acceptodds
Under review as a conference paper at ICLR 2027

Alzheimer's Disease Classification via Score-Derived Spectral Coupling from Longitudinal Cortical Data

Abstract

Longitudinal neuroimaging provides direct observations of within-subject disease-related change, but such data are typically sparse and irregularly sampled. We propose a spherical score-based framework for learning class-conditioned coupling representations from pairwise longitudinal cortical observations drawn from trajectories with two or more timepoints. Given a baseline-follow-up pair, we represent the interval-normalized cortical change in the spherical harmonic (SH) domain and train a class-conditioned score model using a spherical Ornstein-Uhlenbeck diffusion process. Rather than using the learned score for generation, we exploit its local differential structure: the negative score Jacobian defines a state-dependent local dependency operator over SH modes, reducing to the precision matrix in the Gaussian special case. We convert this score-derived operator into a normalized guidance operator and use it to modulate intermediate spectral features of a spherical cortical network under alternative diagnostic hypotheses. The proposed framework is evaluated on longitudinal cortical thickness, FDG-PET, and AV45-PET data from the ADNI dataset, achieving consistently strong classification performance across modalities. Additional cross-cohort and dataset-shift evaluations further support the generalizability of the learned coupling representation.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Spherical Harmonics, Longitudinal Data Analysis, Alzheimer's Disease, Score Matching, Geometric Deep Learning

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

Motivated by the literature that brain connectome organization and large-scale network structure are related to patterns of neurodegenerative vulnerability, this paper proposes a framework that represents interval-normalized cortical changes from baseline to follow-up pairs in the spherical harmonic (SH) domain. A diffusion score model is trained on these SH coefficients, conditioned on the diagnostic class, to estimate the class-conditional score. With the score model frozen, the Jacobian of the score (i.e., the Hessian of the log-density) is interpreted as a dependency operator among SH modes and used as guidance for the classifier under each diagnostic hypothesis. The model shows strong classification performance compared with other methods across modalities and under cross-cohort and phase-transfer settings.

Critical strengths
  1. The idea of representing longitudinal cortical changes in the spherical harmonic domain and modeling the dependencies among these changes is interesting, as this differs from existing approaches.

  2. They evaluated the method not only on a single modality but also on three modalities (cortical thickness, FDG-PET, and AV45-PET). The method also achieves strong performance in cross-cohort and phase-transfer settings. This suggests that the method generalizes beyond a single modality or dataset.

Critical weaknesses
  1. The first sentence of the abstract raises sparse and irregularly sampled longitudinal data as a challenge, so I read the paper somewhat expecting a way to overcome the limited availability of longitudinal data. However, the method still requires a baseline and follow-up pair for classification, so it cannot be applied to patients with only a single scan, which the authors also note as a limitation. There seems to be a slight mismatch between this motivation and the actual contribution. Framing the paper around the importance of learning the coupling structure among longitudinal cortical changes would provide more insight.

  2. Although some ablation results support the role of W, it is still unclear whether the performance gain comes from W itself or simply from using the follow-up information. The zero-change ablation removes the only pathway (W) through which follow-up information reaches the classifier, so the performance drop is somewhat expected. Since W is the core module, comparisons such as using S_f alone or S_f with W (see Requested additional experiments) would help clarify this.

Questions for authors

[Questions]

  1. In the ablation study, in the “Zero-change guidance” experiment, when you replace S_f with S_b, is the first stage (score network training) trained using both S_f and S_b, and only changed in the second stage?
  2. Curious about why you select S_b instead of S_f? Because using S_f with W also seems reasonable, as W still carries the change information from the baseline, so combining it with S_f (the more recent state) could provide more information.
  3. Which time point is the diagnostic label based on (baseline, follow-up, or final diagnosis)?

[Requested additional experiments]

  1. In “Zero-change guidance”, the classifier input is only S_b, and the guidance carries no change information. There is no follow-up information, although the label may be based on the follow-up. It seems that additional experiments would be helpful: (1) only using S_f instead of S_b at the second stage without guidance, (2) using S_f instead of S_b with W guidance.
  2. When conducting the classification using only W, how much does it work? I wonder to what extent W actually plays a role.
  3. In Figure 2, the figure seems to show the result from a randomly selected subject from each group, but how about showing it at the group level (e.g., group-averaged maps)?

[Minor comments]

  1. In Figure 1, instead of only notating J, W, etc., putting the short name of each module might help understanding.
  2. Figure 2 is not referenced in the text of Section 4.

[Suggestion]

  1. It might be interesting to use only S_b to predict or generate the changing information (e.g., SH coefficient c or W).
  2. Regarding Figure 2, if the data has CN (baseline) -> CN (follow-up), CN -> AD, and AD -> AD, it might be interesting to analyze each of those types or use a t-test for each pair or ANOVA.
5 comments
AuthorsOpenReview ↗

Author Response to Reviewer kmzb (Part 1/4)

We sincerely thank the reviewer for their thoughtful evaluation and constructive feedback. We appreciate the reviewer's recognition of the novelty of modeling dependencies among longitudinal cortical changes in the spherical harmonic domain, as well as the comprehensive evaluation across three neuroimaging modalities and cross-cohort/phase-transfer settings. We address the reviewer's concerns regarding the motivation and the specific contribution of the learned coupling structure through the clarifications and additional analyses below.

W1.

We thank the reviewer for this comment and would like to clarify the role of this motivation. We do not aim to resolve the sparsity of longitudinal data itself. Rather, sparse and irregular sampling is the main constraint under which coupling structure among longitudinal cortical changes must be learned: most subjects contribute only a few scans at irregular intervals, which rules out approaches that require dense or regularly sampled trajectories. Our design follows from this constraint. As stated in Section 1 (lines 44–46), we use the baseline–follow-up pair as the minimal unit of longitudinal supervision, so that subjects with only two valid observations contribute directly and longer trajectories can be decomposed into pairwise transitions, while irregular intervals are handled by normalizing change by ΔT\Delta T. The motivation therefore explains why we learn coupling from pairs, not a claim that the method removes data sparsity. To state the central motivation explicitly, we have revised the first sentence of the Abstract: “Longitudinal neuroimaging provides direct observations of within-subject disease-related cortical changes, yet learning the structured dependencies among spatially distributed cortical changes remains challenging.” The sampling constraint is retained in the Introduction as the reason for the pairwise design, and the requirement of two scans at inference is stated as a limitation in Section 5.

W2.

We should have stated this more clearly: the “No guidance” model in Table 3 is not an SbS_b-only model. It receives both SbS_b and SfS_f as input channels and therefore has access to the same baseline and follow-up information as the full model. Its AUC is 87.13%, compared with 91.54% when the follow-up information enters through WW. A second information-matched control that receives SbS_b, (Sf−Sb)/ΔT(S_f-S_b)/\Delta T and ΔT\Delta T directly reaches 85.66% (Appendix D.2). The Zero-change row was intended to show that the guidance depends on each subject’s observed change, not to separate WW from follow-up information; the information-matched comparisons above serve that purpose.

To address this distinction more directly, we conduct the SfS_f-only and Sf+WS_f+W comparisons suggested by the reviewer in Requested Additional Experiments. These experiments allow us to assess the discriminative value of the follow-up cortical measurements themselves and the additional contribution of integrating them with score-derived guidance. We provide the detailed experimental configurations and results in our response to the corresponding request below.

AuthorsOpenReview ↗

Author Response to Reviewer kmzb (Part 2/4)

Q1.

As the reviewer correctly pointed out, in the Zero-change guidance experiment, the score network in the first training stage is still trained using the original longitudinal cortical observation pairs (Sb,Sf)(S_b, S_f). After training, the score network parameters are frozen, and SfS_f is replaced with SbS_b only during guidance construction in the second stage.

Specifically, under the Zero-change condition, we set Sf=SbS_f = S_b, such that the observed longitudinal cortical change becomes zero. Consequently, the clean deformation coefficients become c0=0c_0 = 0. However, the spherical OU perturbation used for guidance computation is applied in the same manner as in the original setting, resulting in the score network input ctg=Σtg1/2ϵc_{t_g} = \Sigma_{t_g}^{1/2}\epsilon. Therefore, this experiment does not remove or modify the trained score network itself, but instead removes only the subject-specific longitudinal change information used to construct the guidance.

The purpose of this experiment is not to directly evaluate the necessity of longitudinal supervision in Stage 1, but rather to examine how downstream classification performance changes depending on whether each subject's actual longitudinal cortical changes are incorporated into the guidance, while keeping the trained score model unchanged.

Recognizing that this distinction was not sufficiently clear in the original manuscript, we clarify this distinction here. We further address the related additional experiments requested by the reviewer in Section R1 below.

Q2.

We intentionally selected the baseline scan SbS_b as the direct input to the classifier to assign distinct roles to the baseline cortical state and the coupling representation estimated from subsequent longitudinal changes. Specifically, SbS_b provides the subject's baseline cortical measurements, while the guidance WW represents the class-conditioned spectral dependency structure estimated from changes observed in the longitudinal pair. Under this design, the proposed model integrates longitudinal information through structural guidance rather than directly using the follow-up scan as a feature input.

Nevertheless, as the reviewer pointed out, incorporating the more recent cortical state SfS_f together with the guidance WW is also a reasonable alternative and may provide additional discriminative information. This configuration can be evaluated within the same proposed architecture simply by replacing SbS_b with SfS_f, without requiring structural modifications to the model. Accordingly, we conducted the reviewer-suggested SfS_f-only and Sf+WS_f+W experiments to more directly assess the discriminative contribution of the follow-up cortical measurements themselves and the additional contribution of score-derived guidance. The detailed experimental configurations and quantitative results are provided in our response to R1 below.

Q3.

We thank the reviewer for this important question. The baseline–follow-up pairs used in our study correspond to "CN to CN" or "AD to AD", with the same diagnostic label at both observation time points. Therefore, the diagnostic label in our study is not assigned retrospectively from one time point to another, but represents the diagnostic status consistently maintained throughout the observed longitudinal interval.

This setting reflects our objective of learning class-conditional coupling structures from longitudinal cortical changes observed within each diagnostic class, rather than predicting transitions between diagnostic states. Accordingly, we define CN and AD as separate diagnostic conditions and train class-specific score distributions using longitudinal observations in which the corresponding diagnostic status remains unchanged. This study design is motivated by previous neuroscience findings suggesting that longitudinal cortical changes and structural relationships observed within specific clinical diagnostic groups provide meaningful information for understanding the characteristics of neurodegenerative diseases [1, 2].

We also agree that analyzing diagnostic transition groups, as suggested by the reviewer, is an important research direction. Extending the current approach to investigate how spectral coupling structures evolve during transitions between diagnostic states could provide further insights into the progression of neurodegenerative diseases. We discuss this direction further in our response to the Suggestion section below.

[1] Xiao Y, Liao L, Huang K, Yao S, Gao L. Coupling Between Hippocampal Parenchymal Fraction and Cortical Grey Matter Atrophy at Different Stages of Cognitive Decline. Journal of Alzheimer’s Disease. 2023;93(2):791-801.

[2] Carmichael O, McLaren DG, Tommet D, Mungas D, Jones RN. Coevolution of brain structures in amnestic mild cognitive impairment. NeuroImage. 2013;66:449–456.

AuthorsOpenReview ↗

Author Response to Reviewer kmzb (Part 3/4)

R1.

We additionally conducted the SfS_f-only and Sf+WS_f+W controlled experiments suggested by the reviewer. These experiments were designed to distinguish the discriminative contribution of the follow-up cortical measurements themselves from the additional effect of incorporating score-derived guidance WW when those measurements are directly provided to the classifier. All experiments were conducted using the same ADNI-CTh data splits, backbone, and training configuration as the main experiments.

The results show that SfS_f-only achieves an ACC of 71.82±3.5371.82\pm3.53%, an AUC of 77.99±5.4177.99\pm5.41%, and an MCC of 0.421±0.0690.421\pm0.069. In contrast, Sf+WS_f+W achieves an ACC of 83.50±4.3183.50\pm4.31%, an AUC of 90.87±5.3290.87\pm5.32%, and an MCC of 0.657±0.0910.657\pm0.091, corresponding to improvements of 11.68 percentage points, 12.88 percentage points, and 0.236, respectively.

These results are consistent with the information-matched comparison above: with guidance, either scan as the direct input yields about 91% AUC (Sb+WS_b+W: 91.54%, Sf+WS_f+W: 90.87%), whereas without guidance even both scans together yield 87.13% (Table 3, “No guidance”).

Furthermore, Sf+WS_f+W achieves an AUC of 90.87%, comparable to the 91.54% obtained by the originally proposed Sb+WS_b+W configuration. This finding is consistent with the original design rationale of assigning distinct roles to the baseline cortical state and longitudinal coupling, while also suggesting that directly using more recent cortical measurements is a viable alternative.

We have added these experiments in Appendix D.2 (Table 9).

R2.

We conducted the additional WW-only classification experiment suggested by the reviewer. The purpose of this experiment was to directly examine whether the proposed score-derived guidance operator merely assists cortical feature extraction or can independently represent diagnostically useful information derived from longitudinal cortical changes.

To this end, we performed classification using only the guidance operator estimated from the ScoreNet trained on baseline and follow-up cortical measurements. We employed L2-regularized logistic regression as the classifier, following the same ADNI-CTh data splits and evaluation protocol as in the main experiments.

The W-only classification achieved ACC 75.21±4.7975.21 ± 4.79%, AUC 82.45±4.0082.45 ± 4.00%, and MCC 0.465±0.1120.465 \pm 0.112. These results indicate that the guidance operator contains diagnostic class-related structural information beyond its role in feature mixing, and that such information can be extracted even with a relatively simple linear classifier.

Meanwhile, as reported in the manuscript, the proposed Full Model outperformed W-only classification by 8.98 percentage points in ACC, 9.09 percentage points in AUC, and 0.204 in MCC. For clarity, we summarize the results reported in the manuscript alongside those of the additional experiment in the table below.

MethodACC (%)AUC (%)MCC
WW-only (Logistic Regression)75.21 ± 4.7982.45 ± 4.000.465 ± 0.112
Full Model84.19 ± 2.1091.54 ± 2.280.669 ± 0.045

These findings suggest that W not only contains discriminative information on its own but also support the effectiveness of the proposed framework's design choice to integrate W into cortical feature extraction.We have added these experiments in Appendix D.2 (Table 9).

R3.

We thank the reviewer for suggesting group-level visualization to assess whether subject-specific coupling patterns exhibit consistent group-level tendencies. Following this suggestion, we aligned each subject’s top-5% spatial coupling participation map to a common FreeSurfer surface template and averaged them within each diagnostic group for CTh and FDG-PET.

As shown in the newly added Figure 4 (Appendix E), CN and AD groups exhibit distinct spatial participation patterns, with relatively higher participation in several posterior and medial cortical regions in the AD group. These regions qualitatively overlap with areas implicated in prior AD studies discussed in the manuscript.

MinorComments.

We thank the reviewer for their careful review and constructive suggestions aimed at improving the readability and clarity of our manuscript.

  1. Following the reviewer's suggestion, we will revise Figure 1 to include brief module names alongside the existing notations, making the overall pipeline of the proposed model more intuitive and easier to follow.
  2. We have added the missing reference to Figure 2 at the appropriate location in Section 4.
Reviewer_hpSWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Paper uses longitudinal cortical scan pairs to learn class-conditioned scores in spherical harmonic space. The negative score Jacobian gives a coupling operator to guide a spherical classifier. Evaluation covers cortical thickness and two PET modalities in ADNI, with transfer to OASIS-2 and between ADNI phases.

Critical strengths
  • Using the score Jacobian to capture dependencies between longitudinal cortical changes is interesting, and the mathematical motivation is clear.
  • Fairly broad evaluation: multiple modalities, 20 baselines, transfer tests, and component ablations.
Critical weaknesses
  • Main benefit of the score-derived operator still is not isolated.
    • Random, diagonal-only, and off-diagonal controls do not replace comparisons with simple class-conditional covariance or precision, or learned spectral attention using the same backbone.
    • Diagonal-only already does resonably well.
  • Generalization claim needs more care, especially for AV45.
    • Only 30 AD subjects in AV45.
    • Method is not best in MCC on AV45.
    • No paired significance analysis for the reported gains.
  • Guidance calculation also uses diffusion noise, but its variability at inference is unclear.
  • Reference errors too: author names are wrong in Jedynak et al. (2012) and Zhao et al. (2019).
Questions for authors
  • How does the method compare with a regularized class-conditional precision matrix or learned spectral affinity, using the same splits and backbone?
  • Is guidance averaged over diffusion-noise draws or computed once? How stable are results across draws and seeds, especially for AV45 and cross-cohort evaluation?
7 comments
AuthorsOpenReview ↗

Author Response to Reviewer hpSW (Part 1/4)

We sincerely thank the reviewer for their careful evaluation and constructive feedback. We particularly appreciate the reviewer's recognition of the novelty and clear mathematical motivation of our approach to modeling dependencies among longitudinal cortical changes, as well as the breadth of our experimental evaluation.

We understand that the reviewer's main concerns are (W1) the need for more controlled comparisons to isolate the specific contribution of the score-derived guidance operator, (W2) the generalization performance, particularly on AV45-PET, and the statistical significance of the reported improvements, and (W3) the stability of guidance under diffusion noise during inference. Below, we address these concerns sequentially through additional experimental results and detailed clarifications.

W1.

We agree with the reviewer's concern that the contribution of the proposed score-derived spectral coupling operator should be distinguished from that of simpler spectral dependency estimation methods. Accordingly, in addition to the existing random guidance and component ablations, we conducted further experiments using two alternative operators: Learned Spectral Affinity and Ledoit–Wolf Precision Guidance. All experiments were conducted using the same ADNI-CTh data splits, downstream classifier backbone, and training configuration. The proposed method achieved an AUC of 91.54%, outperforming Ledoit–Wolf Precision Guidance (84.87%) and Learned Spectral Affinity (77.78%). Detailed experimental configurations and complete results are provided in our response to Q1 below.

We also acknowledge the reviewer's observation that diagonal-only guidance performs reasonably well, whereas off-diagonal-only guidance yields lower performance. On ADNI-CTh, diagonal-only and off-diagonal-only guidance achieved AUCs of 88.23% and 85.86%, respectively, while full guidance achieved 91.54%, outperforming both configurations. These results do not imply that off-diagonal coupling alone accounts for the overall classification performance. Rather, we would like to clarify that the objective of our study is not to establish the independent discriminative power of off-diagonal interactions, but to investigate the utility of a class-conditioned coupling representation that integrates diagonal information with off-diagonal spectral dependencies. In particular, the 3.31 percentage-point AUC improvement of full guidance over diagonal-only guidance suggests that jointly incorporating both components can provide additional benefits for the classification task considered in this study. Section 3.3 of the revised manuscript presents the diagonal and off-diagonal components as complementary.

W2.

We thank the reviewer for raising concerns regarding the limited sample size of AV45-PET and the statistical reliability of the reported performance improvements.

First, we acknowledge that ADNI-AV45 includes only 30 AD subjects, corresponding to approximately six AD subjects per test fold in 5-fold cross-validation. This limited sample size increases the uncertainty of performance estimates and warrants cautious interpretation. We also acknowledge that our method achieves a slightly lower MCC than MS-SiT (0.451 vs. 0.465). However, our method achieves a higher mean AUC (86.76% vs. 85.94%). Therefore, rather than claiming superiority across all metrics on AV45, we emphasize that our method demonstrates competitive classification performance under limited-sample conditions.

To further assess reproducibility, we repeated the AV45-PET and ADNI-CTh to OASIS-2 cross-cohort evaluations using five independent random seeds. On AV45-PET, the additional runs yielded an AUC of 85.76 ± 4.00% and an MCC of 0.452 ± 0.072. For cross-cohort evaluation, the corresponding results were an AUC of 74.21 ± 0.62% and an MCC of 0.378 ± 0.010. These findings support that the mean performance reported in the manuscript is not solely dependent on a particular training initialization. The experimental settings and detailed results are provided in our response to Q2.

Furthermore, to strengthen the statistical evidence for the reported performance improvements, we conducted paired t-tests against 20 baseline methods on ADNI-CTh using identical held-out folds, followed by Holm correction for multiple comparisons. All 20 AUC comparisons remained significant after Holm correction (adjusted p < 0.05). In particular, our method achieved an AUC improvement of 9.85 percentage points over MS-SiT (adjusted p = 0.02464) and 21.64 percentage points over MoNet (adjusted p = 0.00256).

Nevertheless, we recognize that these statistical results on ADNI-CTh do not establish statistical superiority on AV45. Given the limited number of AD subjects and the small AUC difference relative to MS-SiT, Section 3.2 interprets AV45 primarily through AUC and MCC, and the Conclusion lists its small AD sample as a limitation.

AuthorsOpenReview ↗

Author Response to Reviewer hpSW (Part 2/4)

W3.

We thank the reviewer for raising concerns regarding the stability of the guidance operator under inference-time diffusion noise and independent ScoreNet training. To address these concerns, we conducted additional experiments evaluating operator reproducibility across independent OU noise draws and ScoreNet training seeds, as well as downstream classification stability across multiple training seeds. The detailed experimental settings and quantitative results are provided in our response to Q2 below.

W4.

We thank the reviewer for pointing out the incorrect author names in the manuscript. We have verified the correct bibliographic information for the cited references and corrected these errors in the revised manuscript.

Q1.

We thank the reviewer for suggesting comparisons with a regularized class-conditional precision matrix and a learnable spectral affinity. To more directly evaluate the contribution of the proposed score-derived coupling operator, we conducted additional ablation experiments using two alternative guidance constructions.

The first is Attention-based Learned Affinity, which directly learns subject-specific spectral affinity from the SH coefficients of longitudinal changes. The second is Ledoit–Wolf Precision Guidance, which uses a precision matrix estimated by a class-conditional Ledoit–Wolf shrinkage estimator as guidance. All methods were evaluated using the same ADNI-CTh data splits, downstream classifier backbone, and training/evaluation configuration. The results are summarized in the table below.

GuidanceACC (%)MCCAUC (%)
Learned Affinity72.73 ± 4.290.439 ± 0.09077.78 ± 4.92
Ledoit–Wolf Precision75.93 ± 6.800.506 ± 0.14184.87 ± 6.28
Ours84.19 ± 2.100.669 ± 0.04591.54 ± 2.28

The experimental results show that Learned Affinity and Ledoit–Wolf Precision Guidance achieved AUCs of 77.78 ± 4.92% and 84.87 ± 6.28%, respectively, whereas the proposed method achieved 91.54 ± 2.28%, corresponding to AUC improvements of 13.76 and 6.67 percentage points, respectively. Compared with Ledoit–Wolf Precision Guidance, the proposed method also achieved improvements of 8.26 percentage points in ACC and 0.163 in MCC

While Ledoit–Wolf Precision estimates global second-order dependencies in the class-conditional distribution of longitudinal changes, the proposed method constructs a subject-dependent local dependency operator by evaluating the Jacobian of the learned class-conditional score field at each subject's OU-perturbed coefficient state. Furthermore, the comparison with Learned Affinity examines whether directly learned subject-specific spectral interactions from longitudinal changes can reproduce the performance of the proposed method.

Two notes on these comparisons. First, under a Gaussian class-conditional model the score of the OU-smoothed density is linear, so its negative Jacobian is state-independent and equals the regularized class-conditional precision (AtgΓyAtg⊤+Σtg)−1(A_{t_g}\Gamma_y A_{t_g}^\top+\Sigma_{t_g})^{-1} (Appendix A.2). Ledoit–Wolf guidance is thus the Gaussian, state-independent counterpart of our operator, and the gap measures the contribution of subject-dependent structure in the learned score. Second, both alternatives fall below the unguided backbone (87.13%, which receives both SbS_b and SfS_f). The learned affinity must learn an Mg×MgM_g \times M_g interaction matrix from the classification signal alone, whereas our operator is estimated from the density of longitudinal changes; at this sample size, the former does not recover the same benefit.

Beyond the existing component ablations, these results provide additional empirical evidence supporting the utility of the score-derived coupling representation, demonstrating higher classification performance than both alternative guidance constructions under the same downstream architecture. These results are reported in Appendix D.1 (Table 8).

AuthorsOpenReview ↗

Author Response to Reviewer hpSW (Part 3/4)

Q2.

We thank the reviewer for raising important questions regarding the reproducibility of the guidance operator under inference-time diffusion noise and independent score network training initializations, as well as the stability of downstream classification performance. To address these concerns, we conducted additional experiments examining both operator-level and classification-level reproducibility.

At a fixed guidance diffusion time tgt_g, we apply a Spherical Ornstein–Uhlenbeck (OU) perturbation to each subject's longitudinal SH coefficients c0c_0, constructing the evaluation state as

ctg=Atgc0+Σtg1/2ϵ,ϵ∼N(0,I).c_{t_g}=A_{t_g}c_0+\Sigma_{t_g}^{1/2}\epsilon, \epsilon\sim\mathcal N(0,I).

The guidance operator WW is then constructed by evaluating the Jacobian of the learned class-conditional score field at this perturbed state. For each subject and diagnostic hypothesis, we use a single Gaussian noise draw rather than Monte Carlo averaging over multiple draws. The resulting operator is cached for downstream classification.

To evaluate the influence of stochastic perturbations on operator reproducibility, we fixed the trained score network and independently resampled Gaussian noise five times, reconstructing the guidance operator for each draw. Separately, to assess variability arising from independent score network training, we trained five score networks with different random seeds under identical ADNI-CTh data splits and experimental configurations. We measured Pearson correlations between the corresponding estimated guidance operators, as summarized below.

ComparisonPearson Correlation
Across independent noise draws0.995 ± 0.003
Across independent ScoreNet training seeds0.677 ± 0.176

When varying only the inference-time noise draw, the estimated operators exhibited a Pearson correlation of 0.995±0.0030.995\pm0.003, indicating that their relative coupling patterns remained largely unchanged despite stochastic perturbations of the Jacobian evaluation state.

Across independently trained ScoreNets, the correlation was 0.677±0.1760.677 \pm 0.176. Lower element-wise agreement is expected here, since independently trained networks need not converge to identical solutions even when they capture the same structure [1,2]. To test whether subject-specific information is nevertheless preserved, we matched each subject’s operator from one score network against the operators of 117 held-out subjects from another score network: Top-1 identification accuracy was 87.91±16.26%87.91 \pm 16.26\% across all 10 score network pairs (chance ≈ 0.85%; Appendix D.3).

Beyond operator-level analysis, we independently repeated both ADNI-AV45 classification and cross-cohort ADNI-CTh → OASIS-2 evaluation using five additional random seeds, while maintaining the original data splits, training configurations, and evaluation protocols. The results are summarized below.

EvaluationOriginal AUC (%)Additional 5 Seeds AUC (%)Original MCCAdditional 5 Seeds MCC
ADNI-AV4586.7685.76 ± 4.000.4510.452 ± 0.072
ADNI-CTh → OASIS-273.0674.21 ± 0.620.3370.378 ± 0.010

[1] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neu- ral network representations revisited. In Kamalika Chaudhuri and Ruslan Salakhutdinov (eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97 of Proceed- ings of Machine Learning Research, pp. 3519–3529. PMLR, 09–15 Jun 2019.

[2] Yixuan Li, Jason Yosinski, Jeff Clune, Hod Lipson, and John E. Hopcroft. Convergent Learn- ing: Do Different Neural Networks Learn the Same Representations? In International Con- ference on Learning Representations, 2016.

Reviewer_XtJ2OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper models interval-normalized longitudinal cortical changes in spherical harmonic coordinates using a class-conditioned score network. It builds a local operator from the symmetrized negative score Jacobian. It then takes the element-wise absolute value, applies symmetric degree normalization, and uses the result to guide a spherical classifier under alternative diagnostic hypotheses. Evaluation covers cortical thickness and two PET modalities, together with ADNI-to-OASIS-2 cross-cohort and ADNI-1-to-ADNI-2 phase-transfer experiments.

Critical strengths
  1. The method offers a coherent connection between score estimation and spectral feature guidance. The mathematical discussion appropriately distinguishes a local differential operator from a general global dependency graph and explains the Gaussian precision-matrix special case. Evaluating alternative diagnostic hypotheses at inference does not itself require the unknown test label.

  2. The experimental evidence is comparatively substantial.

    • Guidance gain. On cortical thickness, full guidance achieves 91.54% AUC versus 87.13% without guidance.
    • Ablations and transfer. The paper includes component ablations and transfer experiments.
    • Interval control. Appendix B.4 addresses an important potential confound by matching cortical-thickness acquisition intervals between groups (mean 1.52 years for both CN and AD). An interval-only classifier then achieves 48.17% AUC.
    • These checks materially strengthen the central result.
Critical weaknesses
  1. The predictive value of guidance is better established than the accuracy of the proposed dependency representation.
    • Derivative accuracy. Good score-value estimation does not automatically establish accurate score derivatives, particularly with small samples. Symmetrizing the Jacobian does not resolve derivative-estimation error.
    • Missing checks. The paper reports a performance-level sensitivity analysis over t_g and L_g (Appendix C). Beyond that, the experiments do not directly establish operator reproducibility, and they do not compare against simpler class-conditioned covariance/precision guidance.
    • Ablation results. Random guidance is a useful negative control, but its poor performance does not establish the necessity of the score-derived construction. The component ablations also temper the dependency interpretation. Off-diagonal-only guidance (85.86% AUC) falls below the unguided backbone (87.13%), while diagonal-only guidance reaches 88.23% (Table 3). These results do not establish a standalone benefit from off-diagonal guidance. The full operator's advantage over diagonal-only guidance (91.54% vs. 88.23%) does, however, support a complementary contribution.
  2. Some generalization and modality claims remain uncertain.
    • Interval effects. Interval-only AUC is 56.61% for FDG and 59.96% for AV45, and the PET cohorts differ in mean interval by group (FDG CN 2.03 vs. AD 1.71 years; AV45 CN 2.58 vs. AD 2.27 years; Table 4). OASIS also has different mean scan intervals between diagnostic groups (CN 3.32 vs. AD 2.28 years). Separately, on interval-matched cortical thickness, guidance without interval normalization yields 87.04% AUC, essentially equal to the unguided 87.13% (Table 3). This is not itself a confound, but in this reported ablation the AUC gain disappears without interval normalization. That makes the interval handling in the unmatched PET and OASIS cohorts more consequential.
    • Sample size. The AV45 cohort is also highly imbalanced, with only 30 AD participants.
    • These observations do not invalidate the reported AUC, but they limit conclusions about robustness across modalities and acquisition settings. A matched longitudinal baseline given the acquisition interval explicitly would make the attribution to learned coupling more convincing.
Questions for authors
  1. How reproducible are the guidance operators across independent score-network fits? Please specify how the evaluation state c_tg is constructed; if stochastic perturbations are sampled during guidance construction, report sensitivity to those draws. Does operator stability predict classification performance?
  2. How does the method compare with regularized class-conditioned covariance/precision guidance and with a backbone given both longitudinal change and acquisition interval explicitly?
  3. Do the PET and OASIS advantages persist after additional interval adjustment or matching? Beyond the fold-level standard deviations already reported, please include confidence intervals or paired significance tests appropriate to these sample sizes.
8 comments
AuthorsOpenReview ↗

Author Response to Reviewer XtJ2 (Part 1/6)

We sincerely thank the reviewer for their careful evaluation and constructive feedback. We particularly appreciate their positive assessment of the mathematical justification and clarity of our score-derived spectral guidance framework, as well as the extensive experimental validation across multiple neuroimaging modalities, transfer settings, and ablation studies.

We address the reviewer's main concerns regarding (W1) the accuracy, reproducibility, and distinctive contribution of the score-derived dependency operator, including comparisons with alternative guidance methods and longitudinal information-based baselines, and (W2) acquisition interval effects, PET generalization, and statistical significance. Below, we provide additional experimental results and clarifications in response to these concerns and the specific questions (Q1–Q3).

W1.

We thank the reviewer for the specific and constructive comments regarding the estimation accuracy, reproducibility, and distinctive contribution of the score-derived operator. To address these concerns, we conducted additional analyses of operator reproducibility across independent ScoreNet training runs and OU noise perturbations, along with controlled experiments using alternative guidance constructions based on simpler statistical dependencies and learned spectral affinity.

1. Score Jacobian Estimation Accuracy and Operator Reproducibility

We agree with that accurate score estimation does not necessarily guarantee accurate Jacobian estimation, and that symmetrization does not eliminate derivative estimation errors. Accordingly, we distinguish the mathematical interpretation of the ideal score Jacobian from the estimation accuracy and empirical reproducibility of the learned operator, without claiming recovery of the true dependency structure.

To assess reproducibility, we conducted additional analyses across independent ScoreNet training seeds and inference-time OU noise draws, including cross-fit subject identification. These experiments demonstrate high consistency under stochastic perturbations and substantial preservation of subject-specific structure across independently trained models, although reproducibility does not directly establish derivative estimation accuracy. Detailed protocols and quantitative results are provided in our response to Q1 below.

2. Controlled Comparisons with Class-Conditional Precision and Learned Spectral Affinity

We agree that the existing random-guidance experiments and component ablations alone are insufficient to establish the distinctive benefits of the score-derived construction. Accordingly, we conducted controlled comparisons with Attention-based Learned Spectral Affinity and Ledoit–Wolf Precision Guidance under identical ADNI-CTh data splits, downstream backbone, and training configurations.

The proposed method achieved AUC improvements of 13.76 and 6.67 percentage points over Learned Spectral Affinity and Ledoit–Wolf Precision Guidance, respectively. These results provide empirical evidence of the discriminative utility of the proposed score-derived operator compared with alternative guidance constructions. Detailed experimental configurations and quantitative results are provided in our response to Q2 below.

AuthorsOpenReview ↗

Author Response to Reviewer XtJ2 (Part 2/6)

3. Interpretation of Diagonal and Off-Diagonal Component Ablations

We also acknowledge the reviewer's observation that diagonal-only guidance achieves relatively strong performance, whereas off-diagonal-only guidance performs less favorably.

In the original ADNI-CTh component ablation, off-diagonal-only guidance achieved an AUC of 85.8685.86%, which was lower than the unguided backbone at 87.1387.13%. In contrast, diagonal-only guidance achieved 88.2388.23%, while Full Guidance reached 91.5491.54%. We also agree that the relatively poor performance of random guidance alone does not establish the necessity of the score-derived construction.

These findings do not suggest that off-diagonal coupling alone explains the overall classification performance. Rather, we clarify that the objective of our framework is not to demonstrate the standalone discriminative power of off-diagonal interactions, but to evaluate the utility of a class-conditioned coupling representation that integrates diagonal information with non-diagonal spectral dependencies.

Specifically, Full Guidance improved AUC by 3.31 percentage points over diagonal-only guidance, 5.68 percentage points over off-diagonal-only guidance, and 4.41 percentage points over the unguided backbone. The improvement of Full Guidance over diagonal-only guidance does not necessarily imply that off-diagonal components are independently effective as a guidance signal. Instead, it supports their potential complementary contribution when integrated with diagonal information.

This interpretation is consistent with the performance difference between Full Guidance and diagonal-only guidance noted positively by the reviewer. Accordingly, in the revised manuscript, we have clarified the complementary roles of diagonal and off-diagonal components in the score-derived representation, without overstating the independent contribution of off-diagonal coupling.

Overall, the additional operator reproducibility analyses and controlled comparisons with alternative guidance constructions provide more direct empirical evidence for the reproducibility and discriminative utility of the proposed score-derived operator. At the same time, we explicitly distinguish these empirical findings from mathematical guarantees of dependency estimation accuracy and have refined the interpretation of the component ablations in the revised manuscript.

W2.

We thank the reviewer for raising concerns regarding acquisition interval differences and the limited AV45-PET sample size in PET and cross-cohort evaluations.

In longitudinal neuroimaging, irregular acquisition intervals and limited follow-up observations make it difficult to fully match intervals across diagnostic groups. While the relatively abundant CTh data allowed us to control interval differences through baseline–follow-up pair selection, the more limited PET observations left residual imbalances. We therefore included a Scan Interval Control analysis (Appendix B.4), which showed weak diagnostic discrimination from acquisition intervals alone on PET datasets, without ruling out potential interval-related confounding.

To address this concern more directly, we conducted additional interval-matched evaluations on FDG-PET, AV45-PET, and OASIS-2 while retaining all AD subjects. The proposed method maintained competitive classification performance under reduced interval imbalance, although not all comparisons reached statistical significance. Detailed procedures, quantitative results, and paired statistical analyses are provided in our response to Q3 and Appendix D.4.

We acknowledge that the limited sample size, particularly the 30 AD subjects in AV45-PET, restricts statistical reliability. This limitation is also discussed in the Conclusion, with validation on larger longitudinal cohorts remaining an important direction for future work.

AuthorsOpenReview ↗

Author Response to Reviewer XtJ2 (Part 3/6)

Q1.

We thank the reviewer for the questions regarding the inference-time construction of the guidance operator and its reproducibility across independently trained ScoreNets. We believe these questions provide an opportunity to further clarify the proposed framework. To address these concerns, we conducted additional experiments distinguishing operator-level reproducibility from classification-level reproducibility.

In our framework, at a fixed guidance diffusion time tgt_g, we apply a Spherical Ornstein–Uhlenbeck (OU) perturbation to each subject's longitudinal SH coefficients c0c_0 to construct the evaluation state ctgc_{t_g}. The guidance operator WW is then constructed by evaluating the Jacobian of the learned class-conditional score field at this perturbed state. For each subject and diagnostic hypothesis, we use a single noise draw without Monte Carlo averaging over multiple draws. The resulting guidance operator is cached for downstream classification.

To evaluate the effects of Spherical OU perturbations and independent ScoreNet training on operator reproducibility, we conducted two separate analyses. First, we fixed the trained ScoreNet and independently sampled Gaussian noise five times to repeatedly estimate the guidance operator. Second, to assess variability arising from independent training initializations, we trained five ScoreNets with different random seeds under identical data splits and experimental configurations, and compared the operators estimated for the same subjects. We measured Pearson correlations between the corresponding guidance matrices, as summarized below.

ComparisonPearson Correlation
Across independent noise draws0.995 ± 0.003
Across independent ScoreNet training seeds0.677 ± 0.176

When varying only the noise draw while keeping the ScoreNet fixed, the estimated operators exhibited a Pearson correlation of 0.995±0.0030.995\pm0.003, indicating very high consistency. This suggests that the relative coupling patterns of the estimated guidance operators remain largely unchanged despite variations in the Jacobian evaluation states induced by Spherical OU perturbations. In contrast, the correlation across independently trained ScoreNets was 0.677±0.1760.677\pm0.176, indicating that although independent training introduces operator-level variability, a degree of structural consistency is retained across different training runs.

To further examine whether subject-specific information is preserved across independent ScoreNet fits, we performed cross-fit subject identification. Specifically, we compared the operator estimated for each subject using one ScoreNet fit against the operators of 117 subjects estimated using another independently trained fit. Across all 10 pairs of the five independently trained ScoreNets, the Top-1 identification accuracy reached 87.91±16.2687.91\pm16.26%, substantially exceeding the chance-level accuracy of approximately 0.850.85%.

Together, these results indicate that although individual elements of the guidance operator are not estimated identically across independently trained ScoreNets, subject-specific spectral dependency patterns remain substantially identifiable across training runs. We note, however, that these reproducibility measurements do not directly establish the accuracy of the estimated Jacobian relative to the underlying true score Jacobian.

Beyond operator-level reproducibility, we additionally evaluated the effects of independent training initializations on downstream classification performance by repeating the ADNI-AV45 and cross-cohort OOD experiments using five additional random seeds. All experiments maintained the same data splits, training configurations, and evaluation protocols as those reported in the original manuscript.

For AV45-PET, the five additional independent runs achieved an average AUC of 85.76±4.0085.76\pm4.00% and MCC of 0.452±0.0720.452\pm0.072, closely matching the originally reported AUC of 86.7686.76% and MCC of 0.4510.451. Although ACC exhibited relatively greater variability across runs, the average discrimination performance remained comparable to the original results. This variability should be interpreted cautiously given the limited number of AD subjects in the AV45 dataset.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Guide to Author Responses and Manuscript Revisions

We sincerely thank all three reviewers (XtJ2, hpSW, and kmzb) for their careful and constructive feedback. As our individual responses contain extensive experimental results and clarifications, we provide this brief guide to help reviewers locate how their main concerns and questions have been addressed. Detailed evidence and discussion are provided in the corresponding individual responses.

1. Specificity of Score-Derived Guidance and Component Ablations (XtJ2 W1/Q2; hpSW W1/Q1)

We added controlled comparisons with Ledoit–Wolf Precision Guidance and Learned Spectral Affinity under identical experimental settings.

See: XtJ2 Parts 1–2, 4; hpSW Parts 1–2; revised Section 3.3 and Appendix D.1 (Table 8).

2. Contribution of Guidance vs. Follow-Up Information (kmzb W2/R1/R2; XtJ2 Q2)

We clarified that the existing unguided backbone already receives both baseline and follow-up observations. Additional direct-input, SfS_f-only, Sf+WS_f+W, and WW-only experiments further evaluate the discriminative contribution of guidance beyond direct access to longitudinal observations.

See: kmzb Parts 1–3; XtJ2 Part 4; revised Appendix D.2 (Table 9).

3. Jacobian Estimation and Operator Reproducibility (XtJ2 W1/Q1; hpSW W3/Q2)

We clarified the inference-time OU perturbation procedure and evaluated operator stability across independent noise draws and ScoreNet training seeds. Cross-fit subject identification and five additional training-seed evaluations further assess subject-specific structure and downstream classification stability. We distinguish empirical reproducibility from guaranteed Jacobian estimation accuracy.

See: XtJ2 Parts 1, 3–4; hpSW Parts 2–4; revised Appendices B.2 and D.3.

4. Acquisition-Interval Effects and Generalization (XtJ2 W2/Q3)

We conducted additional interval-matched evaluations on FDG-PET, AV45-PET, and OASIS-2 while retaining all AD subjects. We report matched-cohort statistics, classification comparisons, and paired tests.

See: XtJ2 Parts 5–6; revised Appendix D.4.

5. Statistical Reliability (XtJ2 Q3; hpSW W2)

We performed paired statistical tests on interval-matched cohorts and Holm-adjusted AUC comparisons against 20 baselines on ADNI-CTh.

See: XtJ2 Part 6; hpSW Parts 1, 4; revised Section 3.2, Appendix D.4, and Conclusion. The ADNI-CTh Holm-adjusted results are detailed in hpSW Part 4.

6. Motivation and Pairwise Longitudinal Design (kmzb W1)

We revised the Abstract to emphasize learning structured dependencies among longitudinal cortical changes. Sparse and irregular sampling now motivates the pairwise design rather than implying that our method resolves data sparsity. The two-scan inference requirement is explicitly acknowledged.

See: kmzb Part 1; revised Abstract, and Section 5.

7. Additional Clarifications and Suggestions (kmzb Q1–Q3/R3/S1–S2; hpSW W4)

We clarified the Zero-change protocol and diagnostic label definition, added group-level coupling visualizations, discussed single-scan prediction and diagnostic transitions as future directions, and corrected the identified figure and reference issues.

See: kmzb Parts 2–4; hpSW Part 2; revised Figure 1, Section 4, Section 5, and Appendix E (Figure 4).

We hope this guide makes it easier to navigate the detailed responses and revised manuscript. We sincerely appreciate the reviewers' suggestions and the opportunity to strengthen our work.

AuthorsOpenReview ↗

General Rebuttal Summary

We sincerely thank all three reviewers for their thoughtful and constructive feedback. We particularly appreciate their recognition of the methodological contribution, mathematical motivation, and extensive experimental validation of our work.

1. Methodological Novelty and Significance

  • Distinctive methodological contribution: Reviewers recognized our distinctive approach to modeling dependencies among longitudinal cortical changes through score-derived spectral coupling. (XtJ2, hpSW, kmzb)
  • Mathematical grounding: Reviewers highlighted the clear mathematical connection between the score Jacobian and precision structure, alongside inference under alternative diagnostic hypotheses without requiring test labels. (XtJ2, hpSW)
  • Extensive validation: Reviewers positively noted evaluations across three imaging modalities, 20 baselines, and cross-cohort/phase-transfer settings. (XtJ2, hpSW, kmzb)

2. Key Concerns Addressed Through Additional Evidence

  • Guidance contribution: Controlled comparisons with alternative operators and information-matched ablations demonstrated the distinctive discriminative benefit of score-derived guidance beyond longitudinal information alone. (XtJ2, hpSW, kmzb)
  • Reproducibility: Additional noise- and seed-level analyses demonstrated operator stability, preserved subject-specific structure, and reproducible classification performance. (XtJ2, hpSW)
  • Generalization and statistical reliability: Interval-matched evaluations retained performance advantages over selected baselines, while additional paired tests and Holm correction strengthened the statistical evidence. (XtJ2, hpSW)
  • Methodological clarity: We refined the motivation, clarified experimental protocols, and incorporated the requested visualizations and corrections. (kmzb)

We sincerely thank the reviewers and the Area Chair for their time and consideration, and we hope our work will contribute to the ICLR community.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.