acceptodds
Under review as a conference paper at ICLR 2027

Learning beyond Site Bias for OOD Generalization in Brain Networks

Abstract

Graph-based learning from functional magnetic resonance imaging (fMRI) has shown strong potential for brain network analysis. However, existing methods often degrade under cross-site out-of-distribution (OOD) settings, as site-dependent confounder effects can obscure disease-related connectivity patterns and static functional connectivity (FC) does not explicitly capture informative within-scan variations. In this paper, we propose Cross-site OOD Robust brain nEtwork (CORE), a unified framework for brain network learning across unseen sites. First, CORE estimates site-specific confounder effects and aggregates the resulting deconfounders by cross-source reliability for unseen-site correction, while extracting a population scaffold of reproducible label-associated connections from site-wise residualized FC. It then summarizes temporal variations on scaffold edges into compact descriptors for line-graph modeling. Finally, prior-guided subject-adaptive gating modulates message passing, balancing population priors with individual variability. Extensive leave-one-site-out experiments on ABIDE, REST-meta-MDD, SRPBS, and ABCD demonstrate that CORE consistently outperforms competitive baselines, with up to a 10.3% relative improvement in accuracy. These gains also persist across different brain parcellation schemes on ABIDE.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Functional Brain Networks, Functional Connectivity, Cross-site Out-of-Distribution Generalization

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

This paper proposes CORE to combine site-specific deconfounding, cross-site selection of label-associated connections, dynamic FC summaries, and gated line-graph learning.

Critical strengths
  • LOSO evaluation across four datasets directly addresses cross-site distribution shifts.
  • The performance is good.
Critical weaknesses
  • The scaffold’s contribution is insufficiently isolated. The ablations do not separately test supervised scaffold selection, sign consistency, bootstrap stability, and reliability weighting. Add simple classifiers on the same scaffold and a parameter-matched nongraph model to establish the line graph’s added value.
  • Atlas robustness should be stated narrowly. Separate training and evaluation on AAL and CC200 show that the method works with either parcellation. They do not establish transfer to an unseen atlas.
Questions for authors
  • How sensitive is aggregate correction to covariate scaling and the pseudoinverse convention?
  • Existing work explicitly explores demographic-enhanced brain network models [1-3]. Why these works are not included for discussion and comparison?

[1] Classification of brain disorders in rs-fmri via local-to-global graph neural networks. TMI 2022 [2] Knowledge-driven multigraph convolutional network for brain network analysis and potential biomarker discovery. MIA 2025 [3] BrainPrompt+: Multi-Level Brain Prompt Learning for Knowledge-Guided Neurological Disorder Identification. TMI 2026

9 comments
AuthorsOpenReview ↗

Response to Reviewer fiGY (1/6)

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following:

W1: The scaffold's contribution is insufficiently isolated. The ablations do not separately test supervised scaffold selection, sign consistency, bootstrap stability, and reliability weighting. Add simple classifiers on the same scaffold and a parameter-matched nongraph model to establish the line graph's added value.

A1: Thanks for your comment. The reliability weights do not enter scaffold selection. The scaffold is extracted from the site-wise residuals of the training sites, and the weights αe\alpha_e enter only the correction of validation and test subjects in Eq. 8, so we evaluate them separately from the selection criteria.

To isolate each part of the scaffold and the added value of the line graph, we replaced one part of CORE at a time and kept all other components fixed. The variants are constructed as follows:

  • For scaffold selection, we replaced the scaffold with a random edge set, with the edges of highest variance of the site-wise residuals, which uses no labels, with edges selected by a single diagnostic contrast on the pooled source subjects, which uses no cross-site consensus, and with the top consensus contrasts ∣dcom,j∣\vert d_{\mathrm{com},j}\vert without the sign and bootstrap criteria. These four scaffolds have the size of the CORE scaffold in the same fold. We also removed sign consistency, ηE=0\eta_E=0, and both criteria, ηE=ζE=0\eta_E=\zeta_E=0.
  • For the unseen-site correction, we replaced reliability-weighted aggregation with uniform aggregation.
  • On the CORE scaffold and node features, we trained an ℓ2\ell_2-regularized logistic regression, a linear SVM, a random forest, an MLP on the concatenated node features, a Transformer whose tokens are the scaffold pathways, a GINE on the ROI graph restricted to the scaffold edges, and CORE without line-graph message passing, w/o LG. On ABIDE, the MLP, the Transformer and the GINE have 33,814, 34,106 and 33,682 trainable parameters on average over folds, against 33,923 for CORE.

We used ABIDE and REST-meta-MDD with three seeds. Averaged over folds, the scaffold has 412 edges on ABIDE and 486 on REST-meta-MDD, 987 and 1,020 without sign consistency, and 1,334 on both datasets without both criteria. Since τE\tau_E is fixed at the 80th percentile of the absolute consensus contrasts, removing both criteria keeps the top 20% of the 6,670 AAL-116 edges.

Removing bootstrap stability alone leaves the scaffold unchanged at the default thresholds. With seven training sites and ηE=0.75\eta_E=0.75, an edge that passes sign consistency has at most one source site of the opposite sign. The median of seven resampled sites reverses its sign only if that site is drawn at least four times, which gives πj≈0.99\pi_j\approx0.99, far above ζE\zeta_E. Bootstrap stability acts only when sign consistency is removed. Edges whose consensus sign is carried by four of the seven sites then have πj≈0.65\pi_j\approx0.65, and with B=50B=50 about 30% of them pass ζE=0.70\zeta_E=0.70.

The table below reports the LOSO AUC and ACC in %. For each seed, we average the metric over all target sites of a dataset, and each entry is the mean ± SD of the three seed-level values.

AuthorsOpenReview ↗

Response to Reviewer fiGY (2/6)

VariantABIDE AUC↑ABIDE ACC↑REST-meta-MDD AUC↑REST-meta-MDD ACC↑
Scaffold selection
Random scaffold58.17±0.4158.31±0.7758.42±0.6857.95±0.83
Unsupervised, highest-variance edges60.06±0.4259.42±0.8060.37±0.4059.88±0.56
Single contrast on pooled source subjects62.94±0.4562.41±0.5663.95±0.6560.91±0.75
w/o sign consistency, ηE=0\eta_E=063.81±0.4062.19±1.0764.12±0.5162.65±0.49
w/o both, ηE=ζE=0\eta_E=\zeta_E=062.71±0.3361.38±0.4663.30±0.6261.67±0.73
Top consensus contrasts, same size64.45±0.4863.21±0.9964.38±0.6662.97±0.69
Unseen-site correction
Uniform aggregation65.13±0.8364.80±1.2765.52±0.6263.62±0.70
Classifier on the CORE scaffold
Logistic regression, ℓ2\ell_260.92±0.4360.95±0.9262.20±0.6160.41±0.46
Linear SVM61.05±0.7860.31±0.8262.49±0.3960.18±0.81
Random forest59.83±0.8860.12±1.4262.37±0.4559.08±0.97
MLP, parameter-matched62.24±0.4761.73±1.1463.31±0.5260.36±0.60
Transformer, parameter-matched63.02±0.3562.84±0.6963.18±0.3261.79±0.52
ROI-graph GINE, parameter-matched63.58±0.4062.54±0.5964.11±0.4961.40±0.87
CORE w/o LG64.00±0.6663.48±0.7463.96±0.5961.70±0.57
Full model
CORE65.21±0.7864.95±1.1165.58±0.6263.74±0.79

As shown in the above table, supervised selection and cross-site consensus are both needed. A random scaffold falls 5.79 to 7.16 points below CORE, and an unsupervised scaffold of the same size 3.86 to 5.53 points below. Selecting edges by one contrast on the pooled source subjects lowers AUC by 1.63 to 2.27 points and ACC by 2.54 to 2.83 points. Taking the top consensus contrasts of the same size without the sign and bootstrap criteria still falls 0.76 to 1.74 points below CORE, which shows that the criteria select better edges than the contrast magnitude alone. Removing sign consistency more than doubles the scaffold and lowers accuracy by 1.09 to 2.76 points. Without sign consistency, bootstrap stability still removes most edges whose sign is carried by only four of the seven sites, and removing it as well lowers accuracy by a further 0.81 to 1.10 points. Uniform aggregation stays within the variation across seeds of reliability weighting, consistent with its role in the correction only.

AuthorsOpenReview ↗

Response to Reviewer fiGY (3/6)

On the same scaffold and features, the linear classifiers, the random forest and the parameter-matched MLP fall 2.27 to 5.38 points below CORE. The models that capture interactions between pathways perform better, and the Transformer, the GINE and CORE w/o LG lie within about 1 point of each other. Line-graph message passing adds 1.21 to 1.95 points over the best of these three models, which isolates the contribution of the line graph beyond graph or attention models with the same parameter budget.

The scaffold is also reproducible. Over ten runs per fold with different seeds and validation sites, the selected scaffolds have a mean Jaccard index of 0.84 on ABIDE and 0.87 on REST-meta-MDD.

W2: Atlas robustness should be stated narrowly. Separate training and evaluation on AAL and CC200 show that the method works with either parcellation. They do not establish transfer to an unseen atlas.

A2: Thanks for your comment. We agree with this point. CORE and all baselines were trained and evaluated separately with each parcellation. This result shows that the gains hold when the method is applied with either atlas, and it does not show that a trained model transfers to an unseen atlas. We have revised the Abstract, the Introduction, Sec. 4.2 and the Conclusion to state this scope. For example, the Abstract now ends with "CORE can also be trained and evaluated with a second parcellation of ABIDE," and Sec. 4.2 states that each method is trained and evaluated separately with each parcellation, so the result does not address the transfer of a trained model to an unseen atlas.

Q1: How sensitive is aggregate correction to covariate scaling and the pseudoinverse convention?

A3: Thanks for your question. The aggregate correction depends on the covariate scale and the pseudoinverse convention, and this dependence lies mostly in a shift shared by all subjects of a target site. Rescaling the site-level covariates by 10 moves the correction by 0.41 to 0.79 SD of the edge and lowers ACC by 1.49 to 2.41 points, while AUC changes by at most 1 point across all rescalings, and an identified implementation stays within the variation across seeds of the submitted deconfounder.

Within a source site, TR and sequence length take a single value, so their columns in Eq. 5 are multiples of the intercept column. The source data identify the total site offset oe,jo_{e,j} and the within-site slopes, but not how oe,jo_{e,j} is assigned between the intercept and the site-level covariates. The same holds for sex at the two all-male ABIDE sites, SBL and TRINITY. The site-wise residuals in Eq. 6 and the scaffold do not depend on this assignment, while the reliability weights in Eq. 7 and the unseen-site correction in Eq. 8 do. With standardized site-level covariates ae\boldsymbol a_e, the pseudoinverse selects the minimum-norm assignment, under which site ee contributes oe,j(1+ae⊤a∗)/(1+∥ae∥2)o_{e,j}(1+\boldsymbol a_e^{\top}\boldsymbol a_{\ast})/(1+\Vert\boldsymbol a_e\Vert^2) to the correction at a target site with covariates a∗\boldsymbol a_{\ast}. Rescaling ae\boldsymbol a_e changes this assignment, while rescaling a covariate that varies within a site only reparameterizes its slope and leaves the correction unchanged, so the sensitivity comes only from the covariates that are constant within a site.

To characterize where this dependence acts, we have added Proposition 3 to Sec. 3.1. For any aggregate deconfounder, the intercept and the offset of a target site cancel in its diagnostic contrast, and the assignment enters the contrast only through the aggregated slopes of covariates that vary within the target site. It reaches these slopes through the reliability weights and, on ABIDE, through the sex coefficients of SBL and TRINITY. The part of the correction shared by all subjects of the target site moves the operating point of the classifier, so ACC is expected to respond more strongly than AUC.

Reviewer_WV4dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CORE, a unified framework for out-of-distribution generalization in brain network classification across unseen imaging sites. The framework consists of three components: (1) site-aware confounder decoupling using site-specific Huber regression and reliability-weighted aggregation, followed by the extraction of a population-level connectivity scaffold; (2) transient pathway profiling that summarizes dynamic functional connectivity using compact temporal descriptors; and (3) prior-guided subject-adaptive gating for line-graph message passing and prediction. The authors provide two theoretical propositions concerning deconfounding bias and scaffold stability. Experiments on ABIDE, REST-meta-MDD, SRPBS, and ABCD demonstrate improved classification performance over multiple baselines under leave-one-site-out evaluation.

Critical strengths
  1. Relevant and practically important problem. The paper addresses cross-site generalization in brain network analysis, where acquisition heterogeneity and demographic differences present substantial challenges.
  2. Comprehensive empirical evaluation. The authors compare CORE against a broad range of GNN, graph OOD, and brain-network-specific methods across four datasets and an additional brain parcellation setting. The reported results are consistently competitive.
  3. Integrated framework with supporting analyses. The combination of confounder adjustment, compact temporal representations, and adaptive pathway modeling is practically motivated. The additional ablation, sensitivity, and interpretability analyses help characterize the contributions of the proposed components.
Critical weaknesses
  1. Identifiability concerns in site-aware deconfounding. The site-specific regression includes acquisition covariates such as TR and sequence length, which are collinear with the intercept within each site. Although Appendix F adopts a pseudoinverse solution, the resulting coefficients are not uniquely identifiable for unseen acquisition settings. This raises concerns about the validity of reliability-weighted deconfounding on unseen sites.
  2. Insufficient validation of the proposed deconfounding mechanism. The theoretical propositions establish residual-bias and scaffold-stability properties under specific assumptions but provide no guarantee for unseen-site correction. Although the ablation study compares site-aware and global deconfounding, direct evidence of reduced site dependence and an isolated evaluation of reliability-weighted aggregation are still missing.
  3. Limited experimental reliability and reproducibility. The main LOSO experiments use a single run per fold, without paired significance tests to establish the reliability of performance improvements. Furthermore, the source code is not publicly available, limiting independent reproduction of the reported results. The SRPBS evaluation also treats different acquisition batches from the same institution as separate environments, weakening the evidence for generalization to entirely unseen institutions.
  4. Limited methodological novelty and temporal comparability. CORE primarily integrates established techniques, including robust regression, consensus-based feature selection, temporal statistics, and adaptive graph learning. The methodological novelty beyond their integration remains limited. Additionally, fixed-length sliding windows measured in time points may introduce inconsistencies across sites with different TRs, potentially affecting temporal feature comparability.
Questions for authors
  1. How do the authors justify transferring site-specific regression coefficients for TR and sequence length when these covariates are constant within individual sites? How does the pseudoinverse solution affect unseen-site extrapolation?
  2. Can the authors provide direct evidence that CORE reduces residual site dependence, together with comparisons against uniform source aggregation and simpler deconfounding strategies?
  3. Can the authors provide multi-seed results with statistical significance tests and release the source code to facilitate reproducibility? Additionally, can the authors clarify the impact of acquisition-batch splitting on the SRPBS evaluation?
  4. How does CORE account for differences in physical window duration and the number of windows across acquisition protocols? Have the authors evaluated TR-normalized windows or matched scan durations?
15 comments
AuthorsOpenReview ↗

Response to Reviewer WV4d (1/12)

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following:

W1: Identifiability concerns in site-aware deconfounding. The site-specific regression includes acquisition covariates such as TR and sequence length, which are collinear with the intercept within each site. Although Appendix F adopts a pseudoinverse solution, the resulting coefficients are not uniquely identifiable for unseen acquisition settings. This raises concerns about the validity of reliability-weighted deconfounding on unseen sites.

A1: Thanks for your comment. Within a source site, TR and sequence length take a single value, so their columns in Eq. 5 are multiples of the intercept column. The source data identify the total site offset oe,jo_{e,j} and the within-site slopes, but not how oe,jo_{e,j} is assigned between the intercept and the site-level covariates. The same holds for sex at the two all-male ABIDE sites, SBL and TRINITY. The site-wise residuals in Eq. 6 and the scaffold do not depend on this assignment. The reliability weights in Eq. 7 and the unseen-site correction in Eq. 8 do, since both evaluate the deconfounding map of a site at the covariates of other sites, so the validity of reliability-weighted deconfounding at unseen sites depends on how the assignment affects both. With standardized site-level covariates ae\boldsymbol a_e, the pseudoinverse selects the minimum-norm assignment, under which site ee contributes oe,j(1+ae⊤a∗)/(1+∥ae∥2)o_{e,j}(1+\boldsymbol a_e^{\top}\boldsymbol a_{\ast})/(1+\Vert\boldsymbol a_e\Vert^2) to the correction at a target site with covariates a∗\boldsymbol a_{\ast}. Assigning the whole offset to the intercept fits the source data equally well and contributes oe,jo_{e,j}. Propositions 1 and 2 concern the source sites and do not cover this extrapolation.

To characterize the effect of the assignment at an unseen site, we have added Proposition 3 to Sec. 3.1. At an unseen site tt that follows the additive model of Proposition 1, any fixed aggregate intercept bˉ\bar{\boldsymbol b} and coefficients Γˉ\bar{\boldsymbol\Gamma} give a corrected FC rires=ri−bˉ−Γˉ⊤q~i\boldsymbol r_i^{\mathrm{res}}=\boldsymbol r_i-\bar{\boldsymbol b}-\bar{\boldsymbol\Gamma}^{\top}\tilde{\boldsymbol q}_i whose diagnostic contrast satisfies

Dt[rires]=Dt[si⋆]+(Γt⋆−Γˉ)⊤Dt[q~i],\mathrm{D}_t\big[\boldsymbol r^{\mathrm{res}}_i\big]=\mathrm{D}_t[\boldsymbol s^\star_i]+(\boldsymbol\Gamma^\star_t-\bar{\boldsymbol\Gamma})^{\top}\mathrm{D}_t[\tilde{\boldsymbol q}_i],

where the entries of Dt[q~i]\mathrm{D}_t[\tilde{\boldsymbol q}_i] for covariates that are constant within site tt are zero. The result holds for any aggregate deconfounder, so it covers every choice of reliability weights and every assignment of the source offsets. The intercept and the offset of the target site cancel in its diagnostic contrast, and the assignment enters only through the rows of Γˉw\bar{\boldsymbol\Gamma}_w for covariates that vary within the target site. It reaches these rows through the reliability weights and, on ABIDE, through the sex coefficients of SBL and TRINITY. The remaining term vanishes when the diagnostic groups at the target site are matched on these covariates. The part of the correction shared by all subjects of the target site acts on the operating point of the classifier. For a linear readout, it moves all scores of the site by the same amount, which changes ACC and leaves AUC unchanged. For the nonlinear classifier of CORE, this relation holds approximately, so ACC is expected to respond more strongly than AUC.

To test how strongly the reliability-weighted correction depends on the assignment and whether this dependence affects the results, we compared the submitted deconfounder with four variants. The variants are constructed as follows:

  • Default is the submitted deconfounder, which fits each source site with standardized covariates and the pseudoinverse.
  • Default ×0.1 and Default ×10 rescale the site-level covariates by 0.1 and 10 before the same fit. A smaller scale moves the minimum-norm assignment toward the intercept, and a larger scale moves it toward the site-level covariates.
  • Omitted removes the covariates that are constant within a site from the fit of that site. Each within-site slope is aggregated, and enters ueu_e, only over the sites where it is estimated, with renormalized weights. Every coefficient in Eqs. 7 and 8 is identified under this variant, including the reliability weights.
  • Joint estimates the acquisition slopes across source sites, where TR and sequence length vary, by regressing the source intercepts on the site-level covariates with the Huber loss, and applies its prediction at the unseen site. This variant is identified when the site offset is linear in the acquisition parameters, and each fold provides seven training sites for its estimation.
AuthorsOpenReview ↗

Response to Reviewer WV4d (2/12)

All variants yield the same site-wise residuals, scaffold and training features, and differ only in the correction of validation and test subjects. We ran CORE with each variant, BrainOOD and DeCI on ABIDE, REST-meta-MDD and SRPBS with three seeds and the same validation sites. ABCD is excluded, as TR is 0.8 s and the sequence length is identical at all its sites, so all variants coincide.

The table below reports the LOSO AUC and ACC in %. For each seed, we average the metric over all target sites of a dataset, and each entry is the mean ± SD of the three seed-level values. ΔC\Delta_C is the mean absolute deviation of the unseen-site correction of a variant from Default, in units of the across-subject SD of the edge at the target site, averaged over edges, held-out subjects and folds. Since a variant changes both the reliability weights and the aggregated coefficients, ΔC\Delta_C measures their combined effect on the correction.

MethodABIDE AUC↑ABIDE ACC↑ABIDE ΔC\Delta_C↓REST-meta-MDD AUC↑REST-meta-MDD ACC↑REST-meta-MDD ΔC\Delta_C↓SRPBS AUC↑SRPBS ACC↑SRPBS ΔC\Delta_C↓
BrainOOD63.68±0.2260.28±1.13–62.56±0.7458.84±0.59–90.15±0.3677.40±0.41–
DeCI55.90±0.3160.84±0.38–55.06±0.7256.22±0.16–91.58±1.0685.03±0.25–
CORE, Default65.21±0.7864.95±1.11065.58±0.6263.74±0.79092.36±0.5786.77±0.630
CORE, Default ×0.165.09±0.7464.77±0.980.2165.51±0.6363.44±0.560.1392.22±0.5186.26±0.720.30
CORE, Default ×1064.62±0.8963.17±1.120.6465.13±0.4562.25±0.590.4191.55±0.4884.36±0.790.79
CORE, Omitted64.99±0.6664.61±1.250.2365.43±0.6563.36±0.750.1392.13±0.5786.47±0.520.31
CORE, Joint64.66±0.7363.80±1.090.3765.30±0.8263.22±1.030.2591.36±0.4884.33±0.760.49

As shown in the above table, the deviation of the correction grows with the scale of the site-level covariates. ΔC\Delta_C ranges from 0.13 to 0.30 under ×0.1 and from 0.41 to 0.79 under ×10, which confirms that the pseudoinverse correction depends on the covariate scale. The accuracy follows the pattern predicted by Proposition 3. Under ×10, ACC decreases by 1.49 to 2.41 points across the three datasets, while AUC decreases by only 0.45 to 0.81 points, consistent with a shift of the operating point. Under ×0.1, both metrics change by at most 0.51 points. Omitted, under which the reliability weights are identified, shifts the correction by a similar amount as ×0.1, with ΔC\Delta_C between 0.13 and 0.31, and stays within 0.23 points of Default in AUC and within 0.38 points in ACC on all three datasets. These differences are smaller than the SD over three seeds in every column. Joint falls below Omitted, with the largest gap on SRPBS, where ACC drops from 86.47 to 84.33. The training sites of SRPBS cover only three sequence lengths, which limits the estimation of acquisition slopes across sites. Compared with the baselines, Default, ×0.1 and Omitted exceed BrainOOD and DeCI in every column. Under ×10 and Joint, CORE remains above BrainOOD in every column and falls slightly below DeCI on SRPBS, by 0.03 and 0.22 points in AUC and by 0.67 and 0.70 points in ACC.

AuthorsOpenReview ↗

Response to Reviewer WV4d (3/12)

To evaluate the reliability weights directly, we decomposed the deviation of each variant from Default into the part shared by all subjects of a target site and the part that varies across subjects, and compared the reliability weights of each variant with those of Default by their total variation distance and their effective number of sites 1/∑eαe21/\sum_e\alpha_e^2. The decomposition shows that the shared part dominates the deviation. It reaches 0.76 SD of the edge, while the part that varies across subjects stays below 0.11 SD, which agrees with the small changes in AUC in the above table. The comparison of the weights shows that they account for the larger cost of ×10. Under Default, the effective number of sites lies between 6.6 and 6.7 of the seven training sites, so the weights are close to uniform. Scaling by 10 moves the weights by a total variation distance of 0.16 to 0.26 and lowers the effective number of sites to between 4.6 and 5.4, which concentrates the aggregated slopes on fewer sites. Omitted, under which the weights are computed only from identified coefficients, moves them by only 0.05 to 0.07.

To quantify the remaining term in Proposition 3, we computed Rt,j=(γ^t,j−γˉw,j)⊤D^t[q~i]R_{t,j}=(\hat{\boldsymbol\gamma}_{t,j}-\bar{\boldsymbol\gamma}_{w,j})^{\top}\hat{\mathrm{D}}_t[\tilde{\boldsymbol q}_i] of the submitted deconfounder at each target site. The target slopes γ^t,j\hat{\boldsymbol\gamma}_{t,j} are fitted post hoc on the target site with its labels and are never used at inference. Estimation noise in these slopes tends to inflate ∣Rt,j∣\vert R_{t,j}\vert, so the reported values are conservative. We also measured the standardized mean difference, SMD, of age, sex and mean framewise displacement, FD, between the diagnostic groups within each site. The table below reports the median ∣Rt,j∣\vert R_{t,j}\vert in units of the across-subject SD of the edge, the median ratio ρt\rho_t of ∣Rt,j∣\vert R_{t,j}\vert to the observed contrast, the share of edges whose contrast changes sign when Rt,jR_{t,j} is removed, and the same share among edges whose contrast exceeds its standard error, averaged over folds, on scaffold edges and on all edges.

DatasetScaffold edges ∣R∣\vert R\vert↓Scaffold edges ρt\rho_t↓Scaffold edges Flip %↓Scaffold edges Flip>SE %↓All edges ρt\rho_t↓All edges Flip %↓All edges Flip>SE %↓
ABIDE0.0350.123.90.000.175.60.07
REST-meta-MDD0.0270.113.30.050.227.20.37
SRPBS0.0200.040.60.000.134.30.06

As shown in the above table, the remaining term is small on all three datasets, with a median of 0.020 to 0.035 SD of the edge on scaffold edges. It is also smaller relative to the contrast on scaffold edges than on all edges, with ρt\rho_t between 0.04 and 0.12 against 0.13 to 0.22. Removing the remaining term reverses the contrast of 0.6% to 3.9% of the scaffold edges, compared with 4.3% to 7.2% of all edges, and of at most 0.05% of the scaffold edges whose contrast exceeds its standard error. SRPBS shows the smallest remaining term and the fewest sign changes. The small size of the term is consistent with the covariate balance within sites, where the diagnostic groups differ little in age and sex, with a mean ∣SMD∣\vert\mathrm{SMD}\vert between 0.11 and 0.21, and mean FD on ABIDE shows the largest difference, with a mean ∣SMD∣\vert\mathrm{SMD}\vert of 0.31. These results indicate that consensus selection over site-wise residuals retains edges with large and reproducible contrasts, which the remaining term rarely reverses. Reliability-weighted deconfounding gives the same results whether the weights are computed under the pseudoinverse assignment or from identified coefficients, and the gains over the baselines hold under both. The comparison between reliability-weighted and uniform aggregation is reported in our response to W2.

Reviewer_d6U7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CORE, a framework for cross-site out-of-distribution generalization in brain network analysis. The method integrates site-aware confounder decoupling, a population-level scaffold of reproducible diagnostic connections, temporal connectivity descriptors, and prior-guided subject-adaptive gating on a line graph. Experiments are conducted on ABIDE, REST-meta-MDD, SRPBS, and ABCD, with additional evaluation using an alternative ABIDE parcellation.

Critical strengths

1.The paper addresses an important problem in cross-site neuroimaging generalization. The framework considers site-dependent covariate effects, temporal connectivity variations, and subject-specific pathway importance within a unified architecture.

2.The experimental evaluation is relatively comprehensive, covering four datasets, multiple baseline categories, component ablations, temporal representation comparisons, and an alternative brain parcellation.

3.The authors have made substantial efforts to address the methodological concerns raised during the review. In particular, the additional experiments on identifiable deconfounding variants, acquisition-matched temporal descriptors, and repeated LOSO evaluation provide useful evidence and improve the reliability of the experimental conclusions.

Critical weaknesses

1.The methodological novelty and the necessity of individual components remain insufficiently established. CORE combines several existing ideas, including site-specific regression, consensus feature selection, dynamic functional connectivity descriptors, line-graph modeling, and adaptive gating. Although the ablation studies demonstrate the contribution of individual components, the paper could better clarify the methodological distinction from existing brain-network OOD approaches and establish whether the proposed components provide genuinely complementary benefits rather than simply combining established techniques.

2.The contribution of temporal pathway profiling remains dataset-dependent. The additional experiments with fixed-duration windows, percentile-based ranges, and truncated scans are helpful. However, the reported gains are not consistent across all datasets. In particular, the temporal descriptors provide limited improvements on REST-meta-MDD, where the confidence intervals of the paired gains include zero. Therefore, the general effectiveness of temporal profiling across different acquisition settings remains somewhat uncertain.

3.The statistical evidence for consistent superiority over competing methods is still limited. The additional repeated evaluation across three seeds and three validation-site selections is an important improvement. Nevertheless, the reported confidence intervals exclude zero in only a subset of the comparisons, and only two comparisons remain significant after Holm correction. Thus, while CORE demonstrates positive average gains in most settings, the evidence does not fully establish statistically significant superiority across all datasets and evaluation metrics. The conclusions should be calibrated accordingly.

Questions for authors

1.Can the authors further clarify the methodological novelty of CORE compared with existing brain-network OOD methods? In particular, what specific interactions between site-aware deconfounding, temporal pathway profiling, and subject-adaptive gating cannot be achieved by combining existing components independently?

2.Given that the temporal profiling module provides inconsistent improvements across datasets, could the authors discuss the conditions under which its contribution is most reliable and whether its effectiveness depends on acquisition characteristics or dataset properties?

3.Given that most repeated paired comparisons do not reach statistical significance after multiple-testing correction, could the authors moderate the claims regarding the consistent superiority of CORE over existing methods?

The paper addresses a meaningful problem and presents a reasonably comprehensive framework with encouraging experimental results. The authors have also responded constructively to the major methodological concerns through substantial additional analyses. These efforts have significantly improved the soundness of the work.

However, some concerns remain regarding the methodological novelty, the dataset-dependent contribution of temporal profiling, and the statistical strength of the claimed performance improvements. Overall, I consider the work technically promising, but its current evidence does not fully establish the broad superiority and methodological significance claimed in the paper.

10 comments
AuthorsOpenReview ↗

Response to Reviewer d6U7 (1/8)

We are truly grateful for the time you have taken to review our paper and your insightful review. Here we address your comments in the following:

W1: Acquisition-effect transfer is not identified by the source regressions. Site-constant covariates are collinear with the intercept (Appendix F). Pseudoinverse fitting selects coefficients, but equivalent within-site fits yield different cross-site weights and corrections in Equations 7–8. The oracle contrast-preservation result does not justify this extrapolation.

A1: Thanks for your comment. Within a source site, TR and sequence length take a single value, so their columns in Eq. 5 are multiples of the intercept column. The source data identify the total site offset oe,jo_{e,j} and the within-site slopes, but not how oe,jo_{e,j} is assigned between the intercept and the site-level covariates. The same holds for sex at the two all-male ABIDE sites, SBL and TRINITY. The site-wise residuals in Eq. 6 and the scaffold do not depend on this assignment, while the reliability weights in Eq. 7 and the unseen-site correction in Eq. 8 do, since both evaluate the deconfounding map of a site at the covariates of other sites. With standardized site-level covariates ae\boldsymbol a_e, the pseudoinverse selects the minimum-norm assignment, under which site ee contributes oe,j(1+ae⊤a∗)(1+∥ae∥2)o_{e,j} (1+\boldsymbol a_e^ {\top} \boldsymbol a_*) (1+\Vert\boldsymbol a_e\Vert^2) to the correction at a target site with covariates a∗\boldsymbol a_*. Assigning the whole offset to the intercept fits the source data equally well and contributes oe,jo_{e,j}. Propositions 1 and 2 concern the source sites and do not cover this extrapolation.

To characterize the effect of the assignment at an unseen site, we have added Proposition 3 to Sec. 3.1. At an unseen site tt that follows the additive model of Proposition 1, any fixed aggregate intercept bˉ\bar{\boldsymbol b} and coefficients Γˉ\bar{\boldsymbol\Gamma} give a corrected FC rires=ri−bˉ−Γˉ⊤q~i\boldsymbol r_i^{\mathrm{res}}=\boldsymbol r_i-\bar{\boldsymbol b}-\bar{\boldsymbol\Gamma}^{\top}\tilde{\boldsymbol q}_i whose diagnostic contrast satisfies

Dt[rires]=Dt[si⋆]+(Γt⋆−Γˉ)⊤Dt[q~i],\mathrm{D}_t\big[\boldsymbol r^{\mathrm{res}}_i\big]=\mathrm{D}_t[\boldsymbol s^\star_i]+(\boldsymbol\Gamma^\star_t-\bar{\boldsymbol\Gamma})^{\top}\mathrm{D}_t[\tilde{\boldsymbol q}_i],

where the entries of Dt[q~i]\mathrm{D}_t[\tilde{\boldsymbol q}_i] for covariates that are constant within site tt are zero. The result holds for any aggregate deconfounder, so it covers every choice of weights and every assignment of the source offsets. The intercept and the offset of the target site cancel in its diagnostic contrast, and the assignment enters only through the rows of Γˉw\bar{\boldsymbol\Gamma}_w for covariates that vary within the target site. It reaches these rows through the reliability weights and, on ABIDE, through the sex coefficients of SBL and TRINITY. The remaining term vanishes when the diagnostic groups at the target site are matched on these covariates. The part of the correction shared by all subjects of the target site acts on the operating point of the classifier. For a linear readout, it moves all scores of the site by the same amount, which changes ACC and leaves AUC unchanged. For the nonlinear classifier of CORE, this relation holds approximately, so ACC is expected to respond more strongly than AUC.

To test how strongly the correction depends on the assignment and whether this dependence affects the results, we compared the submitted deconfounder with four variants. The variants are constructed as follows:

  • Default is the submitted deconfounder, which fits each source site with standardized covariates and the pseudoinverse.
  • Default ×0.1 and Default ×10 rescale the site-level covariates by 0.1 and 10 before the same fit. A smaller scale moves the minimum-norm assignment toward the intercept, and a larger scale moves it toward the site-level covariates.
  • Omitted removes the covariates that are constant within a site from the fit of that site. Each within-site slope is aggregated, and enters ueu_e, only over the sites where it is estimated, with renormalized weights. Every coefficient in Eqs. 7 and 8 is identified under this variant.
  • Joint estimates the acquisition slopes across source sites, where TR and sequence length vary, by regressing the source intercepts on the site-level covariates with the Huber loss, and applies its prediction at the unseen site. This variant is identified when the site offset is linear in the acquisition parameters, and each fold provides seven training sites for its estimation.
AuthorsOpenReview ↗

Response to Reviewer d6U7 (2/8)

All variants yield the same site-wise residuals, scaffold and training features, and differ only in the correction of validation and test subjects. We ran CORE with each variant, BrainOOD and DeCI on ABIDE, REST-meta-MDD and SRPBS with three seeds and the same validation sites. ABCD is excluded, as TR is 0.8 s and the sequence length is identical at all its sites, so all variants coincide.

The table below reports the LOSO AUC and ACC in %. For each seed, we average the metric over all target sites of a dataset, and each entry is the mean ± SD of the three seed-level values. ΔC\Delta_C is the mean absolute deviation of the unseen-site correction of a variant from Default, in units of the across-subject SD of the edge at the target site, averaged over edges, held-out subjects and folds. Since a variant changes both the reliability weights and the aggregated coefficients, ΔC\Delta_C measures their combined effect on the correction.

MethodABIDE AUC↑ABIDE ACC↑ABIDE ΔC\Delta_C↓REST-meta-MDD AUC↑REST-meta-MDD ACC↑REST-meta-MDD ΔC\Delta_C↓SRPBS AUC↑SRPBS ACC↑SRPBS ΔC\Delta_C↓
BrainOOD63.68±0.2260.28±1.13–62.56±0.7458.84±0.59–90.15±0.3677.40±0.41–
DeCI55.90±0.3160.84±0.38–55.06±0.7256.22±0.16–91.58±1.0685.03±0.25–
CORE, Default65.21±0.7864.95±1.11065.58±0.6263.74±0.79092.36±0.5786.77±0.630
CORE, Default ×0.165.09±0.7464.77±0.980.2165.51±0.6363.44±0.560.1392.22±0.5186.26±0.720.30
CORE, Default ×1064.62±0.8963.17±1.120.6465.13±0.4562.25±0.590.4191.55±0.4884.36±0.790.79
CORE, Omitted64.99±0.6664.61±1.250.2365.43±0.6563.36±0.750.1392.13±0.5786.47±0.520.31
CORE, Joint64.66±0.7363.80±1.090.3765.30±0.8263.22±1.030.2591.36±0.4884.33±0.760.49

As shown in the above table, the deviation of the correction grows with the scale of the site-level covariates. ΔC\Delta_C ranges from 0.13 to 0.30 under ×0.1 and from 0.41 to 0.79 under ×10, which confirms that the pseudoinverse correction depends on the covariate scale. The accuracy follows the pattern predicted by Proposition 3. Under ×10, ACC decreases by 1.49 to 2.41 points across the three datasets, while AUC decreases by only 0.45 to 0.81 points, consistent with a shift of the operating point. Under ×0.1, both metrics change by at most 0.51 points. Omitted shifts the correction by a similar amount as ×0.1, with ΔC\Delta_C between 0.13 and 0.31, and stays within 0.23 points of Default in AUC and within 0.38 points in ACC on all three datasets. These differences are smaller than the SD over three seeds in every column. Joint falls below Omitted, with the largest gap on SRPBS, where ACC drops from 86.47 to 84.33. The training sites of SRPBS cover only three sequence lengths, which limits the estimation of acquisition slopes across sites. Compared with the baselines, Default, ×0.1 and Omitted exceed BrainOOD and DeCI in every column. Under ×10 and Joint, CORE remains above BrainOOD in every column and falls slightly below DeCI on SRPBS, by 0.03 and 0.22 points in AUC and by 0.67 and 0.70 points in ACC.

AuthorsOpenReview ↗

Response to Reviewer d6U7 (3/8)

To identify how the assignment changes the correction, we decomposed the deviation of each variant from Default into the part shared by all subjects of a target site and the part that varies across subjects, and compared the reliability weights of each variant with those of Default by their total variation distance and their effective number of sites 1/∑eαe21/\sum_e\alpha_e^2. The decomposition shows that the shared part dominates the deviation. It reaches 0.76 SD of the edge, while the part that varies across subjects stays below 0.11 SD, which agrees with the small changes in AUC in the above table. The comparison of the weights shows that they account for the larger cost of ×10. Under Default, the effective number of sites lies between 6.6 and 6.7 of the seven training sites, so the weights are close to uniform. Scaling by 10 moves the weights by a total variation distance of 0.16 to 0.26 and lowers the effective number of sites to between 4.6 and 5.4, which concentrates the aggregated slopes on fewer sites. Omitted moves the weights by only 0.05 to 0.07.

To quantify the remaining term in Proposition 3, we computed Rt,j=(γ^t,j−γˉw,j)⊤D^t[q~i]R_{t,j}=(\hat{\boldsymbol\gamma}_{t,j}-\bar{\boldsymbol\gamma}_{w,j})^{\top}\hat{\mathrm{D}}_t[\tilde{\boldsymbol q}_i] of the submitted deconfounder at each target site. The target slopes γ^t,j\hat{\boldsymbol\gamma}_{t,j} are fitted post hoc on the target site with its labels and are never used at inference. Estimation noise in these slopes tends to inflate ∣Rt,j∣\vert R_{t,j}\vert, so the reported values are conservative. We also measured the standardized mean difference, SMD, of age, sex and mean framewise displacement, FD, between the diagnostic groups within each site. The table below reports the median ∣Rt,j∣\vert R_{t,j}\vert in units of the across-subject SD of the edge, the median ratio ρt\rho_t of ∣Rt,j∣\vert R_{t,j}\vert to the observed contrast, the share of edges whose contrast changes sign when Rt,jR_{t,j} is removed, and the same share among edges whose contrast exceeds its standard error, averaged over folds, on scaffold edges and on all edges.

DatasetScaffold ∣R∣\vert R\vert↓Scaffold ρt\rho_t↓Scaffold Flip %↓Scaffold Flip>SE %↓All ρt\rho_t↓All Flip %↓All Flip>SE %↓
ABIDE0.0350.123.90.000.175.60.07
REST-meta-MDD0.0270.113.30.050.227.20.37
SRPBS0.0200.040.60.000.134.30.06

As shown in the above table, the remaining term is small on all three datasets, with a median of 0.020 to 0.035 SD of the edge on scaffold edges. It is also smaller relative to the contrast on scaffold edges than on all edges, with ρt\rho_t between 0.04 and 0.12 against 0.13 to 0.22. Removing the remaining term reverses the contrast of 0.6% to 3.9% of the scaffold edges, compared with 4.3% to 7.2% of all edges, and of at most 0.05% of the scaffold edges whose contrast exceeds its standard error. SRPBS shows the smallest remaining term and the fewest sign changes. The small size of the term is consistent with the covariate balance within sites, where the diagnostic groups differ little in age and sex, with a mean ∣\vertSMD$\vert$ between 0.11 and 0.21, and mean FD on ABIDE shows the largest difference, with a mean ∣\vertSMD$\vert$ of 0.31. These results indicate that consensus selection over site-wise residuals retains edges with large and reproducible contrasts, which the remaining term rarely reverses. The assignment selected by the pseudoinverse acts mainly through a shift shared by all subjects of a target site, and the gains over the baselines hold under both the pseudoinverse assignment and the identified Omitted variant.

W2: Temporal features can reintroduce acquisition bias. Windows contain 30 time points, so their duration varies with TR; the max–min range also depends on the number of windows. These features bypass static residualization. Table 9 compares temporal encoders but lacks duration- and scan-length-matched controls.

A2: Thanks for your comment. The window length W=30W=30 and the stride Sw=5S_w=5 are fixed in time points, so the window duration scales with the TR, and the number of windows LiL_i grows with the scan length. On SRPBS, $T_i\in\lbrace 107,143,240\rbrace $ gives $L_i\in\lbrace 16,23,43\rbrace $. The variance of a windowed correlation depends on the window duration [1], and the expected max–min range fi,jf_{i,j} grows with LiL_i. The submitted descriptors are computed on time-point windows and are not residualized.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of revision

We sincerely thank all reviewers for their constructive comments. We have revised the manuscript accordingly, and all changes are highlighted in blue. The main revisions are summarized below.

1. Correction at an unseen site and covariates constant within a site

  • Added Proposition 3 on the diagnostic contrast at an unseen site under any aggregate deconfounder to Sec. 3.1, with its proof in Appendix D, and described the pseudoinverse fit for covariates constant within a site in Sec. 3.1. (Reviewers d6U7, WV4d, fiGY)
  • Added the rescaling of the site-level covariates by 0.1 to 10, the rescaling of TR or sequence length alone, the identified Omitted variant and the Joint variant, with the deviation of the correction and comparisons with BrainOOD and DeCI, in Appendix J.9 (Table 21). (Reviewers d6U7, WV4d, fiGY)
  • Added the magnitude of the remaining term of Proposition 3 at the target sites and its effect on the sign of the diagnostic contrast in Appendix J.9 (Table 22). (Reviewers d6U7, WV4d)

2. Repeated evaluation and paired statistics

  • Added paired per-site comparisons of CORE with BrainOOD and DeCI over three seeds and three validation-site draws per target site, with two-level bootstrap confidence intervals, Holm-corrected Wilcoxon tests, and the SD of the gain across target sites, seeds and validation-site draws, in Appendix J.2 (Table 9). (Reviewers d6U7, WV4d)
  • Revised the Abstract, Introduction and Sec. 4.2 so that the relative improvement is computed from the nine-run means. (Reviewers d6U7, WV4d)

3. Generalization to unseen institutions

  • Added an institution-level evaluation of SRPBS, where all batches of an institution are held out together, in Appendix J.3 (Table 10), and stated in Appendix E.1 that the batch-level split is less strict than holding out entire institutions. (Reviewer WV4d)

4. Site dependence and deconfounding strategies

  • Added a comparison of no deconfounding, global deconfounding, and site-aware deconfounding with uniform or reliability-weighted aggregation, with site probes on the corrected FC and on the representation used for prediction and the heterogeneity of the source contrasts, in Appendix J.6 (Tables 15 and 16). (Reviewer WV4d)
  • Revised the Abstract, Introduction and Conclusion so that reliability weighting is no longer presented as a main component. (Reviewer WV4d)

5. Acquisition dependence of temporal descriptors

  • Added windows fixed in seconds, a 10–90% range, scans truncated to a common duration and residualized descriptors, with site probes on the descriptors and on the representation, in Appendix J.7 (Table 17). (Reviewers d6U7, WV4d)
  • Narrowed the claim on temporal descriptors, and described them as fixed-dimensional in the Introduction. (Reviewers d6U7, WV4d)

6. Isolating the scaffold and the line graph

  • Added variants that replace scaffold selection, the unseen-site correction or the classifier, including parameter-matched MLP, Transformer and GINE models, together with an established pipeline and ComBat with a GCN, in Appendix J.5 (Table 14). (Reviewers WV4d, fiGY)

7. Related work and scope of claims

  • Added demographic-enhanced brain network models, LG-GNN, KMGCN and BrainPrompt+, to the Related Work. (Reviewer fiGY)
  • Narrowed the claim on atlas robustness in the Abstract, Introduction, Sec. 4.2 and Conclusion to training and evaluation with either parcellation. (Reviewer fiGY)

We hope these revisions address the reviewers' concerns, and we are happy to provide further clarification during the discussion period.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.