." However, does the choice of the sex classification task truly align with this standard, or does it appear somewhat arbitrary? Could the authors consider alternative downstream tasks that may better align with exploring brain-biology and brain-cognition associations?

6. In the MODULARITY EXPERTS ANALYSIS section, the authors claim that "C can be 5 or 7, which are both reasonable and consistent with the typical number of functional modules in neuroscience." Could the authors provide relevant citations from neuroscience literature to substantiate this statement?

7. In the STATE EXPERTS ANALYSIS section, while the authors compare performance differences with varying numbers of state experts, could they include a more thorough and well-reasoned analysis of the experimental results? How does the number of state experts impact the model's performance?

8. Should the experiments in the paper also consider imaging from different brain region templates, such as AAL 90/116, CC200, and BNA246? How might incorporating these templates enhance the study's findings?

1. To my understanding, C and K, # of modularity experts and state experts, are critical hyperparameters of the proposed model. I assume those two parameters were somehow optimized. In Table 1, what were C and F of dFCExpert? Given results in Table 2 and Figure 4, I feel like the performance of dFCExpert is comparable to other models. Does it suggest that the superior performance of dFCExpert largely comes from its optimized hyperparameter rather than conceptual advances?

Below are specific questions related to it:

• Please clarify C and K values used for dFCExpert in Table 1
• Please describe hyperparameter optimization process in Table 1 and other results. In other words, how do we know this optimized parameter set not overfitted to the specific datasets?
• Please provide results comparing dFCExpert to other models when using similar hyperparameter settings. This would help distinguish between gains from conceptual advances versus hyperparameter tuning.

2. As stated in weakness, are there any unique aspects that can be achieved by the proposed model? Observations in Figure 4 are already reported with conventional methods (dFC + K-mean clustering). In addition to increment in accuracy, I suggest the authors perform additional analysis to support the unique strength of the proposed model.

• Please compare the interpretability of learned states between dFCExpert and conventional methods.
• Demonstrate the underlying mechanism how dFCExpert captures individual differences in brain modularity.
• Analyze the stability of identified states across different runs or subsets of data.

3. The authors reported different C/K between HCP and ABCD datasets. I suspect this inconsistency simply suggest that the proposed method is unstable. If the authors think this discrepancy in C and K reflects important neurophysiological difference between two datasets, the authors should further support such results.

• Explain the author's rationale why using different C/K values for each dataset.
• Provide results using consistent C/K values across both datasets to demonstrate stability
• If the authors believe the different values reflect neurophysiological differences in datasets, please consider to provide additional results to support this claim.

Major concerns

• Please consider adding an interpretation of each state and what can be learned from the difference between the female and male state assignment pattern in Section 4.4 (Figure 4). Appendix G.2 visualizes the functional connectivity of each state and the difference between the two sexes, but it is still difficult to interpret the neuroscientific implication because each matrix is small and the colorbar is not centered to zero.
• Please consider adding a subfigure that color-codes the modularity-matched ROIs versus unmatched ROIs with the Schaefer 7 networks. It is not straightforward to see which areas of the brain show alignment with the Schaefer 7 networks in its current form.
• Please clarify which atlas was used for the experiments along with the number of ROIs.

Minor concerns

• Extending the experiments to a clinical fMRI dataset (e.g., ABIDE) would be of interest to the readers in the field.
• The typefaces that denote functions, vectors, and matrices are often inconsistent, which lowers readability. Please consider revising the typefaces of notations to be consistent throughout the paper. For example, $\mathbf{P}$ is a matrix, while $\mathbf{G}$ is a set. $T$ is a scalar, $X_{t}$ is a matrix, and $F_{a}$ is a function, and so on.
• Typo: The S"c"haefer atlas is often referred to as the Shaefer atlas.

Recommendation

I find the strengths of this work outweigh the above concerns. Thus, I recommend accept of the paper.

Will the MoE scheme work for other GNN architectures as well?