NeuroH-TGL: Neuro-Heterogeneity Guided Temporal Graph Learning Strategy for Brain Disease Diagnosis

To Reviewer 4GxD:

We sincerely appreciate the reviewer's recognition of the proposed method and the experimental performance, as well as the valuable comments provided. We provide detailed responses to the constructive comments.

Answers to weaknesses:

W1: Sensitivity Analysis of T and S.

We understand your concerns about the sliding window. In fact, the number of overlapping windows T and window length S jointly determine the construction of the brain network. Specifically, T changes within the set {5, 6, 7, 8, 9, 10}, and S varies within the set {60, 70, 80, 90, 100}. We conducted a grid search on both S and T to ensure the optimal parameter combination. Given space limitations, we explore the impact of T and S on accuracy using the NC vs. MCI and NC vs. TDPD tasks as examples, with the results recorded in Tables R1 and R2, respectively. For NC vs. MCI, the best diagnostic performance is achieved when T=6 and S=90. For NC vs. TDPD, the best diagnostic performance is achieved when T=8 and S=80. Larger values of T and S lead to longer overlapping sequences across windows, which might smooth out valuable dynamic information in the brain. Conversely, smaller values of T and S result in shorter time sequences per window, potentially making the statistical correlation between brain regions unreliable. Therefore, a moderate window size can balance reliable statistical correlation with temporal evolution. We will discuss this in the revised version.

Table R1 Impact of S and T on accuracy for the NC vs. MCI task (%).

Table R2 Impact of S and T on accuracy for the NC vs. TDPD task (%).

W2: Model Stability.

Thanks for the valuable comment. We agree that the proposed method exhibits higher variance, particularly in PD diagnosis tasks. This might be attributed to the following reasons. First, the sample size of the PD dataset is relatively small, which leads to unstable performance of the model on the test set. Moreover, as shown in Tables 4 and 5 in Appendix A, the standard deviation of dynamic brain network analysis methods is generally higher than that of static brain network analysis methods. This phenomenon may be due to the fact that dynamic analysis methods require simultaneous optimization of spatio-temporal feature extraction, leading to a loss function with more local optima. Consequently, the higher standard deviation associated with dynamic methods is a common characteristic in this field. Notably, although our method has a slightly higher standard deviation than the baseline methods, its average diagnostic performance is significantly better than theirs. Overall, the higher standard deviation of our method is acceptable. In future research, we plan to increase the sample size and the number of modalities to further enhance the model's stability.

Answers to Questions:

Q1: Motivation of CNN.

Thanks for the great advice. We chose 2D CNN to process the output of GCN for two main reasons. First, the spatio-temporal heterogeneity of the brain is essentially a local property. After GCN extracts local spatial dependencies, 2D CNN can further model the local temporal correlations of the brain. However, MLP treats all node features equally, which leads to the neglect of local temporal dependencies. Therefore, combining 2D CNN with GCN can capture complementary spatio-temporal associations. Second, in order to intuitively compare the performance of CNN and MLP, we conduct ablation experiments on two datasets. As shown in Tables R3 and R4, CNN outperforms MLP in diagnostic accuracy. The evidence indicates that CNN is a more reasonable choice.
Table R3 Ablation results on the ADNI dataset (%).

Table R4 Ablation results on the PD dataset (%).

Q2: Concatenation Operation.

In fact, we have considered using established recurrent frameworks or attention to obtain global spatio-temporal features. Tables R5 and R6 illustrate the impact of GConvLSTM and attention-based fusion methods on accuracy. The results show that the simple concatenation operation performs comparably to the aforementioned methods, and even outperforms them in some cases. This is likely because the proposed TPGCN can effectively encode the dynamic influence of historical brain states on the current network, and concatenating the outputs of TPGCN directly is already sufficient to characterize global spatio-temporal features. If an additional complex framework is employed, it might cause the model to overfit to the noise within the data, thereby losing critical spatio-temporal patterns and increasing computational costs. Consequently, we employ the efficient concatenation operation to obtain the global spatio-temporal features.

Table R5 Ablation results on the ADNI dataset (%).

Table R6 Ablation results on the PD dataset (%).

Q3: Temporal Resolution.

Thanks for the insightful comment. As you mentioned, the ADNI is a multi-site dataset. NeuroH-TG does not address this temporal inconsistency. We mitigate this challenge from the perspective of data processing. The fMRI data used are acquired by 3T Philips scanners with consistent parameters, and all have the same acquisition duration. The specific scanning parameters are as follows: repetition time=3000ms, echo time=30ms, slice thickness=3.31mm, and the whole-brain images consist of 48 slices. After obtaining the raw fMRI data, we employ the same preprocessing pipeline for brain signal extraction. The preprocessing steps included correction, realignment, and detrending techniques to mitigate the interference from external factors. Finally, we use the same AAL template to parcellate all the fMRI data into 90 brain regions with 140 time points. Therefore, we ensure the temporal resolution consistency of brain graphs across different subjects from the data processing perspective.

To Reviewer 9ryA

We sincerely appreciate the reviewer's assessment that the proposed method is innovative and reasonable, and for providing valuable suggestions. We provide detailed responses to the constructive comments.

Answers to Weaknesses:

W1: Threshold.

The threshold $\alpha$ for dynamic brain networks is the same. In the experiments, $\alpha$ is varied within {0.3, 0.4, 0.5, 0.6, 0.7, 0.8}, and we use grid search to determine the optimal value. The experimental results show that optimal performance can be achieved when $\alpha$=0.6 for different diagnostic tasks.

W2: Comparison based on Graph Transformer.

Following your suggestion, we have added Graphormer and ALTER as comparison experiments. As can be seen from Tables R1 and R2, our proposed method continues to demonstrate superior performance. This is because although the aforementioned two methods can depict the complex spatial structure of the brain, they both overlook the dynamic characteristics of brain networks. In contrast, our method not only effectively captures the inherent neural heterogeneity but also simulates the temporal propagation mechanism of the brain, thus achieving superior results.

Table R1 Classification results on the ADNI dataset (%). task1: NC vs. MCI vs. AD; task2: NC vs. MCI; task3: NC vs. AD; task4: MCI vs. AD.

Table R2 Classification results on the PD dataset (%). task1: NC vs. TDPD vs. PGPD; task2: NC vs. TDPD; task3: NC vs. PGPD; task4: TDPD vs. PGPD.

W3: Source of Comparison Results.

Thanks for the thoughtful suggestion. Due to the use of different datasets, we are unable to directly excerpt experimental results from the original papers. The codes of the comparison methods are all open source, and we use the open source code to obtain experimental results on the datasets used in this paper. To ensure fairness, the hyperparameters of the comparison methods are also determined through grid search.

W4: Explanation of Figure 5.

In Figure 5, different colors indicate the relative importance of brain regions. This importance represents the spatio-temporal connectivity strength of the brain regions. It is worth noting that the statistical p-values of these brain regions are all less than 0.05. This means that the spatio-temporal connectivity strength of these brain regions can effectively distinguish different groups.

Answers to Questions:

Q1: Symbol Definition.

Thanks for your question. Unfortunately, your question is not fully displayed. Considering that there are the same symbols $\parallel \parallel_2$ in Eq (4), Eq (5), and Eq (6), we speculate that you are asking about the meaning of this symbol. $\parallel \parallel_2$ represents the $L_2$ norm. We will clarify this in the revised version to facilitate reader understanding.

Q2: Size of Convolution Kernel.

The size of the convolution kernel is 3×3. We will clarify this in the implementation details.

To Reviewer k1cj:

We sincerely appreciate the reviewer's recognition of the novelty of the proposed method and the comprehensiveness of the experimental analysis. We provide detailed responses to your constructive comments.

Answers to Weaknesses:

W1: Terminology.

Thanks for this kinder reminder. Neuro-heterogeneity indicates that certain brain regions have more extensive connections and more pronounced temporal changes. Topological consistency networks refers to the brain connections that maintain a stable state. Temporal trend networks refer to the brain will dynamically adjust the connections according to cognitive demands while maintaining stable connections of certain nodes. This spatio-temporal coordination forms the neural basis that supports complex cognitive functions. We will provide intuitive definitions in revised version to facilitate understanding for readers outside of neuroimaging.

W2: Figure Caption.

In the revised version, we will provide more detailed explanations for the figures to enhance their readability.

W3: Generalization.

We understand your concern about the model's generalization. NeuroH-TGL shows excellent performance across many diseases, including MCI, AD, TDPD and PGPD, which proves its generalizability to some extent. In future research, we will extend this method to applications such as brain age prediction and brain decoding.

Answers to Questions:

Q1: Small-Scale Ablation.

Following your suggestion, we replaced TPGCN with GRU and LSTM for an ablation study. From Tables R1 and R2, We can find that TPGCN achieves superior performance. This is because GRU and LSTM focus more on the temporal evolution and overlook the topological properties of the brain. TPGCN is capable of modeling the temporal influence of historical brain topology on the current network structure, thereby more comprehensively capturing spatio-temporal dependencies.

Table R1 Results on the ADNI dataset (%). task1: NC vs. MCI vs. AD; task2: NC vs. MCI; task3: NC vs. AD; task4: MCI vs. AD.

Table R2 Results on the PD dataset (%). task1: NC vs. TDPD vs. PGPD; task2: NC vs. TDPD; task3: NC vs. PGPD; task4: TDPD vs. PGPD.

Q2: Subject-Level Interpretability.

Thanks for your valuable feedback. NeuroH-TGL offers subject-level interpretability by calculating the spatio-temporal heterogeneity across all brain regions for each subject, identifying areas with significant changes. Additionally, the spatio-temporal heterogeneity of corresponding brain regions varies among different subjects, which can characterize the diverse impacts of brain diseases on the same brain regions across subjects. Therefore, NeuroH-TGL can offer personalized biomarkers. We apologize that image or file uploads are not permitted in this rebuttal, preventing us from providing visual results.

Q3: Computational Efficiency.

Thanks for the insightful comment. We compare the average time for 1 inference, number of model parameters, and floating-point operations (FLOPs) between NeuroH-TGL and some spatio-temporal models. The results are shown in Table R3. The inference time for NeuroH-TGL is 0.0236 seconds, longer than other spatio-temporal models. This may be because it performs spatio-temporal heterogeneity mining and weighting for multiple brain networks separately, and performs a large number of convolutions during feature extraction. Notably, NeuroH-TGL only requires 0.0924 million parameters and 0.2479 million FLOPs, making it the most efficient among the compared methods. Overall, NeuroH-TGL can achieve superior performance with fewer resources, indicating good scalability. We will clarify this in the discussion section.

Table R3 Computational Efficiency Analysis.

Q4: Sensitivity Analysis of T and S.

We understand your concerns about the parameters related to the window. In fact, the number of overlapping windows T and window length S jointly determine the construction of the brain network. Specifically, T changes within the set {5, 6, 7, 8, 9, 10}, and S varies within the set {60, 70, 80, 90, 100}. We conducted a grid search on both S and T to ensure the optimal parameter combination. Given space limitations, we explore the impact of T and S on accuracy using the NC vs. MCI and NC vs. TDPD tasks as examples, with the results recorded in Tables R4 and R5, respectively. For NC vs. MCI, the best performance is achieved when T=6 and S=90. For NC vs. TDPD, the best performance is achieved when T=8 and S=80. Larger values of T and S lead to longer overlapping sequences across windows, which might smooth out valuable dynamic information in the brain. Conversely, smaller values of T and S result in shorter time sequences per window, potentially making the statistical correlation between brain regions unreliable. Therefore, a moderate window size can balance reliable statistical correlation with temporal evolution. We will discuss this in the revised version.
Table R4 Impact of T and S on accuracy for the NC vs. MCI task (%).

Table R5 Impact of T and S on accuracy for the NC vs. TDPD task (%).

To Reviewer 15Vy:

We sincerely appreciate the reviewer for acknowledging that the proposed method is interesting and for offering valuable suggestions. We provide detailed responses to the constructive comments.

Answers to Weaknesses:

W1: Comparison of Self-Attention.

As you suggested, we employ the self-attention mechanism (SA) to capture spatio-temporal heterogeneity. Tables R1 and R2 show that SA does not effectively improve diagnostic accuracy. In addition, the NeuroH-TGL based on cosine similarity requires only 0.09 million model parameters and 0.25 million FLOPs. After introducing SA, the model parameters and FLOPs increase to 0.11 million and 1.60 million, respectively. Therefore, we chose the efficient cosine similarity to compute spatio-temporal heterogeneity.

Table R1 Accuracy of different methods on the ADNI dataset (%).

Table R2 Accuracy of different methods on the PD dataset (%).

W2: Interpretability Methods.

Thanks for pointing this out. We compare NeuroH-TGL with two interpretability methods: ALTER [1] and MGNN [2]. The results are summarized in Tables R3 and R4. Evidently, NeuroH-TGL outperforms both baselines. ALTER captures long-range and short-range dependencies, but it overlooks the dynamic properties of the brain. MGNN offers module-level interpretability, but it ignores spatio-temporal heterogeneity, leading to less precise characterization of brain activity. By contrast, NeuroH-TGL not only identifies heterogeneous regions that drive network reconfiguration but also models the temporal propagation mechanism of neural activity, thereby achieving the best performance. These results will be included in the revised version.

Table R3 Classification results on the ADNI dataset (%). task1: NC vs. MCI vs. AD; task2: NC vs. MCI; task3: NC vs. AD; task4: MCI vs. AD.

Table R4 Classification results on the PD dataset (%). task1: NC vs. TDPD vs. PGPD; task2: NC vs. TDPD; task3: NC vs. PGPD; task4: TDPD vs. PGPD.

W3: Sensitivity Analysis of T and S.

Thanks for the kinder reminder. The number of overlapping windows T and window length S jointly determine the construction of the brain network. Specifically, T changes within the set {5, 6, 7, 8, 9, 10}, and S varies within the set {60, 70, 80, 90, 100}. We conducted a grid search on both S and T to ensure the optimal parameter combination. Given space limitations, we explore the impact of T and S on accuracy using the NC vs. MCI and NC vs. TDPD tasks as examples, with the results recorded in Tables R5 and R6, respectively. For NC vs. MCI, the best diagnostic performance is achieved when T=6 and S=90. For NC vs. TDPD, the best diagnostic performance is achieved when T=8 and S=80. Larger values of T and S lead to longer overlapping sequences across windows, which might smooth out valuable dynamic information in the brain. Conversely, smaller values of T and S result in shorter time sequences per window, potentially making the statistical correlation between brain regions unreliable. Therefore, a moderate window size can balance reliable statistical correlation with temporal evolution. We will discuss this in the revised version.

Table R5 Impact of T and S on accuracy for the NC vs. MCI task (%).

Table R6 Impact of T and S on accuracy for the NC vs. TDPD task (%).

Answers to Questions:

Q1: $x_i$.

As you expected, $x_i$ represents the time series signal of the $i^{th}$ brain region, and we will clarify this in the revised version.

Q2: Determination of Hyperparameters.

For other hyperparameters, such as the brain network threshold $\alpha$, we also determine it by grid search. Specifically, $\alpha$ is searched within {0.3, 0.4, 0.5, 0.6, 0.7, 0.8}. We incorporated these parameters into an outer loop during model training so that the optimal setting could be identified through grid search. The results show that optimal performance can be achieved when $\alpha$=0.6.

Q3: Mitigating Overfitting.

We understand your concern about overfitting. In fact, to mitigate overfitting, we have taken the following measures. First, we adopt an early stopping mechanism that 80 epochs patience in total 300 epochs. Second, stratified cross-validation is used to evaluate model performance. It ensures that the proportions of different classes are consistent in both the training and validation sets, thereby preventing the model from overfitting to specific subsets.

[1] Long-range Brain Graph Transformer. NeurIPS 2024

[2] Leveraging Brain Modularity Prior for Interpretable Representation Learning of fMRI. IEEE TBME 2024