acceptodds
Under review as a conference paper at ICLR 2027

Context-Conditioned Anomalous Subgraph Representation Learning for Brain Disease Detection

Abstract

Existing approaches for brain network analysis typically perform whole-graph classification or rely on dynamic functional connectivity sequences, overlooking the fact that healthy brains exhibit a stable contextual regularity where local subnetworks are consistently constrained by their specific functional contexts. We discover that this fundamental organizational principle is largely preserved in neuropsychiatric disorders, with disease-related alterations manifesting primarily as subtle perturbations to its concrete instantiation. Context-conditioned anomalous subgraph representation construction module is proposed to localize abnormal brain subnetwork. Role-aware brain subnetwork guidance encoder is introduced to encode multi-modal deviation directions, while anomaly-guided signed brain graph network is designed to propagate subgraph-level signals across the whole brain. Theoretically, we establish controlled false-positive bounds, abnormal detectability conditions, and region recovery guarantees via total-variation-regularized residual fields, showing that CCA-SRL provably distinguishes normal from abnormal brains when local subnetworks exhibit sufficiently strong conditional-representation shifts. Empirically, our model achieves new state-of-the-art performance on three public datasets—ADHD-200, OpenNeuro-ds002424, and ABIDE-II—outperforming 13 baselines.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain Disease Detection, Subgraph Representation, Graph Neural Network

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_n9EuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CCA-SRL, which models how local subnetwork features depend on their functional context, then uses deviations from a control-trained model to guide signed graph learning.

Critical strengths
  • Context-conditioned deviations provide a useful alternative to treating local subnetworks independently.
  • Matched/mismatched contexts, corruption experiments, split-half analyses, and held-out-site checks support the existence of contextual regularity.
Critical weaknesses
  • Candidates are mined from diagnostic differences and then selected by disorder–control anomaly-score separation. This is a supervised diagnostic pipeline with normative features; demonstrate that localization remains meaningful beyond selecting discriminative regions.
  • Contextual predictability is not necessarily disease-specific. Internal and context features come from the same correlation matrix and may share global strength, motion, or acquisition effects.
Questions for authors
  • Are score-reference distributions calibrated separately by candidate, or pooled across heterogeneous candidates?
  • Existing work explicitly explores subgraph encodings for brain networks [1-3]. Why these works are not included for discussion and comparison?

[1] Contrastive Graph Pooling for Explainable Classification of Brain Networks. TMI 2024 [2] Long-range brain graph transformer. NeurIPS 2024 [3] Geo-Mamba: Geometry-informed state-space learning of functional brain organization. MIA 2026

5 comments

Weakness 1

Thank you for this question. Diagnostic labels are needed in our candidate-discovery procedure to identify connectivity differences between the disorder and control groups and to select candidates with greater anomaly-score separation. For each selected candidate, the model then compares each subject’s internal subnetwork representation with the expectation given its functional context, using a healthy-control reference. We provide three complementary observations on the disease-related meaning of these deviations:

Patient–control differences. In ADHD-200, the representative four-ROI DMN–DAN candidate has higher mean conditional NLL in patients than in healthy controls (13.248 vs. 6.099; Welch BH-adjusted q=0.00853q=0.00853; paper Appendix I, Table 17). This is a group-level difference for a candidate selected using diagnostic labels.

Within-patient symptom association. Symptom scales were not used for candidate discovery, normality-model fitting, or classifier training. In the OpenNeuro-ds002424 ADHD cases with available symptom scores, we tested the covariate-adjusted association between the conditional anomaly measure and symptom severity (Table N1).

The positive association in Table N1 provides clinical evidence beyond the binary diagnostic labels used for candidate selection.

Table N1. Within-patient symptom association in OpenNeuro-ds002424 ADHD cases.

CasesPhenotypePartial Spearman ρ\rho95% CIPermutation ppBH qq
34ADHD_Total_Raw (T1)+0.4286[0.0108, 0.7116]0.02140.0428

External literature. Liu et al. [1] report ADHD-related connectivity differences involving bilateral angular gyri and superior parietal lobules in an independent adult cohort, corresponding to the four ROI endpoints in our representative candidate. Norman et al. [2] report altered DMN–DAN connectivity in a large ADHD mega-analysis, consistent with our network-level finding; their cohort includes ADHD-200 participants. These observations provide ROI- and network-level convergence, not independent replication of the exact extracted edges or clinical ground-truth localization.

References

[1] Liu, L., Chen, D., Huang, F., et al. Interference of default mode on attention networks in adults with attention-deficit/hyperactivity disorder and its association with genetic variants and treatment outcomes. CNS Neuroscience & Therapeutics, 30(8):e14900, 2024. doi:10.1111/cns.14900.

[2] Norman, L. J., Sudre, G., Price, J., Shastri, G. G., and Shaw, P. Evidence from “big data” for the default-mode hypothesis of ADHD: a mega-analysis of multiple large samples. Neuropsychopharmacology, 48:281–289, 2023. doi:10.1038/s41386-022-01408-z.

Weakness 2

Thank you for this concern. CCA-SRL identifies disorder-related candidate subnetworks using training-set disorder–control differences. Internal and context features derived from the same functional connectivity matrix may share global, motion, or acquisition-related influences. This research mainly focuses on the role of the conditional relationship between a subnetwork's internal functional organization and its surrounding functional context in brain disease diagnosis. Unlike conventional approaches that primarily characterize connectivity patterns themselves, we investigate how deviations from this healthy context–subnetwork relationship provide disease-relevant information. Specifically, we learn the conditional relationship from healthy controls and quantify subject-specific departures from the expected internal organization given the surrounding functional context, using these deviations to guide whole-brain classification.

Question 1

Calibration is candidate-specific. Each subject’s score for candidate kk is compared with healthy-control scores for the same candidate kk.

Question 2

Thank you for pointing out these relevant works. Table N2 reports our comparison with ContrastPool, ALTER, and Geo-Mamba on all three datasets. CCA-SRL has the highest ACC on all three datasets and the highest AUC on OpenNeuro-ds002424 and ABIDE-II; on ADHD-200, its AUC is 63.04%, compared with 63.08% for ALTER. Results are four-fold mean ± standard deviation. pp-values compare each baseline with CCA-SRL using a two-sided paired tt-test on corresponding four-fold results (n=4n=4).

Table N2. Comparison with additional baselines.

MethodOpenNeuro-ds002424 ACC (%)pACCp_{\mathrm{ACC}}OpenNeuro-ds002424 AUC (%)pAUCp_{\mathrm{AUC}}ADHD-200 ACC (%)pACCp_{\mathrm{ACC}}ADHD-200 AUC (%)pAUCp_{\mathrm{AUC}}ABIDE-II ACC (%)pACCp_{\mathrm{ACC}}ABIDE-II AUC (%)pAUCp_{\mathrm{AUC}}
ALTER60.89±5.410.00087767.46±2.400.01083357.83±4.110.01071563.08±0.830.99085256.90±4.680.04788759.10±3.960.035624
ContrastPool54.81±7.320.00583654.74±5.880.00076756.86±2.120.00023259.86±2.000.27751554.13±4.060.01284156.67±3.800.010957
Geo-Mamba55.96±5.840.00379561.12±2.580.00275253.15±4.880.01192853.72±5.930.03509758.54±3.740.04782362.94±1.910.041107
CCA-SRL70.38±4.76—74.52±4.51—64.33±2.14—63.04±5.40—63.07±0.73—64.19±1.18—

Dear Reviewer n9Eu,

Thank you for highlighting the questions surrounding the clinical interpretation of the identified subnetworks and the relationship of our approach to existing brain graph methods. Following your suggestions, we have provided additional evidence and methodological clarifications:

  • Clinical relevance (W1): provided patient–control differences in conditional anomaly scores, a within-patient association with ADHD symptom severity, and supporting evidence from external neuroimaging studies.
  • Context-conditioned representations (W2): clarified the motivation and methodological scope of modeling deviations from the healthy relationship between a subnetwork's internal organization and its surrounding functional context.
  • Candidate-specific calibration (Q1): clarified that each candidate's anomaly score is calibrated against its corresponding healthy-control reference distribution.
  • Additional baselines (Q2): added comparisons with ContrastPool, ALTER, and Geo-Mamba across all three datasets, including paired statistical tests (Table N2).

We would be grateful to know whether these responses address your principal concerns, and we would be happy to provide further clarification on any remaining questions.

Best regards,
The Authors

Reviewer_qzZ5OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses the needs of brain disease detection and abnormal brain network localization in computational neuroimaging. It proposes and validates a fundamental organizational principle in healthy brain functional networks, termed subnetwork context constraint (SCC). Specifically, the internal connectivity patterns of local subnetworks are systematically constrained by their surrounding functional context, boundary interactions, and network roles. Neurological and psychiatric disorders do not completely disrupt this fundamental organizational principle but primarily manifest as subtle, localized perturbations in its specific manifestations. Building on this key insight, the authors develop a context-conditioned anomalous subgraph representation learning (CCA-SRL) framework. The framework sequentially employs a context-conditioned anomalous subgraph construction module to identify candidate abnormal brain regions, a role-aware subgraph-guided encoder to characterize multimodal deviation directions, and an anomaly-guided signed brain graph network to propagate subgraph-level anomaly signals throughout the whole brain. Finally, a reliability-calibrated gating mechanism is introduced to fuse features for disease classification. Theoretical analyses rigorously establish that the model provides controlled false-positive bounds, anomaly detectability conditions, and regional recovery guarantees. Extensive experiments on three public datasets, ADHD-200, OpenNeuro-ds002424, and ABIDE-II, demonstrate that the proposed method outperforms 13 baseline methods and achieves state-of-the-art performance. Additional experiments systematically validate the robustness of the SCC principle, the necessity of individual model components, and the effectiveness of abnormality localization.

Critical strengths
  1. The first key strength of this work is the introduction of subnetwork context constraint (SCC) as a fundamental organizational principle of brain functional networks, based on which the authors develop the context-conditioned anomalous subgraph representation learning framework CCA-SRL. Moving beyond conventional whole-graph classification and isolated subgraph metrics, the proposed approach defines abnormalities in terms of the consistency between local subnetworks and their functional context. This perspective better captures the hierarchical and relational nature of brain functional organization, establishes a new paradigm for network-based brain disease detection, and constitutes the central innovative contribution of the work, substantially enhancing its academic value.

  2. The second key strength lies in the completeness and rigor of the evidential chain. At the theoretical level, the authors formally establish the model's controlled false-positive rate, anomaly detectability, and abnormal region recovery guarantees, providing explicit statistical bounds. At the experimental level, the method achieves state-of-the-art (SOTA) performance on three public datasets, while more than ten supplementary experiments systematically validate the robustness of the core principle, the necessity of individual components, and the effectiveness of abnormality localization. These analyses rule out multiple alternative explanations, lending strong credibility to the conclusions and demonstrating a high level of technical maturity. Collectively, they provide a solid foundation for the paper's acceptance.

Critical weaknesses
  1. Lack of independent validation for anomalous subgraph localization. The paper evaluates its localization capability primarily through controlled synthetic perturbations of brain graphs. Although these experiments demonstrate that the method can recover artificially introduced anomalies, they do not establish whether the identified subnetworks correspond to clinically meaningful abnormalities in real patient populations. The brain-region visualizations are descriptive post hoc analyses rather than independent validation, leaving the real-world clinical relevance and reliability of the proposed localization mechanism insufficiently established.

  2. Dependence on disorder labels during candidate subgraph discovery. Despite framing the method around context-consistency-based anomaly detection, CCA-SRL uses disorder–control differences to construct and select candidate subnetworks, introducing label supervision into the anomaly discovery process. Although the ablation studies show that candidate statistics alone cannot explain the full model's performance gains, they do not isolate the effect of label-informed candidate selection or establish that the approach remains effective without diagnostic labels. The method's applicability to unseen disease subtypes, rare disorders, or settings without labeled control groups therefore remains insufficiently demonstrated.

Questions for authors
  1. How can we confirm that the identified anomalous subnetworks correspond to clinically meaningful disease-related abnormalities, rather than merely discriminative patterns between diagnostic groups? The synthetic perturbation experiments validate localization under controlled conditions, but the real-data visualizations are explicitly descriptive post-hoc analyses and do not provide independent confirmation of disease-related localization.

  2. To what extent does the reliance on disorder–control labels during candidate subgraph discovery limit the applicability of CCA-SRL as a general context-conditioned anomaly detection framework? While ablation studies show that candidate statistics alone do not account for the full model’s performance, it remains unclear whether the learned representations can identify anomalous subnetworks in unseen disorders or settings without diagnostic labels.

  3. While the subnetwork contextual constraint (SCC) has been validated to generalize across acquisition sites, does the full CCA-SRL classification pipeline (including candidate subgraph discovery and guided graph encoding) generalize to previously unseen sites? An explicit leave-one-site-out evaluation of classification performance would help disentangle disease-related patterns from site-specific confounders.

7 comments

Weakness 1 / Question 1

Thank you for this question. We provide further clarification and supporting evidence regarding the clinical relevance of the identified subnetworks:

(1) Diagnostic contribution. CCA-SRL uses context-conditioned abnormal-subgraph evidence to guide whole-brain classification. In the component ablation (Table 2 of the paper), removing the subgraph branch reduces performance, indicating that this evidence contributes to diagnosis. CCA-SRL achieves state-of-the-art performance across the three datasets, including the highest ACC among the compared methods.

(2) Perturbation analysis. As reported in the paper (Appendix F, Table 12), after a controlled perturbation is introduced into a target subnetwork, the target ranks first with accuracy 0.939–0.996 across the three datasets; the corresponding random non-target controls are 0.021–0.049. This indicates that the anomaly score can identify the perturbed subnetwork under controlled conditions.

(3) Patient–control differences and individual scores. In ADHD-200, the representative four-ROI DMN–DAN candidate has higher mean conditional NLL in patients than controls (13.248 vs. 6.099; Welch BH-adjusted q=0.00853q=0.00853; paper Appendix I, Table 17). This shows a group difference in conditional deviation for this candidate. Its score is calculated from each subject's own functional connectivity and context, providing subject-specific evidence rather than a single group-level score.

(4) Symptom association. To examine clinical relevance beyond diagnostic labels, we tested the association between conditional anomaly scores and symptom severity within patients. Symptom scales were not used for candidate discovery, reference-model fitting, or classifier training. Table Q1 reports the covariate-adjusted association in the OpenNeuro-ds002424 ADHD cases with available symptoms.

As shown in Table Q1, higher conditional anomaly scores are associated with greater ADHD symptom severity within the available patient sample.

Table Q1. Within-patient symptom association in OpenNeuro-ds002424 ADHD cases.

CasesPhenotypePartial Spearman ρ\rho95% CIPermutation ppBH qq
34ADHD_Total_Raw (T1)+0.4286[0.0108, 0.7116]0.02140.0428

(5) Convergence with external studies. Liu et al. [1] report ADHD-related connectivity differences involving bilateral angular gyri and superior parietal lobules in an independent adult cohort, corresponding to the four ROI endpoints in our representative candidate. Norman et al. [2] report altered DMN–DAN connectivity in a large ADHD mega-analysis, consistent with our network-level finding. The Norman cohort includes ADHD-200 participants. These findings support ROI- and network-level convergence, rather than independent replication of the exact extracted edges or clinical ground-truth localization.

References

[1] Liu, L., Chen, D., Huang, F., et al. Interference of default mode on attention networks in adults with attention-deficit/hyperactivity disorder and its association with genetic variants and treatment outcomes. CNS Neuroscience & Therapeutics, 30(8):e14900, 2024. doi:10.1111/cns.14900.

[2] Norman, L. J., Sudre, G., Price, J., Shastri, G. G., and Shaw, P. Evidence from “big data” for the default-mode hypothesis of ADHD: a mega-analysis of multiple large samples. Neuropsychopharmacology, 48:281–289, 2023. doi:10.1038/s41386-022-01408-z.

Weakness 2 / Question 2

Thank you for this question. CCA-SRL is a supervised diagnostic framework with context-conditioned anomaly representations. As described in Section 3.1, diagnostic labels are used to construct candidate subnetworks from disorder–control connectivity differences and to select candidates according to anomaly-score separation.

After candidate selection, the normality model is fitted on healthy controls. For each subject, it measures how the candidate subnetwork’s internal representation deviates from the pattern expected given that subject’s functional context. Thus, labels determine which candidates are examined, while the conditional residual provides subject-specific deviation evidence for those candidates. This design requires labeled training data for candidate discovery and healthy controls to establish the reference model.

Question 3

Yes. We evaluated the complete CCA-SRL classification pipeline, including candidate discovery and guided graph encoding, using leave-one-site-out (LOSO) evaluation on ADHD-200 and ABIDE-II.

For LOSO, CCA-SRL and the five baselines use the same random seed, training/validation/test partitions and held-out subjects within each dataset. Neither model fitting nor validation uses the held-out site. Tables Q2–Q3 report site-macro and individual-site ACC/AUC. ACC is averaged over all seven ADHD-200 and fifteen ABIDE-II sites; AUC is averaged over the six and fourteen sites, respectively, that contain both classes.

Tables Q2 and Q3 report LOSO results for individual sites and site-macro averages. CCA-SRL has the highest site-macro AUC among the compared methods on ADHD-200 (55.07%) and ABIDE-II (61.32%). On ABIDE-II, Naive Bayes has a slightly higher site-macro ACC (54.60% vs. 54.13%), while CCA-SRL achieves a higher site-macro AUC (61.32% vs. 51.25%).

Table Q2. ADHD-200 held-out sites (ACC / AUC, %; — denotes undefined AUC for single-class sites).

Held-out siteCCA-SRLNaive BayesHGNNBrainGNNBrainGBBrainOOD
KKI71.28 / 55.5460.64 / 46.7273.40 / 44.0671.28 / 44.7559.57 / 44.1772.34 / 55.77
NYU45.53 / 55.5142.02 / 46.8542.80 / 49.2443.19 / 56.0942.02 / 47.7642.80 / 49.99
NeuroIMAGE55.56 / 61.5050.00 / 43.7550.00 / 50.0851.39 / 50.6952.78 / 51.7044.44 / 40.28
OHSU61.06 / 47.9747.79 / 28.2155.75 / 34.5561.95 / 52.6661.06 / 43.3958.41 / 36.28
Peking57.96 / 51.6051.02 / 44.3251.02 / 42.7751.84 / 40.3957.55 / 44.5359.18 / 51.64
Pittsburgh83.51 / 58.3358.76 / 72.1883.51 / 50.0061.86 / 79.8495.88 / 40.3292.78 / 75.54
WashU98.36 / —72.13 / —93.44 / —83.61 / —88.52 / —98.36 / —
Site-macro mean67.61 / 55.0754.62 / 47.0064.28 / 45.1260.73 / 54.0765.34 / 45.3166.90 / 51.58

Table Q3. ABIDE-II held-out sites (ACC / AUC, %; — denotes undefined AUC for single-class sites).

Held-out siteCCA-SRLNaive BayesHGNNBrainGNNBrainGBBrainOOD
BNI57.14 / 59.5746.43 / 45.7946.43 / 44.2646.43 / 51.7950.00 / 42.0939.29 / 37.88
EMC52.83 / 49.2947.17 / 48.0849.06 / 53.9952.83 / 57.9849.06 / 50.4349.06 / 49.00
GU54.29 / 52.9152.38 / 52.7349.52 / 57.4949.52 / 51.2052.38 / 55.4554.29 / 54.33
IP (only 3 samples)66.67 / 100.0033.33 / 0.0033.33 / 0.0066.67 / 50.0066.67 / 0.0066.67 / 0.00
IU65.00 / 68.0055.00 / 60.2557.50 / 53.2550.00 / 54.7552.50 / 61.5032.50 / 37.75
KKI31.75 / 51.3967.77 / 55.8838.39 / 52.3453.55 / 49.4169.19 / 55.8154.03 / 54.19
KUL28.57 / —42.86 / —14.29 / —28.57 / —0.00 / —25.00 / —
NYU44.76 / 51.8771.43 / 50.0037.14 / 53.8744.76 / 41.2928.57 / 46.4450.48 / 48.89
OHSU53.76 / 56.3248.39 / 49.9040.86 / 37.6943.01 / 37.5056.99 / 56.1844.09 / 36.00
ONRC67.31 / 59.0657.69 / 57.2761.54 / 55.1663.46 / 69.2244.23 / 62.3457.69 / 53.28
TCD55.56 / 75.0066.67 / 65.0044.44 / 45.0044.44 / 60.0055.56 / 75.0055.56 / 50.00
UCD50.00 / 42.8650.00 / 44.6443.75 / 36.5146.88 / 43.6543.75 / 50.0056.25 / 55.16
UCLA78.57 / 83.6778.57 / 83.6757.14 / 79.5971.43 / 79.5964.29 / 89.8057.14 / 85.71
UM46.43 / 52.8260.71 / 64.1032.14 / 33.8553.57 / 56.4153.57 / 60.5142.86 / 36.41
USM59.38 / 55.6940.62 / 40.2037.50 / 44.7143.75 / 47.4550.00 / 43.1434.38 / 36.86
Site-macro mean54.13 / 61.3254.60 / 51.2542.87 / 46.2650.59 / 53.5949.12 / 53.4847.95 / 45.39

Thank you for the detailed responses and additional analyses. The newly reported within-patient symptom association provides useful evidence beyond binary diagnostic discrimination, while the candidate-mining controls further demonstrate that the performance gains cannot be solely attributed to disease-informed candidate selection. I also appreciate the clarification that CCA-SRL is intended as a supervised diagnostic framework with context-conditioned anomaly representations, rather than a fully unsupervised anomaly detection method. This distinction makes the methodological scope of the proposed approach clearer.

Overall, the additional analyses and methodological clarification have substantially addressed my original concerns. The dependence on diagnostic labels remains an inherent limitation when extending the framework to unseen disorders or settings without labeled diagnostic data, but the authors have adequately clarified the intended scope of the proposed approach.

Reviewer_X4Q6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CCA-SRL models healthy subnetwork features conditioned on their functional context, then injects residual anomaly evidence into a signed graph classifier. The paper evaluates diagnosis on three datasets and contextual consistency through mismatching, perturbation, temporal, and cross-site analyses.

Critical strengths
  1. The contextual-consistency hypothesis is investigated extensively. Hard negatives, fixed contexts, within-subject swaps, temporal splits, and held-out-site analyses address several plausible alternative explanations (Appendices C-G).
  2. Matched candidate-mining controls and component ablations support an empirical contribution beyond supervised subnetwork selection alone (Appendix G.3; Tables 2-3).
Critical weaknesses
  1. The recovery guarantee is not established for the implemented method. Theorem 3 and Appendix A.7 compare similar-size discrete regions without handling the soft masks or size penalty in Equation (18). Moreover, actual candidates come from supervised edge mining, rather than this auxiliary TV optimization.
  2. Classifier calibration may reflect fitting artifacts. Algorithms 1-3 fit the normality model and empirical control tails using the same controls, risking a mismatch between in-sample and unseen-subject scores. Appendix B specifies cross-fitting for SCC validation, but its use in the final classifier is not established.
  3. Diagnostic robustness is less supported than contextual consistency. The leave-one-site-out experiments measure context preference, not diagnostic accuracy. The classification tables give one-sample tests rather than paired evidence of superiority over baselines, leaving the reliability of the diagnostic gains unresolved.
Questions for authors
  1. Are classifier training scores and control-tail calibration cross-fitted? If not, how do fully cross-fitted results compare?
  2. Can the authors provide paired diagnostic improvements on held-out sites and clarify which implemented procedure the recovery theorem guarantees?
8 comments

Weakness 1 / Question 2

We thank the reviewer for this constructive comment. We appreciate the opportunity to clarify the scope of Theorem 3 and its relationship to the implemented CCA-SRL pipeline. We will make the corresponding assumptions, theoretical guarantees, and methodological distinctions explicit in the manuscript.

First, regarding which procedure the recovery theorem guarantees: Theorem 3 / Appendix A.7 is a conditional guarantee for the residual-field recovery procedure, i.e., the TV-regularized region search over node-level context residuals. It characterizes the recovery of anomalous regions from contextual residuals under signal-separation, boundary-penalty, and noise conditions. In the implemented CCA-SRL pipeline, supervised candidate discovery is performed within the training folds using disorder–control connectivity differences. We will explicitly distinguish these complementary theoretical results and the supervised candidate-discovery process, highlighting that Theorem 3 provides a conditional recovery analysis for the auxiliary residual-field region search under the stated signal-separation, boundary-penalty, and noise assumptions.

Second, regarding soft masks and the size penalty in Eq. (18): The original proof sketch analyzes spatial recovery under residual-separation and noise assumptions. The soft-mask relaxation and size penalty in Eq. (18) allow continuous region membership and explicitly regularize region size, providing a flexible graph-regularized formulation for spatial recovery. We will clarify the scope of the theoretical analysis and its relationship to this broader formulation, distinguishing the recovery conditions characterized by the proof from the additional flexibility introduced by the relaxation. The theoretical analysis highlights that sufficiently separated contextual residuals, relative to graph-boundary penalties and noise, provide a principled basis for identifying spatially localized abnormalities. The controlled-perturbation experiments (Appendix F, Table 12) and diagnostic classification evaluations (Tables 1–2 and Appendix G.3) further support the practical utility of context-conditioned residual deviations for identifying abnormal subnetworks and guiding brain disease diagnosis.

The paired diagnostic improvements on held-out sites requested in Question 2 are addressed separately in our response to Weakness 3 / Question 2.

Weakness 2 / Question 1

In the original classifier, the normality model and empirical control-tail references are fitted using the same training healthy controls, so training anomaly scores and tail calibration are not fully cross-fitted.

To evaluate cross-fitted training scores and independent tail calibration, we designed a five-fold, three-role cross-fitting experiment and evaluated it on one dataset for each disorder: ADHD-200 for ADHD and ABIDE-II for ASD. Within each outer training fold, we apply five-fold, three-role cross-fitting. In each rotation, one group is scored, healthy controls from three groups fit the normality model, and healthy controls from the remaining group form the independent control-tail reference. Across five rotations, every training record receives out-of-fold anomaly features. We train the original classifier on these features and evaluate it under the outer four-fold protocol, keeping supervised candidate mining unchanged.

Table X1 reports four-fold results with out-of-fold anomaly features and independent healthy-control tail references: CCA-SRL retains diagnostic discrimination on both datasets.

Table X1. Classification performance with five-fold, three-role cross-fitting (four-fold mean, %).

DisorderDatasetACCAUC
ADHDADHD-20063.1562.65
ASDABIDE-II58.3161.94

Weakness 3 / Question 2

To address these points, we evaluated the full classification pipeline under leave-one-site-out (LOSO) splits and performed two-sided paired tt-tests on corresponding four-fold classification results.

For LOSO, CCA-SRL and the five baselines use the same random seed, training/validation/test partitions and held-out subjects within each dataset. Neither model fitting nor validation uses the held-out site. Tables X2–X3 report site-macro and individual-site ACC/AUC. ACC is averaged over all seven ADHD-200 and fifteen ABIDE-II sites; AUC is averaged over the six and fourteen sites, respectively, that contain both classes.

Tables X2 and X3 report LOSO results for individual sites and site-macro averages. CCA-SRL has the highest site-macro AUC among the compared methods on ADHD-200 (55.07%) and ABIDE-II (61.32%). On ABIDE-II, Naive Bayes has a slightly higher site-macro ACC (54.60% vs. 54.13%), while CCA-SRL achieves a higher site-macro AUC (61.32% vs. 51.25%).

Table X4 reports two-sided paired tt-tests on corresponding four-fold ACC and AUC differences (n=4n=4). The results show statistically significant differences for multiple baseline comparisons across the three datasets.

Table X2. ADHD-200 held-out sites (ACC / AUC, %; — denotes undefined AUC for single-class sites).

Held-out siteCCA-SRLNaive BayesHGNNBrainGNNBrainGBBrainOOD
KKI71.28 / 55.5460.64 / 46.7273.40 / 44.0671.28 / 44.7559.57 / 44.1772.34 / 55.77
NYU45.53 / 55.5142.02 / 46.8542.80 / 49.2443.19 / 56.0942.02 / 47.7642.80 / 49.99
NeuroIMAGE55.56 / 61.5050.00 / 43.7550.00 / 50.0851.39 / 50.6952.78 / 51.7044.44 / 40.28
OHSU61.06 / 47.9747.79 / 28.2155.75 / 34.5561.95 / 52.6661.06 / 43.3958.41 / 36.28
Peking57.96 / 51.6051.02 / 44.3251.02 / 42.7751.84 / 40.3957.55 / 44.5359.18 / 51.64
Pittsburgh83.51 / 58.3358.76 / 72.1883.51 / 50.0061.86 / 79.8495.88 / 40.3292.78 / 75.54
WashU98.36 / —72.13 / —93.44 / —83.61 / —88.52 / —98.36 / —
Site-macro mean67.61 / 55.0754.62 / 47.0064.28 / 45.1260.73 / 54.0765.34 / 45.3166.90 / 51.58

Table X3. ABIDE-II held-out sites (ACC / AUC, %; — denotes undefined AUC for single-class sites).

Held-out siteCCA-SRLNaive BayesHGNNBrainGNNBrainGBBrainOOD
BNI57.14 / 59.5746.43 / 45.7946.43 / 44.2646.43 / 51.7950.00 / 42.0939.29 / 37.88
EMC52.83 / 49.2947.17 / 48.0849.06 / 53.9952.83 / 57.9849.06 / 50.4349.06 / 49.00
GU54.29 / 52.9152.38 / 52.7349.52 / 57.4949.52 / 51.2052.38 / 55.4554.29 / 54.33
IP (only 3 samples)66.67 / 100.0033.33 / 0.0033.33 / 0.0066.67 / 50.0066.67 / 0.0066.67 / 0.00
IU65.00 / 68.0055.00 / 60.2557.50 / 53.2550.00 / 54.7552.50 / 61.5032.50 / 37.75
KKI31.75 / 51.3967.77 / 55.8838.39 / 52.3453.55 / 49.4169.19 / 55.8154.03 / 54.19
KUL28.57 / —42.86 / —14.29 / —28.57 / —0.00 / —25.00 / —
NYU44.76 / 51.8771.43 / 50.0037.14 / 53.8744.76 / 41.2928.57 / 46.4450.48 / 48.89
OHSU53.76 / 56.3248.39 / 49.9040.86 / 37.6943.01 / 37.5056.99 / 56.1844.09 / 36.00
ONRC67.31 / 59.0657.69 / 57.2761.54 / 55.1663.46 / 69.2244.23 / 62.3457.69 / 53.28
TCD55.56 / 75.0066.67 / 65.0044.44 / 45.0044.44 / 60.0055.56 / 75.0055.56 / 50.00
UCD50.00 / 42.8650.00 / 44.6443.75 / 36.5146.88 / 43.6543.75 / 50.0056.25 / 55.16
UCLA78.57 / 83.6778.57 / 83.6757.14 / 79.5971.43 / 79.5964.29 / 89.8057.14 / 85.71
UM46.43 / 52.8260.71 / 64.1032.14 / 33.8553.57 / 56.4153.57 / 60.5142.86 / 36.41
USM59.38 / 55.6940.62 / 40.2037.50 / 44.7143.75 / 47.4550.00 / 43.1434.38 / 36.86
Site-macro mean54.13 / 61.3254.60 / 51.2542.87 / 46.2650.59 / 53.5949.12 / 53.4847.95 / 45.39

Comments on the paper

2 comments

Thank you to the reviewer who has already followed up on the authors' responses. For reviewers who have not yet done so, please post a brief follow-up acknowledging the authors' responses within the next 24 hours, so the authors have time to respond before discussion closes. In your reply, please indicate which main concerns have been addressed, what remains, and whether your overall assessment changes or stays the same.

Summary of the Rebuttal

We sincerely thank all three reviewers for their careful evaluation of our work and constructive feedback. We greatly appreciate the opportunity to clarify our contributions and strengthen the manuscript through the rebuttal process.

During the rebuttal, we carefully addressed the concerns raised by all three reviewers regarding theoretical guarantees, calibration reliability, diagnostic robustness, clinical relevance, label dependence, and comparisons with existing methods. We provided targeted theoretical and methodological clarifications, together with additional experiments and statistical analyses to strengthen the evidence supporting CCA-SRL.

We clarified the scope and assumptions of Theorem 3, distinguishing the auxiliary residual-field recovery analysis from the implemented supervised candidate-discovery procedure, and explained the relationship between the theoretical analysis and the soft-mask formulation. We also clarified the supervised nature of candidate discovery, the role of context-conditioned anomaly representations, and candidate-specific calibration, while further explaining the methodological motivation for modeling healthy context–subnetwork relationships.

We conducted five-fold, three-role cross-fitting experiments to examine calibration reliability, full-classifier LOSO evaluations on ADHD-200 and ABIDE-II to assess cross-site generalization, and paired statistical tests to evaluate diagnostic improvements. We further introduced two independent external test cohorts, OpenNeuro-ds000030 (162 subjects) and ABIDE-I (406 subjects), with CCA-SRL achieving the highest AUC among the compared methods on both datasets. To strengthen the clinical interpretation of identified subnetworks, we supplemented patient–control comparisons with within-patient symptom associations and supporting neuroimaging literature. Finally, we added comparisons with ContrastPool, ALTER, and Geo-Mamba across all three original datasets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.