acceptodds
Under review as a conference paper at ICLR 2027

Spiking-Temporal Directed Hierarchical Propagation Network for Brain Disorders Diagnosis

Abstract

The brain organizes its functions hierarchically. Clarifying how local activity in brain disorders is transformed and propagated across functional hierarchies is essential for pursuing more precise neurobiological markers. Existing diagnosis methods typically construct graphs from inter-regional correlations or aggregate regions into higher-level representations, but they often decouple connectivity from ongoing regional dynamics and rely on undirected interactions, leaving the pathway from local temporal activity to system-level information flow insufficiently modeled. Thus, we propose Spiking-Temporal Directed Hierarchical Propagation Network (ST-DHPNet), a biologically inspired framework for multi-level brain network modeling for brain disorder diagnosis. ST-DHPNet first introduces a Spiking Temporal Brain Encoder that decomposes regional BOLD signals and processes consecutive observations through a stateful spiking encoder, producing activity-sensitive regional representations. It then employs an activity-guided spatial encoder that conditions functional-connectivity attention on regional activity states. Finally, a directional hierarchical propagation module aggregates regions into functional subnetworks using anatomical priors and performs effective-connectivity-guided propagation within and between subnetworks. This design explicitly couples regional temporal dynamics, activity-dependent spatial interactions, and directed subnetwork propagation. Experiments on three brain disorder datasets demonstrate performance improvements over other state-of-the-art methods in brain disorder diagnosis. ST-DHPNet reveals candidate biomarkers at three complementary levels: regional latent spike representations, inter-regional connectivity, and subnetwork coordination, providing an interpretable computational framework for brain disorder classification.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spiking Neural Networks, Brain Disorders Diagnosis

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_vW4UOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper attempts to integrate Spiking Neural Networks (SNNs), functional connectivity (FC), and effective connectivity (EC) with the hierarchical structure of brain function to establish a unified modeling framework—ranging from brain region activity to functional subnetworks—and validates this framework using the ABIDE, ADNI, and REST-meta-MDD datasets.

Critical strengths
  1. The research problem is clearly defined, aiming to unify local temporal dynamics, spatial interactions, and hierarchical propagation.
  2. Systematic comparative experiments were conducted on three datasets, supplemented by multiple sets of ablation studies.
  3. The authors examined the impact of various factors—such as functional network priors, brain region parcellations, and effective connectivity parameters—on the results.
Critical weaknesses
  1. Whether the mathematical implementation of functional network priors truly serves a constraining role.
  2. Whether the so-called "directed information propagation" supports the paper's interpretation of causality and information flow.
  3. Whether the visualization of spiking activity substantiates the conclusions regarding neurobiological markers.
Questions for authors
  1. I notice that the output X_Tout from the Spiking Temporal Brain Encoder module is expected to be a spike train, whereas X_Sout from the Activity-Guide Spatial Brain Encoder has not undergone spiking conversion. How are these two fused in the Directional Hierarchical Propagation Module? X_Tout is a binary sequence (0 or 1) resulting from spiking conversion, while X_Sout appears to be a non-spiking tensor; could the authors please clarify how these two types of data are fused?

  2. In Section 3.3, the paper constrains the associations between brain regions and functional subnetworks by employing an anatomical prior matrix, D. The authors argue that this mechanism guides brain regions to aggregate into functionally coherent subnetworks that are neurobiologically plausible. However, a critical issue regarding this design requires clarification: if the prior matrix utilizes binary (0/1) encoding, setting the scores of disallowed associations to zero is not equivalent to completely masking those associations. For instance, a disallowed brain region–subnetwork association might yield a score of zero after multiplication, yet still receive a non-zero weight following the Entmax operation. Consequently, the current formulation may not effectively achieve the prior-based constraint claimed by the authors. Alternatively, if the prior matrix employs continuous similarity values ​​rather than a binary mask, the authors need to explain the value range, normalization method, and the impact on attention allocation.

  3. The paper employs transfer entropy (TE) to estimate directed statistical dependencies between brain regions and applies it to analyze information propagation both within and between sub-networks. As I understand it, transfer entropy characterizes directed statistical dependency conditioned on historical information and cannot be automatically equated with true neural causality. Currently, the paper does not sufficiently detail the critical implementation aspects of the TE estimation process.

  4. The description of the activity guidance mechanism in the paper mainly stays at the structural level, and it has not yet clearly explained what substantive constraints it introduces compared with the conventional cross-attention mechanism. Since Query and Key come from different data representations, the author should also clarify the input and output dimensions of each projection matrix, as well as the specific brain area relationship corresponding to the attention score.

If the author addresses the questions above, I will raise my score.

3 comments
AuthorsOpenReview ↗

Author Response (1/2)

We are grateful for the reviewers’ insightful comments. Please find our detailed responses below. We hope our responses adequately address the concerns raised and welcome any further questions.

Reply to W1 & Q2. About the role and effectiveness of functional network priors.

We thank the reviewer for pointing out this important distinction. We would like to clarify that the anatomical prior matrix DD is a continuous-valued soft prior which contains continuous spatial similarity scores rather than binary constraints. It modulates the learned ROI-to-subnetwork attention scores before Entmax normalization, thereby introducing an anatomical inductive bias without strictly prohibiting associations with low prior values. Specifically, D∈R200×8D\in\mathbb{R}^{200\times8} represents the associations between 200 ROIs and eight functional subnetworks, including seven cortical networks and one cerebellar/subcortical community. Before attention computation, we apply row-wise L2 normalization across the subnetwork dimension:

$ \widehat D_{ik}=\frac{D_{ik}}{\max(|D_{i,:}|_2,\epsilon)}, $

where the normalized values lie in [0,1][0,1]. The normalized prior is then used to modulate the learned attention scores:

$ \widetilde E=E\odot\widehat D^\top,\qquad W_{\mathrm{attn}}=\operatorname{Entmax}_{}(\widetilde E). $

Therefore, this mechanism provides soft anatomical guidance. The continuous-valued prior modulates the learned attention scores before Entmax normalization, allowing anatomical information to influence the attention allocation without strictly restricting ROI-to-subnetwork associations.

Reply to W2 & Q3. About the interpretation and implementation details of directed information propagation.

We thank the reviewer for this important comment. We agree that transfer entropy (TE) estimated from BOLD signals characterizes directed statistical dependencies rather than true neural causality. Accordingly, the directional propagation in ST-DHPNet refers to computational message passing guided by estimated directional dependencies, rather than direct physiological information flow. Regarding the implementation, TE is estimated from regional BOLD time series during data preprocessing, independently of model training. Specifically, each subject's preprocessed ROI time series is independently standardized and discretized using five states quantile-based binning. Pairwise TE is then estimated in both directions using PyInform, with diagonal entries set to zero. The resulting asymmetric matrix (WEC)(W^{EC}) satisfies (WijEC=TE(i→j))(W_{ij}^{EC}=TE(i\rightarrow j)), where the estimated dependence from ROI ii to ROI jj is used to guide computational message aggregation.

To further examine the contribution of directional connectivity, we conducted additional experiments using four alternative EC configurations: shuffled, symmetric, transposed, and uniform connectivity matrices. Across ABIDE, ADNI, and REST-meta-MDD, the original TE-generated EC consistently achieved the highest mean ACC, AUC, and SEN. In particular, reversing the connectivity directions or removing directional asymmetry generally reduced classification performance, suggesting that preserving the estimated directional structure provides useful discriminative information. We have added explanations for the TE estimation process and the direction propagation in the revised version.

Dataset: ABIDEACC(%)AUC(%)SEN(%)
Shuffled72.27±2.6577.58±2.5277.22±6.60
Symmetric72.32±1.9777.35±1.8577.42±7.90
Transpose72.56±1.7277.75±2.4674.43±5.82
Uniform72.05±2.3977.14±2.4777.31±7.03
Original73.20±2.0678.02±2.2178.51±4.97
Dataset: ADNIACC(%)AUC(%)SEN(%)
Shuffled73.59±3.2578.85±3.1975.26±6.57
Symmetric72.95±3.1678.88±2.1575.00±7.47
transpose73.59±3.4578.68±3.2779.58±9.12
Uniform73.97±1.9178.70±2.3877.11±9.34
Original74.87±2.2079.71±2.3480.52±6.98
Dataset: REST-meta-MDDACC(%)AUC(%)SEN(%)
Shuffled67.57±1.4870.69±2.0171.17±5.06
Symmetric67.70±1.2471.17±1.0371.57±6.12
Transpose67.74±1.2170.34±1.5070.60±5.82
Uniform67.70±0.9970.87±1.3370.98±4.42
Original68.02±2.0171.67±1.7071.81±4.17
AuthorsOpenReview ↗

Author Response (2/2)

Reply to W3. About the neurobiological interpretation of model representations and biomarkers.

We thank the reviewer for this important comment. We agree that model-generated spikes and attention weights should not be directly equated with neuronal firing rates or physiological connectivity strength. Instead, these quantities represent learned temporal responses and interaction weights within the proposed computational framework. Specifically, the spikes generated by our spiking temporal encoder are latent computational events generated from BOLD signals, rather than recorded neuronal action potentials. Each BOLD frame corresponds to a computational timestep rather than a physiological spike timescale. Similarly, attention weights characterize learned interactions between regional activity and connectivity representations, rather than directly measuring physiological connectivity strength. Our interpretability analysis identifies regional and network-level differences that correspond to brain regions and functional systems implicated in previous neuroimaging studies, providing indirect support for their neurobiological plausibility.

Reply to Q1. About the fusion of temporal and spatial representations.

We thank the reviewer for raising this question. We would like to clarify that the output of the Spiking Temporal Brain Encoder, XToutX_T^{out}, is not a binary spike train, but a continuous-valued representation obtained by integrating spike-based and continuous temporal features. Specifically, as shown in Eq. (8), the binary spike representation XIntraX_{\mathrm{Intra}} is first transformed through a learnable linear projection WW, producing continuous-valued features. These projected features are then combined with the continuous inter-segment representation XInterX_{\mathrm{Inter}} through element-wise multiplication:$ X_T^{out}=X_{\mathrm{Trend}}+ X_{\mathrm{Inter}}\odot(X_{\mathrm{Intra}}W).$

In this operation, the projected spike features modulate the inter-segment representations, allowing local spike-based temporal responses to interact with long-range temporal dependencies. The resulting features are further combined with the continuous trend representation through addition, producing the final continuous-valued temporal output XToutX_T^{out}. Subsequently, in the Directional Hierarchical Propagation Module, XToutX_T^{out} is projected into the same feature space as the non-spiking spatial representation XSoutX_S^{out}. The two representations are then fused through element-wise addition:$ X_{st}=W_TX_T^{out}+X_S^{out}, $ where WTW_T is a learnable projection matrix. This operation integrates complementary temporal dynamics and spatial information into a unified ROI-level representation for subsequent hierarchical propagation. Therefore, the fusion is performed between dimensionally aligned continuous-valued representations.

Reply to Q4. About the mechanism and formulation of activity-guided attention.

We thank the reviewer for this valuable comment. We agree that AGA follows the standard cross-attention formulation. Its distinctive contribution lies in the choice of information sources: queries are derived from regional BOLD temporal representations, whereas keys and values are derived from FC representations. Consequently, the attention weight between ROIs ii and jj is conditioned on the temporal representation of ROI ii and the connectivity profile of ROI jj. This introduces an activity-conditioned inductive bias, allowing spatial aggregation weights to depend jointly on regional temporal characteristics and connectivity patterns, rather than relying solely on FC features. This design specifies the information sources used for attention weighting and feature aggregation, providing an architectural inductive bias rather than an additional constraint.

Regarding the implementation details, let Xts∈RN×TX_{ts}\in\mathbb{R}^{N\times T} and Xfc∈RN×NX_{fc}\in\mathbb{R}^{N\times N} denote the BOLD time series and FC matrix, respectively. The temporal and FC embedding layers map T→dT\rightarrow d and N→dN\rightarrow d, producing Hts,Hfc∈RN×dH_{ts},H_{fc}\in\mathbb{R}^{N\times d}. Queries are generated from temporal embeddings, while keys and values are generated from FC embeddings:

$ Q=H_{ts}W_Q,\qquad K=H_{fc}W_K,\qquad V=H_{fc}W_V, $

where WQ,WK,WV∈Rd×dW_Q,W_K,W_V\in\mathbb{R}^{d\times d}. For each head, attention is computed as

Aij=Softmaxj(qiTkjdh)A_{ij} = \text{Softmax}_{j} \left( \frac{q_i^T k_j}{\sqrt{d_h}} \right)

oi=∑j=1NAijvjo_i = \sum_{j=1}^{N} A_{ij} v_j

For each attention head, AijA_{ij} specifies the weight assigned to ROI jj’s FC-derived value when updating ROI ii, conditioned on ROI ii’s BOLD temporal features. Here, the FC representation of ROI jj encodes its connectivity profile with all ROIs. The weights are normalized over key/value ROIs jj, and aggregation is performed across ROIs within each head. The head outputs are then concatenated and projected to obtain the final spatial representation.

Thank you for addressing my questions—best of luck to you! I will increase my rating.

Reviewer_chHjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ST-DHPNet for fMRI-based brain disorder classification. The method combines spiking temporal encoding, activity-guided spatial attention, and directed propagation within and between functional subnetworks. It constructs hierarchical representations using BOLD time series, functional connectivity, and transfer-entropy-based effective connectivity, and evaluates classification performance on three datasets. The paper also examines model-generated spike activity and attention weights as candidate biomarkers.

Critical strengths

S1. The paper addresses a practically meaningful problem of brain disorder classification and attempts to jointly model regional activity, inter-regional connectivity, and functional subnetwork structure.

S2. Using regional temporal information to complement functional connectivity is intuitively motivated, and hierarchical aggregation provides a reasonable way to incorporate prior knowledge of brain functional organization.

S3. The experiments cover three datasets and include component ablations and temporal encoder comparisons. The NYU site-holdout experiment provides generalization evidence beyond random subject splits.

Critical weaknesses

W1. According to Eqs. (17)–(18), intra-subnetwork propagation followed by mean pooling is equivalent to weighting source features by their out-strength. Consequently, it cannot distinguish connectivity structures with the same node out-strengths but different internal routing patterns. Moreover, removing EC also removes connectivity weights and propagation operations, so this ablation cannot determine whether the improvement comes from directionality rather than additional connectivity information or weighted aggregation.

W2. I am not yet convinced that the experiments establish the contribution beyond the inherited components. BrainHGT is acknowledged, but the decomposition also closely follows Leddam [1], including its LD initialization code, without attribution. The relationship to temporal-spatial cross-attention in STDCformer [2] and hierarchical directed modeling in NH-GCAT [3] should be discussed. More importantly, removing EC removes weighted propagation as well as direction, so this ablation does not establish the benefit of directionality. Table 7 is helpful, but the SNN improves ABIDE accuracy over Transformer by only 0.64 points, without a matched-capacity analysis or uncertainty for this difference.

W3. Changes in the model’s internal representations cannot be directly interpreted as physiological changes in the brain. For example, fewer model-generated spikes do not necessarily imply lower neuronal firing rates, and higher attention weights do not necessarily indicate stronger physiological connectivity between brain regions. The NYU experiment reproduces some regional findings, but their physiological meaning still requires further validation.

References

[1] Yu et al. Leddam. ICML 2024. https://proceedings.mlr.press/v235/yu24s.html

[2] Zhang et al. STDCformer. Heliyon, 2024. https://pmc.ncbi.nlm.nih.gov/articles/PMC11734066/

[3] Chen et al. NH-GCAT. arXiv:2511.17622, 2025. https://arxiv.org/abs/2511.17622

Questions for authors

Q1. Are the temporal encoders in Table 7 matched in parameter count and training budget? Does the spiking encoder retain its advantage over these alternatives on ADNI and REST-meta-MDD?

Q2. What evidence supports linking the model-generated spikes and attention weights to physiological activity or connectivity? Does the NYU analysis also reproduce the ROI-level and subnetwork-level connection patterns, in addition to the regional spike differences?

6 comments
AuthorsOpenReview ↗

Author Response (1/3)

We are grateful for the reviewers’ insightful comments. Please find our detailed responses below. We hope our responses adequately address the concerns raised and welcome any further questions.

Reply to W1 & W2. About the intra-subnetwork propagation and directional connectivity.

Thank you for the valuable suggestion. We agree that, for fixed regional features and subnetwork assignments, combining Eqs. (17)–(18) yields Z~2,k=1∣Gk∣∑i∈Gk(∑j∈GkWijEC)xi.\widetilde Z_{2,k} =\frac{1}{|G_k|}\sum_{i\in G_k} \left(\sum_{j\in G_k}W_{ij}^{EC}\right)x_i. This indicates that the intra-subnetwork propagation followed by mean pooling is mathematically equivalent to an out-strength-weighted aggregation of regional features. Therefore, this operation captures the contribution of each ROI through its outgoing effective-connectivity strength, rather than preserving the complete internal routing topology. We have clarified this mathematical property and revised the corresponding descriptions to avoid overstating its ability to characterize fine-grained information routing.

To evaluate whether the estimated directional structure contributes to the model's performance, we therefore conducted additional controlled experiments using transposed, symmetrized, shuffled, and uniform-weight EC matrices, while retaining the same network architecture and weighting and propagation operations. As reported in the tables below, the original EC achieves the highest mean ACC, AUC, and SEN across all three datasets. These results support the utility of the estimated directional weights. Furthermore, the original EC outperforms both shuffled and uniform weights, indicating that its benefits extend beyond simply introducing a weighted aggregation operation.

Dataset: ABIDEACC(%)AUC(%)SEN(%)
Shuffled72.27±2.6577.58±2.5277.22±6.60
Symmetric72.32±1.9777.35±1.8577.42±7.90
Transpose72.56±1.7277.75±2.4674.43±5.82
Uniform72.05±2.3977.14±2.4777.31±7.03
Original73.20±2.0678.02±2.2178.51±4.97
Dataset: ADNIACC(%)AUC(%)SEN(%)
Shuffled73.59±3.2578.85±3.1975.26±6.57
Symmetric72.95±3.1678.88±2.1575.00±7.47
transpose73.59±3.4578.68±3.2779.58±9.12
Uniform73.97±1.9178.70±2.3877.11±9.34
Original74.87±2.2079.71±2.3480.52±6.98
Dataset: REST-meta-MDDACC(%)AUC(%)SEN(%)
Shuffled67.57±1.4870.69±2.0171.17±5.06
Symmetric67.70±1.2471.17±1.0371.57±6.12
Transpose67.74±1.2170.34±1.5070.60±5.82
Uniform67.70±0.9970.87±1.3370.98±4.42
Original68.02±2.0171.67±1.7071.81±4.17
AuthorsOpenReview ↗

Author Response (2/3)

Reply to W2. About tthe relationship between our method and existing approaches.

We thank the reviewer for highlighting these relevant studies and providing the opportunity to further clarify the relationship between our method and existing approaches.

We apologize for inadvertently omitting the citation to Leddam [1] in the original submission. We have now added the corresponding reference and clarified its relationship to our method in the revised manuscript. Importantly, our primary contribution lies not in the decomposition strategy itself, but in the design of the overall dual-path spiking temporal encoder, which integrates decomposition-based temporal representations with complementary spiking dynamics to capture diverse temporal dependencies in BOLD signals. Unlike Leddam, which primarily focuses on multivariate time series forecasting, our temporal encoder is specifically designed to learn complementary temporal representations for brain disorder classification.

We thank the reviewer for highlighting these relevant studies. Although STDCformer and NH-GCAT explore related concepts, our work differs in its modeling focus and architectural design. Regarding STDCformer [2], both methods employ temporal-spatial cross-attention to integrate temporal and spatial information. However, their implementations differ in how these representations interact. STDCformer focuses on temporal-spatial feature interaction, whereas our Activity-Guided Attention (AGA) specifically uses regional BOLD temporal representations to construct queries and functional connectivity (FC) representations to construct keys and values. This allows regional temporal activity to guide the aggregation of inter-regional connectivity features. Furthermore, unlike NH-GCAT [3], which explores hierarchical directed brain-network modeling through variational latent causal attention, our Directional Hierarchical Propagation Module explicitly uses transfer entropy (TE)-estimated directed statistical dependencies to guide feature aggregation within and between functional subnetworks, together with anatomical-prior-guided ROI-to-subnetwork aggregation. Thus, the two approaches differ in their mechanisms for constructing and utilizing directed dependencies in hierarchical brain representations.

Importantly, our ablation studies further demonstrate the respective contributions of these components within the proposed framework. As reported in the table below, removing the spiking temporal encoder, AGA, or hierarchical aggregation consistently degrades performance across the three datasets. In particular, removing hierarchical aggregation leads to substantial performance drops, while removing AGA also results in consistent degradation across all datasets. Furthermore, removing the decomposition component or the TE-guided directional propagation also reduces performance. These results indicate that the overall performance is not attributable to the decomposition strategy or any individual existing operation alone, but benefits from the complementary roles of temporal encoding, activity-guided spatial modeling, and hierarchical representation learning.

Dataset: ABIDEACC(%)AUC(%)SEN(%)
w/o AGA72.71±2.3877.58±2.6975.06±7.19
w/o Decomposition72.02±1.5977.41±2.7873.19±7.91
w/o STBE72.36±1.9877.23±2.1775.49±7.06
w/o Aggregation67.93±2.5171.22±1.2568.06±6.15
w/o EC71.53±1.9677.38±1.8674.72±7.93
ST-DHPNet73.20±2.0678.02±2.2178.51±4.97
Dataset: ADNIACC(%)AUC(%)SEN(%)
w/o AGA74.61±1.9779.26±2.6378.42±5.61
w/o Decomposition72.43±4.5679.39±2.7671.84±7.89
w/o STBE73.59±2.3178.08±2.4280.26±6.78
w/o Aggregation68.72±3.9373.01±4.5769.74±3.80
w/o EC73.20±2.4679.16±2.6175.52±4.99
ST-DHPNet74.87±2.2079.71±2.3480.52±6.98
Dataset: REST-meta-MDDACC(%)AUC(%)SEN(%)
w/o AGA67.27±1.1571.60±1.1371.40±4.47
w/o Decomposition67.53±1.2770.98±1.2471.45±5.88
w/o STBE67.23±0.9371.51±1.0870.12±6.63
w/o Aggregation63.22±1.3465.56±1.4668.35±9.28
w/o EC67.81±1.2371.35±1.1271.23±4.27
ST-DHPNet68.02±2.0171.67±1.7071.81±4.17

We have modified the related work and methodology sections to clarify these relationships and provide a more explicit explanation of the respective roles of these components.

AuthorsOpenReview ↗

Author Response (3/3)

Reply to Q1 & W2. About parameter count and training budget and performance of temporal encoders.

We thank the reviewer for this valuable comment. To ensure a fair comparison, we replaced only the SNN temporal encoder with LSTM, GRU, TCN, or Transformer encoders, while keeping the remaining architecture, feature dimensions, data splits, and training settings unchanged. The resulting models have comparable parameter counts, ensuring a closely matched model capacity. We further evaluated their computational costs and total training times on ABIDE, as summarized below.

MethodParams (M)Total Training Time (s)FLOPs (×108\times 10^{8})
LSTM5.431028.3
GRU5.421067.5
TCN5.42968.1
Transformer5.4631819.9
Spiking temporal encoder5.521395.7

The spiking temporal encoder achieves the lowest computational cost in terms of FLOPs. Although its training time is longer than those of LSTM, GRU, and TCN, it remains substantially shorter than that of Transformer. These results suggest that the proposed spiking encoder offers a favorable balance between computational complexity and training efficiency under comparable model capacities.

To further evaluate its generalizability, we extended the temporal encoder comparisons to ADNI and REST-meta-MDD, using the same experimental settings. The results demonstrate that the proposed spiking temporal encoder maintains the highest mean ACC, AUC, and SEN across both datasets. These findings suggest that the proposed spiking temporal encoder provides competitive computational efficiency and classification benefits across different brain disorder datasets.

Model (Dataset: ADNI)ACC(%)AUC(%)SEN(%)
GRU74.62±2.2179.27±2.9477.11±9.42
LSTM73.72±2.5278.03±2.2475.26±10.86
TCN73.72±1.9278.57±2.3573.16±9.02
Transformer72.69±3.1477.22±2.8172.37±8.34
Spiking temporal encoder74.87±2.2079.71±2.3480.52±6.98
Model (Dataset: REST-meta-MDD)ACC(%)AUC(%)SEN(%)
GRU66.16±1.9369.83±2.1470.73±4.13
LSTM66.40±1.1470.66±1.8171.47±5.54
TCN65.84±1.1569.50±1.5869.11±5.96
Transformer65.46±2.1668.70±2.2666.13±7.16
Spiking temporal encoder68.02±2.0171.67±1.7071.81±4.17

Reply to Q2 & W3. About the physiological interpretability.

We thank the reviewer for this important comment. We agree that model-generated representations should not be directly equated with physiological measurements. Specifically, the spikes generated by our spiking temporal encoder are latent computational events generated from BOLD signals, rather than recorded neuronal action potentials. As described in Section 3.1, each BOLD frame corresponds to a computational timestep rather than a physiological spike timescale. Therefore, differences in model-generated spike rates reflect variations in learned temporal representations, rather than direct changes in neuronal firing rates. Similarly, attention weights in our activity-guided spatial encoder characterize learned interactions between regional activity and connectivity representations, rather than directly measuring physiological connectivity strength. These representations provide a means to investigate disease-related patterns learned by the model. Our interpretability analysis identifies regional and network-level differences that correspond to brain regions and functional systems implicated in previous neuroimaging studies, providing indirect support for their neurobiological plausibility. Furthermore, the held-out NYU evaluation reproduces significant group differences in several frontal regions, supporting the cross-site reproducibility of the model-generated patterns.

To further assess the cross-site reproducibility of the learned interaction patterns, we extended the NYU analysis to both ROI-level and subnetwork-level attention. Specifically, we compared patient–control mean-difference matrices between the primary analysis and the NYU cohort. At the subnetwork level, 71.43% of directed connections exhibited consistent difference directions. At the ROI level, the directional agreement was 54.40%. These results suggest that the model-generated attention difference patterns are partially reproduced at both levels, with stronger consistency at the subnetwork level. In the revised manuscript, we have added ROI-level and subnetwork-level visualizations of the NYU site results to further illustrate the reproduce of the model-generated attention patterns.

Reviewer_Du9xOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a biologically inspired framework for multi-scale brain disorder diagnosis. The method integrates three components: a spiking temporal brain encoder, an activity-guided spatial encoder, and a directional hierarchical propagation module. It couples regional temporal dynamics, activity-dependent spatial interactions, and subnetwork propagation for comprehensive analysis. The proposed ST-DHPNet method is evaluated on three datasets. It achieves the highest reported mean accuracy, AUC, and sensitivity among the compared methods across all three datasets.

Critical strengths

A main strength of this paper is the integration of regional temporal dynamics, functional connectivity, and hierarchical directed propagation within a unified architecture. Rather than relying exclusively on static functional connectivity, the method incorporates temporal information into spatial representations and uses estimated directed dependencies to model interactions across functional subnetworks. This is a meaningful architectural contribution for brain network representation learning. The experimental evaluation is also reasonably comprehensive; it covers three disorders and a diverse set of baselines. The held-out NYU-site experiment and the comparisons with alternative temporal encoders provide useful additional evidence beyond standard random-split classification results.

Critical weaknesses

A concern is whether the proposed directional propagation captures meaningful brain information flow. The method uses transfer entropy estimated from BOLD signals as effective connectivity, but transfer entropy represents directed statistical dependence rather than necessarily causal neural interactions. Given the temporal resolution and hemodynamic characteristics of fMRI, interpreting these dependencies as biological information propagation requires caution. Although the authors examine different history lengths for transfer entropy, they do not directly establish the reliability or physiological validity of the estimated directions. Comparisons with e.g. reversed, symmetrized, or randomly permuted directional connectivity would help determine whether the directional structure itself is responsible for the reported improvements.

Questions for authors
  1. How sensitive is the proposed directional propagation to the quality of transfer entropy estimation? Could you compare the original EC matrix against its transpose, a symmetrized version, and a randomly shuffled version to demonstrate that the estimated directionality is important?

  2. The improvements from activity-guided attention are relatively small, particularly on ADNI. Could you explain what additional information this mechanism captures and how significant are the improvements?

  3. What is the computational overhead of estimating transfer entropy and training the spiking temporal encoder?

3 comments
AuthorsOpenReview ↗

Author Response (1/2)

We are grateful for the reviewers’ insightful comments. Please find our detailed responses below. We hope our responses adequately address the concerns raised and welcome any further questions.

Reply to W & Q1. About the directional propagation.

We thank the reviewer for this insightful comment. We agree that transfer entropy estimated from BOLD signals reflects directed statistical dependencies rather than necessarily causal neural interactions.

To evaluate whether the estimated directional structure contributes to the model's performance, we conducted additional experiments on three datasets using four alternative EC configurations: (1) Shuffled, where connectivity entries are randomly permuted to disrupt the original connectivity structure; (2) Symmetric, where the EC matrix is symmetrized to remove directional asymmetry; (3) Transpose, where the directions of connectivity are reversed; and (4) Uniform, where uniform connectivity weights are used to eliminate differences in connection strength.

Dataset: ABIDEACC(%)AUC(%)SEN(%)
Shuffled72.27±2.6577.58±2.5277.22±6.60
Symmetric72.32±1.9777.35±1.8577.42±7.90
Transpose72.56±1.7277.75±2.4674.43±5.82
Uniform72.05±2.3977.14±2.4777.31±7.03
Original73.20±2.0678.02±2.2178.51±4.97
Dataset: ADNIACC(%)AUC(%)SEN(%)
Shuffled73.59±3.2578.85±3.1975.26±6.57
Symmetric72.95±3.1678.88±2.1575.00±7.47
transpose73.59±3.4578.68±3.2779.58±9.12
Uniform73.97±1.9178.70±2.3877.11±9.34
Original74.87±2.2079.71±2.3480.52±6.98
Dataset: REST-meta-MDDACC(%)AUC(%)SEN(%)
Shuffled67.57±1.4870.69±2.0171.17±5.06
Symmetric67.70±1.2471.17±1.0371.57±6.12
Transpose67.74±1.2170.34±1.5070.60±5.82
Uniform67.70±0.9970.87±1.3370.98±4.42
Original68.02±2.0171.67±1.7071.81±4.17

Across all three datasets, the original EC consistently achieves the highest mean ACC, AUC, and SEN. In particular, reversing the connectivity directions or removing directional asymmetry generally leads to performance degradation, suggesting that the original directed dependencies provide useful discriminative information. The results with shuffled and uniform connectivity further indicate that preserving the estimated connectivity structure is beneficial for classification.

Reply to Q2. About the activity-guided attention.

We thank the reviewer for this valuable comment. On ADNI, although the accuracy gain is modest, activity-guided attention increases sensitivity from 78.42% to 80.52%, an improvement of 2.10%. Its benefit is therefore more apparent in patient detection than in overall accuracy. Specifically, AGA allows regional temporal information to directly guide spatial feature aggregation. Queries are constructed from regional BOLD time series, while keys and values are constructed from FC features. Since FC summarizes inter-regional statistical dependencies without explicitly preserving their temporal organization, BOLD-based queries provide complementary temporal information that enables the model to adaptively aggregate connectivity features according to regional activity patterns.

To further assess the statistical significance of these improvements, we conducted two-sided paired tt-tests between the full ST-DHPNet and its variant without AGA across ten matched runs using identical random seeds and data partitions. The results show that AGA improves the mean ACC, AUC, and SEN, with the improvement in SEN reaching statistical significance (p=0.034<0.05p=0.034<0.05). These findings support the effectiveness of AGA, particularly in enhancing patient identification.

AuthorsOpenReview ↗

Author Response (2/2)

Reply to Q3. About the computational overhead of transfer entropy estimation and spiking temporal encoder training.

Thank you for asking about computational overhead. Transfer entropy (TE) is estimated during data preprocessing, independently of model training, and therefore does not introduce additional computational overhead during the training of the spiking temporal encoder. We benchmarked 10 randomly selected subjects from each dataset using 200 ROIs. The measured preprocessing times were 0.408±0.009 s per subject on ABIDE, 0.401±0.003 s on ADNI, and 0.413±0.020 s on REST-meta-MDD.

To evaluate the computational efficiency of the spiking temporal encoder, we compared it with LSTM, GRU, TCN, and Transformer alternatives on the ABIDE dataset. For a fair comparison, we replaced only the temporal encoder while keeping the remaining architecture, feature dimensions, data splits, and training settings unchanged. All models have comparable parameter counts, ensuring closely matched model capacities. We measured their parameter counts, FLOPs, and total training times, as summarized below.

MethodParams (M)Total Training Time (s)FLOPs (×108\times 10^{8})
LSTM5.431028.3
GRU5.421067.5
TCN5.42968.1
Transformer5.4631819.9
Spiking temporal encoder5.521395.7

The spiking temporal encoder achieves the lowest FLOPs among all alternatives, indicating reduced theoretical computational complexity. Although its training time is longer than those of LSTM, GRU, and TCN, it is substantially shorter than that of Transformer. These results demonstrate that the proposed spiking encoder offers a favorable trade-off between computational complexity and training time under comparable model capacities.

Thanks for the detailed responses to my comments. I appreciate your efforts to address the concerns and questions raised in my review. The responses provide some useful explanations. I'm generally positive about the overall quality of this paper, but not all my concerns have been sufficiently resolved. I maintain my original scores.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.