acceptodds
Under review as a conference paper at ICLR 2027

DevelopmentODE: Structured Neural ODEs for Early Brain Development Dynamics Across a Decade

Abstract

Understanding how individual brain development unfolds over childhood requires modeling developmental trajectories from sparse longitudinal observations. Long-term neurodevelopmental forecasting is not simply extending short-term prediction to longer horizons as each child is typically observed at only a few irregularly spaced visits and developmental dynamics vary substantially with individuals and age. Generic continuous-time models naturally accommodate irregular timing, but typically absorb these sources of variation into a single flexible transition function, providing little structure for how population-level progression, individual variability, and developmental age should shape the dynamics. We propose DevelopmentODE, a structured continuous-time framework that organizes population and subject-specific variation within a shared developmental geometry while allowing the governing dynamics to evolve with developmental age. The model establishes this geometry around a developmental canal representing the population-level trajectory, whose local direction provides a reference for organizing subject-specific variation. Subject deviation velocities are constrained relative to this developmental direction, while a shared nonlinear deviation field captures individual developmental motion without disrupting population-level progression. Building on the same geometry, DevelopmentODE captures developmental non-stationarity through ordered age-dependent deformations of the shared vector field. These deformations progressively adapt a common underlying dynamical structure as developmental age changes, while elapsed time determines how long the resulting dynamics are integrated. The resulting formulation leverages broadly supported population-level developmental structure to guide learning from sparse individual trajectories, while allowing the governing dynamics to evolve smoothly across developmental age. We evaluate DevelopmentODE on longitudinal fMRI data by predicting future functional connectivity observations of the same child from earlier developmental observations. DevelopmentODE consistently outperforms competing baselines across short- and long-horizon predictions. These results support the benefit of explicitly structuring developmental dynamics for long-horizon neurodevelopmental forecasting.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Neural ODEs, Longitudinal Modeling, Neurodevelopment, Functional Connectivity

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BxwhOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DevelopmentODE, a structured Neural ODE for predicting a child’s future brain functional connectivity. It represents connectivity in a PCA latent space and combines a population developmental trajectory with individual deviation dynamics. Experiments on a longitudinal cohort report modest improvements over some baselines that include direct predictors, sequence models, and previous Neural ODE methods.

Critical strengths
  • The proposed method seems to be well-motivated. Separating population-level development from individual deviations, while allowing dynamics to vary with age, is a reasonable inductive bias that could be beneficial given the longditudal forcasting task the model is trying to solve.

  • The sensitivity analysis in section 4.6 show that the proposed method is reasonably robust to the different hyperparameters and moving parts within the model. These results suggest that the proposed method is relatively robust to these choices and we could expect it to work similarly no matter the choice we make.

Critical weaknesses
  • The improvement of the proposed method in the headline experiment (table 1) over the baselines is modest. The amount of improvement over the best baseline is relatively small, ranging around 1% ~ 5% across different metrics. Due to this it is difficult to know if the supposed improvement of the proposed method is statstically significant, and even if so, the improvement may be too incremental for this work to be considered a practically useful contribution for the given task.

  • The predictive contribution of the complementary-space projection, which is the geometric structure the paper presents as its central contribution, is not clearly isolated from the experiments provided in the paper. The “w/o Canal” ablation removes the population drift and the projection together, so the reported 1.2% MSE degradation cannot establish the projection’s specific contribution. Appendix C.6 provides a useful sensitivity analysis, but its learned α\alpha values (0.15–0.20) retain substantial attenuation of the population-aligned component. The small performance changes therefore show robustness to a modest relaxation, rather than establish whether projection improves prediction. A matched variant retaining the canal drift and stage modulation but fixing α=1\alpha = 1, thereby removing the state/output projections, would directly test the predictive value of the proposed geometric constraint.

Questions for authors
  • Given that the proposed method show relatively small improvements agasint the baselines for table 1, some within the standard deviation margin, could the authors perform statistical tests to see if the improvements are statstically significant? For example, the authors may perform paired significance tests to see if the proposed method is better than the best baseline method.

  • The task has two trivial solutions (which could act as a performance floor for prediction models), where the prediction is (1) identical to the previous timepoint (x^(tb)=x(ta)\hat{x}(t_b) = x(t_a)), or (2) simply the population average for all subjects. For these two cases, what is the performance?

6 comments

Dear Reviewer Bxwh,

We thank you for the detailed and constructive feedback. We have conducted additional experiments addressing the statistical reliability of our forecasting improvements, the contribution of complementary-space geometry, and the performance of simple forecasting baselines.

W1 and Q1. Statistical significance and practical relevance

We agree that improvements over competitive forecasting baselines should be supported by rigorous statistical evidence. We therefore conducted three paired significance tests, together with bootstrap confidence intervals, against two of the strongest learning-based baselines, Time-MLP and MoE-NODE. We additionally examined the consistency of improvements across subjects and prediction horizons.

Table 1. Statistical significance against competitive learning-based baselines.

ComparatorMSE improvementFold t-test pSubject t-test pPermutation pBootstrap 95% CI
Time-MLP3.29%1.14e-43.87e-974.00e-6[0.0009393, 0.0010624]
MoE-NODE2.72%1.14e-49.78e-584.00e-6[0.0007576, 0.0008957]

All comparisons use matched held-out predictions. Fold-level paired t-tests use five matched folds. Subject-level tests aggregate prediction-pair errors, while permutation tests use subject-level paired sign flips. P-values are Holm-corrected over the original four-comparator family. Confidence intervals represent unscaled MSE reductions and are obtained through fold-stratified subject-clustered bootstrap resampling. All three statistical tests confirm significant improvements over both baselines, with bootstrap confidence intervals entirely above zero.

Beyond statistical significance, three additional findings support the relevance of these improvements.

  • The gains are broadly distributed across subjects. DevelopmentODE achieves lower subject-level MSE than Time-MLP and MoE-NODE for 94.53% and 93.23% of held-out subjects, respectively. Thus, the improvements are not driven by a small number of subjects or prediction pairs.

  • The improvements extend across forecasting horizons. Relative to Time-MLP, DevelopmentODE reduces MSE by 2.92% on short-term and 3.57% on long-term predictions. Relative to MoE-NODE, the improvements are 4.67% and 1.37%, respectively. All four comparisons are significant under Holm-corrected fold-level paired tests (p < 0.001).

  • The task addresses a scientifically important and data-limited problem. Longitudinal neuroimaging studies have documented the difficulty of modeling developmental trajectories from sparse observations [1], heterogeneous age-dependent FC changes [2], and measurement variability in functional connectivity [3].

These characteristics motivate approaches that combine broadly supported population developmental information with limited individual longitudinal supervision. DevelopmentODE explicitly organizes population progression, shared deviation dynamics, and age-dependent adaptation within a continuous-time framework.

W2. The contribution of complementary-space geometry

We appreciate your observation that the original w/o Canal ablation jointly removes the population developmental reference and its associated geometric constraints. Following your suggestion, we performed a strictly matched fixed α = 1 experiment, retaining the population canal drift, shared deviation field, and ordered age-dependent adaptation while removing both complementary-space projections.

Table 2. Matched sensitivity analysis of the geometric constraint.

ConfigurationαAll MSE (×100)Adjacent MSE (×100)
Strict complementary02.9403 ± 0.03262.9451 ± 0.0262
Learnable alignedLearned2.9390 ± 0.03212.9428 ± 0.0266
Fully aligned12.9395 ± 0.03232.9444 ± 0.0260

Together with the learnable α experiment in the Appendix, the results demonstrate that forecasting performance remains highly stable across substantially different geometric constraints. Moving from strict complementary dynamics (α = 0) to fully retained population-aligned dynamics (α = 1) changes all-forward MSE by only a little bit. This finding further clarifies that the principal value of complementary-space geometry is not an isolated reduction in prediction error, but the explicit organization and interpretability it provides for developmental dynamics.

We highlight two contributions of this geometric formulation.

  • Explicit organization of population and individual dynamics. The population canal defines a local developmental direction through its tangent. Complementary projection provides an explicit mechanism for separating population-aligned progression from deviation motion. This constrains the degrees of freedom available to the individual deviation field and reduces its ability to freely re-explain population-level progression. Consequently, the model incorporates an explicit population–individual organization. Importantly, the α-relaxation experiment demonstrates that this organization can accommodate different assumptions about population-aligned individual motion without substantially affecting forecasting accuracy.

  • Interpretable and empirically analyzable developmental variation. The geometric formulation also enables direct examination of developmental quantities that would otherwise remain entangled within an unconstrained vector field. In the Appendix, we analyze the population canal velocity, individual deviation magnitude, and the population-aligned and complementary components of observed longitudinal changes. For example, the population-aligned fraction of observed displacement varies substantially across developmental intervals, from 53.0% for the neonatal-to-1-year transition to 13.8% for the 6-to-10-year transition.

These analyses demonstrate how the proposed geometry provides a common reference for characterizing age-dependent population progression and individual variation. This modeling principle is consistent with scientific machine learning approaches that incorporate domain-motivated structure into flexible differential-equation models [4], and is biologically motivated by developmental canalization [5].

The structural ablations further support the broader formulation:

  • Removing the population-referenced canal formulation increases MSE by 1.22%.
  • Removing ordered age-dependent adaptation increases MSE by 1.97%.

We revised the manuscript to distinguish the role of complementary geometry as a structural inductive bias from the isolated predictive effect of strict orthogonality, and to make the corresponding analyses more prominent.

Q2. Identity and population-average prediction baselines

We thank the reviewer for suggesting these important reference baselines. We evaluated the simple alternatives:

  • Identity: directly carries forward the observed source FC, Xi(tb)=Xi(ta).
  • Population average predicts the mean FC computed from the population average for all subjects.

Table 3. Comparison with Identity and population-mean baselines.

ComparatorMSE reduction by OursFold t-test pSubject t-test pPermutation pBootstrap 95% CI
Identity49.79%7.15e-92.32e-3015.00e-6[0.028734, 0.029581]
population average3.73%8.22e-53.78e-1004.00e-6[0.0010675, 0.0012088]

The reported p-values are Holm-corrected; the Identity results use the expanded five-comparator family, whereas the population-mean results use the original four-comparator family. Bootstrap intervals report unscaled paired MSE reductions and are conditional on the evaluated checkpoints. DevelopmentODE substantially outperforms Identity and the population average.

This demonstrates that simply retaining the previous FC observation or population average is insufficient for multi-year developmental forecasting. The population average is a stronger baseline than Identity, consistent with the substantial population-level developmental regularity observed in longitudinal neuroimaging studies [1, 2].

We hope these clarifications and additional findings address your concerns. Please let us know if you have any further questions.

References

[1] Modeling Sparse Longitudinal Data in Early Neurodevelopment.

[2] Longitudinal Developmental Trajectories of Functional Connectivity Reveal Regional Distribution of Distinct Age Effects in Infancy.

[3] A Decade of Test-Retest Reliability of Functional Connectivity: A Systematic Review and Meta-Analysis.

[4] Universal Differential Equations for Scientific Machine Learning.

[5] Waddington's Canalization Revisited: Developmental Stability and Evolution.

Reviewer_rWeXOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DevelopmentODE, a structured continuous-time framework that organizes population and subject-specific variation within a shared developmental geometry while allowing the governing dynamics to evolve with developmental age.

Critical strengths
  1. The paper addresses a meaningful and challenging problem of long-term brain development forecasting from sparse and irregular longitudinal fMRI observations.
Critical weaknesses
  1. DevelopmentODE is deterministic and trained with an MSE loss, which may encourage predictions toward average developmental trajectories. Given that similar initial FC states may lead to different developmental outcomes, how does the model handle this ambiguity?

  2. Two important ablation studies are missing: a) Population canal vs. deviation dynamics: whether the model truly distinguishes population-level progression from individual variation. b) Shared deviation field vs. age-dependent adaptation: whether the model separates age-shared dynamics from age-specific corrections.

  3. Although the evaluation metrics are computed for individual prediction pairs, the reported results are aggregated across subjects and do not directly demonstrate that the model captures personalized developmental trajectories. Besides, the improvements are limited.

Questions for authors
  1. Regarding to weakness 1, a baseline using the population-mean FC at the target age, computed from training subjects, would help assess whether the model learns meaningful subject-specific dynamics.
  2. Regarding Weakness 2, could the authors provide additional ablations?
  3. Regarding Weakness 3, could the authors provide subject-level trajectory visualizations or quantitative analyses to demonstrate that DevelopmentODE captures individual developmental patterns rather than primarily following population-average trajectories?
  4. The population canal is learned using an MLP initialized from age-specific population means. Since its primary role is to provide a smooth and differentiable population trajectory, could a fixed interpolation of population means achieve similar performance? Such an approach may also be more interpretable, less prone to overfitting, and reduce potential compensation between population and individual dynamics.
7 comments

Dear Reviewer rWeX,

We thank you for the thoughtful assessment and for recognizing the importance of forecasting long-term brain development from sparse longitudinal observations. We have conducted additional population-baseline comparisons, structural ablations, and subject-level analyses to address these concerns. We also wish to clarify the biological motivation and methodological Informativeness underlying DevelopmentODE.

Biological motivation and methodological Informativeness

DevelopmentODE is motivated by two complementary characteristics of early neurodevelopment.

  • Population-level developmental regularity coexists with substantial individual variability. Inspired by developmental canalization [1], we introduce an explicit population developmental reference and organize subject-conditioned dynamics relative to this shared trajectory.

  • Developmental dynamics are non-stationary across age. Longitudinal neuroimaging studies document heterogeneous developmental trajectories and different age-dependent patterns of functional connectivity maturation [2,3].

The methodological novelty lies in how these developmental characteristics are organized within continuous-time dynamics. Generic and age-conditioned Neural ODEs typically absorb population progression, individual variation, and age dependence into a flexible transition function. DevelopmentODE instead introduces three structured components:

  • A population developmental canal providing a shared, differentiable developmental reference.
  • A nonlinear deviation field pooling individual developmental transitions across subjects and ages.
  • Ordered age-dependent deformations continuously adapting the shared dynamics without learning independent vector fields for each developmental stage.

W1, W3, Q1, and Q3. Population-average behavior and subject-specific forecasting

We appreciate the concern that deterministic MSE optimization may encourage predictions toward population-average developmental trajectories.

Importantly, the optimal deterministic predictor under squared error is the conditional expectation given the available information, not necessarily the population mean at the target age: $ \hat X_i^*(t_b)= \mathbb E[X_i(t_b)\mid X_i(t_a),t_a,t_b]. $ DevelopmentODE incorporates both developmental-age information and the individual's observed FC to learn how these information sources jointly inform future-state prediction.

To directly address your concern about subject-specific forecasting, we have additional experiments, and our additional analyses establish three findings:

  • Beyond population-average prediction: DevelopmentODE significantly outperforms both overall and target-age population-mean baselines, demonstrating additional predictive value beyond shared developmental progression.
  • Broad subject-level benefits: Improvements are distributed across individuals rather than driven by a few prediction pairs. DevelopmentODE improves forecasting for 78.65% of subjects relative to the exact-age mean and more than 93% relative to two competitive learning-based baselines.
  • Effective utilization of individual information: Replacing the correct source observation with an age-matched observation from another subject worsens predictions, with the correct input benefiting 74.99% of subjects.

1.Population-mean baseline comparisons. We evaluated the overall training-set mean and exact-target-age training mean baselines.

Table 1. Improvements over population-mean forecasting baselines and paired statistical significance.

ComparatorR² improvementFold t-test pSubject t-test pPermutation pBootstrap 95% CI
Overall training mean+7.22%8.22e-53.78e-1004.00e-6[0.0010675, 0.0012088]
Exact-age training mean+0.65%3.98e-51.84e-304.00e-6[0.0000868, 0.0001078]

All three paired tests consistently support significant improvements over both population-mean baselines, with bootstrap confidence intervals entirely above zero. The age-conditioned population means are a particularly strong baseline because shared developmental progression explains an important part of future FC structure. Nevertheless, DevelopmentODE consistently improves upon this population-only reference with statistical significance.

Implementation Details: R² improvements are relative changes over each comparator. Statistical tests and confidence intervals evaluate paired MSE differences, defined as comparator MSE minus DevelopmentODE MSE. Fold-level paired t-tests use five matched folds. Subject-level tests aggregate prediction-pair errors within each subject, and permutation tests use subject-level paired sign flips. P-values are Holm-corrected over the original four-comparator family. Bootstrap intervals report unscaled MSE reductions using fold-stratified subject-clustered resampling.

2. Subject-level forecasting performance. We conducted supplementary analyses over held-out subjects, aggregating prediction errors within each subject.

Table 2. Subject-level forecasting improvements.

ComparatorSubjects improved
Time-MLP94.53%
MoE-NODE93.23%
Exact-age Mean78.65%

These results demonstrate that DevelopmentODE improves forecasting for the great majority of held-out subjects compared to baselines. In particular, more than 93% of subjects benefit relative to the two competitive learning-based baselines. This provides direct subject-level evidence that the forecasting improvements are broadly distributed.

3.Predictive contribution of individual source information. We further investigated whether the model utilizes subject-specific information beyond developmental age. Using the same trained model, we compared predictions from the correct individual observation against predictions obtained by replacing the source FC with either the same-age population means or another age-matched subject's FC.

Table 3. Predictive contribution of correctly matched source observations.

Input comparisonMSE improvement from correct sourceSubjects benefiting
Correct source vs. same-age mean0.102%65.60%
Correct source vs. age-matched subject swap0.154%74.99%

The corresponding subject-level bootstrap 95% confidence intervals are [0.067%, 0.139%] and [0.119%, 0.192%], respectively. These controlled comparisons demonstrate that correctly matched individual observations provide additional predictive information beyond the population developmental clock. When developmental ages are preserved, but individual source information is replaced, forecasting performance consistently deteriorates.

W2 and Q2. Population canal, shared deviation field, and age-dependent adaptation

We appreciate the suggestion to examine whether the population reference, shared deviation dynamics, and age-dependent corrections provide distinct predictive contributions.

The submitted manuscript already contains ablations for the population-referenced formulation and age-dependent adaptation. We additionally evaluated a w/o Shared Field configuration.

Table 4. Structural contributions to forecasting performance.

VariantMSE (×100) ↓Relative MSE degradation
Full DevelopmentODE2.9390—
w/o Canal2.9750+1.22%
w/o Stage2.9968+1.97%
w/o Shared Field2.9470+0.27%

The three ablations examine different aspects of the proposed developmental structure:

  • Population-referenced modeling: Removing the canal formulation increases MSE by 1.22%, supporting the benefit of an explicit population developmental reference.
  • Age-dependent adaptation: Removing the stage corrections increases MSE by 1.97%, demonstrating that a shared deviation field alone does not achieve the performance of the full model.
  • Shared deviation dynamics: Removing the shared nonlinear field increases MSE by 0.27%, providing complementary evidence for the contribution of age-shared dynamics.

The larger degradation from removing stage adaptation is consistent with a central motivation of DevelopmentODE: developmental transformations can change substantially with absolute age, making a stationary shared vector field insufficient. The shared field and ordered stage corrections have different modeling roles. The former pools information across individuals and ages, while the latter adapts the shared dynamics to developmental non-stationarity without introducing fully independent stage-specific networks.

Q4. Population interpolation versus learned developmental dynamics

We appreciate the suggestion to compare the learned developmental formulation with simpler interpolation-based alternatives. We therefore evaluated an additional interpolation-based forecasting baseline compared with DevelopmentODE.

Table 5. DevelopmentODE versus interpolation.

ModelMSE (×100) ↓R² (×100) ↑MAE (×10) ↓PCC (×100) ↑
Interpolation baseline3.0257 ± 0.036834.324 ± 0.6791.2237 ± 0.005659.087 ± 0.518
DevelopmentODE2.9390 ± 0.032336.156 ± 0.5171.2009 ± 0.004660.514 ± 0.386

DevelopmentODE reduces MSE by 2.87% and improves R² by 5.34% relative to interpolation. These results demonstrate that the interpolation-based alternative does not achieve the forecasting performance of DevelopmentODE. Importantly, the population canal is not used solely to interpolate expected FC values across ages. Its derivative enters the governing dynamics, while subject-conditioned deviation fields model how individual states evolve relative to the population reference.

We thank the reviewer for motivating these additional analyses and clarifications. Please let us know if you have any further questions.

References

[1] Waddington's Canalization Revisited: Developmental Stability and Evolution.

[2] Longitudinal Developmental Trajectories of Functional Connectivity Reveal Regional Distribution of Distinct Age Effects in Infancy.

[3] Modeling Sparse Longitudinal Data in Early Neurodevelopment.

[4] Universal Differential Equations for Scientific Machine Learning.

Reviewer_W4WYOpenReview ↗
Rating
3(+1)
Confidence
4(+3)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces a framework, termed as DevelopmentODE, for forecasting longitudinal functional connectivity (FC) of early childhood brain development. The authors developed a network that is expected to model population-level developmental trajectories and individual subject-level variability in an encoded latent space. The proposed DevelopmentODE also considers the challenges of sparse and irregularly sampled observations by accounting for age-dependent dynamics. Experiments are validated on longitudinal resting-state fMRI data from 484 children (with at least two time points available) in the Early Brain Development Study (EBDS), spanning ages 1–10 years, and compare it with existing time-series models for longitudinal prediction. The results show that incorporating population-level developmental structure and age-dependent dynamics can improve longitudinal FC forecasting.

Critical strengths
  • The paper addresses the challenging task of long-term brain developmental forecasting from sparse and irregularly sampled longitudinal neuroimaging fMRI data.

  • The proposed method accounts for age-dependent developmental changes by introducing smoothly weighted, stage-specific corrections to a shared neural vector field. This also considers the realistic situation of sparse and irregularly collected fMRI data.

  • The paper is well organized and generally easy for readers to follow.

Critical weaknesses
  • Child brain functional networks derived from fMRI exhibit highly nonlinear developmental changes, with potentially rapid transitions during specific developmental stages. The assumption that whole-brain developmental dynamics can be linearly decomposed into population-level development and subject-specific variation may not be appropriate, as it does not seem to account for the highly nonlinear underlying geometry.
  • In addition, the proposed geometric decomposition projects the shared neural field onto the space orthogonal to the population developmental trajectory. However, individuals may develop faster or slower while following the same population trajectory. Ignoring these parallel velocity deviations may limit the model’s ability to capture the real individual differences in developmental rates.
  • The derivation from Eq. (6) to Eq. (7) is unclear and appears to involve approximations. This is important because Eq. (7) seems to rely heavily on network-based parameterization learned from the training data. As a result, the interpretability and modeling rationale of DevelopmentODE may be obscured by the learned network. The authors should clarify the derivation and explain how the proposed modeling assumptions are preserved in the final implementation.
  • The key Fig. 1 should be simplified and clarified. Several mathematical expressions and notations are introduced before being defined, making the figure difficult to follow.
  • The population-level and subject-level effects in Eq. (9) may be entangled and compensate for one another. The weighting parameter dd appears to play a critical role, but its definition and influence on the learned dynamics are not sufficiently explained.
  • The long-term prediction results (Tab. 1) show merely improved performance over several competing methods. It is difficult to justify if the proposed method can effectively addresses sparse and missing longitudinal data with statistically significant gains.
Questions for authors
  • Why is it appropriate to impose ODE-based dynamics in the latent space? Do the learned latent trajectories correspond to meaningful developmental dynamics when mapped back to the fMRI FC space through the decoder?
  • The derivation from Eq. (6) to Eq. (7) is unclear and appears to involve approximation steps before the full network FΘF_\Theta is introduced. Could the authors clarify this derivation?
  • How are the weighting parameters in Eq. (9), particularly dd, defined or selected?
  • Of the 484 subjects in the dataset, how many have more than two time points? This information is important for assessing the amount of longitudinal data available to learn individual-specific developmental dynamics.
  • How are the baseline methods selected? Do they account for irregular age intervals and missing scans, and are they evaluated under the same observation patterns as DevelopmentODE?
  • The comparison may be missing recent generative approaches for predicting longitudinal or 4D brain fMRI dynamics. Examples include but not limited to [1]. Could the authors explain why this line of research methods were not included in the comparison?
    • [1] BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics.
5 comments

Dear Reviewer W4WY,

We thank you for the constructive feedback and for recognizing the importance of forecasting early brain development from sparse and irregular longitudinal observations. We provide the following clarifications for your concerns.

W1. Nonlinear developmental dynamics and population–individual decomposition

We agree that early brain development exhibits highly nonlinear and heterogeneous changes, including potentially rapid transitions across developmental stages. This is precisely one of the central challenges motivating DevelopmentODE: learning nonlinear, age-dependent developmental dynamics from sparse individual longitudinal observations.

Our framework addresses this challenge through three complementary components:

  • Nonlinear population progression. The developmental canal (t) is represented by a nonlinear neural function, allowing the shared developmental trajectory to change nonlinearly with age.
  • Nonlinear individual dynamics. The shared deviation field f models state-dependent individual variation through nonlinear dynamics, rather than assuming fixed or linear residual trajectories.
  • Developmental non-stationarity. Ordered age-dependent deformations continuously adapt the governing dynamics across developmental stages while preserving shared structure across sparsely observed ages.

Importantly, the decomposition zi(t)=(t)+ri(t) is an exact representation of a subject's latent state relative to a population reference, not an assumption that the underlying dynamics are linear. This design is motivated by longitudinal neurodevelopmental studies documenting heterogeneous and nonlinear age-dependent FC changes [1, 2]. DevelopmentODE explicitly incorporates them into a structured continuous-time formulation.

W2. Individual developmental rates and population-aligned velocity

We appreciate the concern that individuals may develop at different rates along a shared population trajectory. This issue is specifically examined in Appendix C.6, "Sensitivity to a Learnable Population-Aligned Component," where we introduce a parameterized relaxation of the complementary-space constraint. This formulation allows individual deviation dynamics to retain a population-aligned velocity component. We have explained more clearly in the main text how this experiment relates to individual developmental-rate variation and provided an explicit reference to Appendix C.6.

W3 and Q2. Derivation from Eq. (6) to Eq. (7)

We apologize for the insufficient explanation. Eq. (7) follows directly by integrating the differential equation in Eq. (6) over the developmental interval [ta,tb]:

z^i(tb)=zi(ta)+∫tatbFΘ(zi(t),t) dt,\hat z_i(t_b)=z_i(t_a)+\int_{t_a}^{t_b}F_\Theta(z_i(t),t)\,dt,

where F denotes the complete right-hand side of Eq. (6), including the population canal derivative, shared nonlinear deviation field, and age-dependent corrections. There is no additional approximation in this mathematical transition. Numerical approximation occurs only when evaluating the integral using an ODE solver. Our default implementation uses RK4 with normalized step size 0.02.

W4. Clarity of Figure 1

We agree that Figure 1 introduces excessive mathematical notation before explaining the corresponding concepts. We remove unnecessary expressions, improve the order of definitions, and ensure consistent notation with Section 3. We also add a dedicated notation and parameter table to the appendix, covering the major variables, learnable parameters, and their roles in the model.

W5 and Q3. Population–individual compensation and parameter sensitivity

The parameter d in Eq. (9) denotes the latent dimensionality. The factor 1/d normalizes the population-centering loss rather than directly weighting population and individual velocities. We have already examined the influence of d in Section 4.6 and Figure 4(b), considering d={8,16,32,64}. The results show relatively stable forecasting performance across these dimensions. We clarify these parameter definitions and their influence on the learned dynamics in the revised manuscript and appendix parameter table.

W6. Statistical significance and sparse longitudinal forecasting

We agree that improvements over competitive baselines should be supported by formal statistical evidence. We therefore conducted three complementary paired significance tests against two strong learning-based baselines, Time-MLP and MoE-NODE, using identical held-out prediction pairs.

Table 1. Statistical significance of all-forward forecasting improvements.

ComparatorMSE improvementFold paired t-test pSubject paired t-test pPermutation pBootstrap 95% CI
Time-MLP3.29%1.14e-43.87e-974.00e-6[0.0009393, 0.0010624]
MoE-NODE2.72%1.14e-49.78e-584.00e-6[0.0007576, 0.0008957]

All three statistical tests support significant improvements over both baselines, with bootstrap confidence intervals entirely above zero.

Additional evidence further supports the reliability of the improvements:

  • Consistency across prediction horizons. DevelopmentODE improves MSE relative to Time-MLP by 2.92% on short-term and 3.57% on long-term predictions. The corresponding improvements relative to MoE-NODE are 4.67% and 1.37%. All four comparisons are significant under Holm-corrected fold-level tests (p<0.001).
  • Broad subject-level benefits. DevelopmentODE achieves lower subject-level MSE than Time-MLP and MoE-NODE for 94.53% and 93.23% of held-out subjects, respectively.
  • Robustness to sparse supervision. Section 4.4 separately investigates reductions in population-level and longitudinal supervision. DevelopmentODE maintains its advantage over representative baselines under reduced longitudinal supervision, directly addressing the sparse-data setting motivating the framework.

Implementation Details: Confidence intervals report unscaled MSE reductions. Fold-level paired tests use five matched folds. Subject-level tests aggregate prediction errors, and permutation tests use subject-level paired sign flips. P-values are Holm-corrected over the original four-comparator family. Bootstrap intervals use fold-stratified subject-clustered resampling.

Q1. Why use latent ODE dynamics, and are the trajectories meaningful in FC space?

An ODE provides a natural model of continuous developmental evolution: the observed initial state determines a trajectory, while the learned vector field describes how that state evolves with developmental age. Unlike a direct endpoint predictor, this formulation defines a consistent continuous-time transition process, including for irregular and previously unobserved prediction intervals. Neural ODEs and Latent ODEs provide established foundations for learning continuous-time state evolution from irregularly sampled observations [3, 4].

In DevelopmentODE, the vector field additionally depends on absolute developmental age, allowing different developmental periods to follow different dynamics. The predicted trajectories are decoded into FC space, where both the forecasting objective and evaluation metrics are computed. Thus, the learned dynamics are directly evaluated through future FC prediction rather than only latent-space reconstruction. We interpret these trajectories as predictive representations of developmental change, without claiming that every intermediate state corresponds to an independently verified biological mechanism.

Q4. Number of subjects with more than two observations

Among the 484 eligible subjects, 277 (57.2%) have more than two longitudinal visits, while 207 have exactly two. The median is three visits per subject, with a range of 2–7.

Q5. Baseline selection and evaluation fairness

All learning-based baselines use the same subject-disjoint splits, fold-specific representations, source–target prediction pairs, FC-space objectives, and validation-based checkpoint selection.. All methods are evaluated on identical observed longitudinal pairs without access to additional intermediate scans.

Q6. Comparison with recent generative approaches

We thank the reviewer for pointing out recent generative approaches. We will expand our related-work discussion to include BrainWorld [5] and other related work such as diffusion transformer [6].

BrainWorld uses structural MRI as subject-level anatomical context to generate 4D fMRI dynamics. The diffusion-transformer approach generates voxelwise 4D fMRI sequences conditioned on cognitive tasks. Both studies investigate short-timescale spatiotemporal fMRI generation. In contrast, DevelopmentODE predicts future functional connectivity observations of the same child across multi-year developmental intervals, conditioned on an earlier FC observation and developmental ages. The approaches therefore differ in prediction targets, temporal scales, and conditioning information. Direct comparisons would require substantial adaptation to the same subject-specific longitudinal FC forecasting protocol. We will explicitly discuss these generative approaches as complementary directions and clarify why they are not directly comparable under the current task formulation.

We thank the reviewer again for the constructive suggestions. Please let us know if you have any further questions.

References

[1] Modeling Sparse Longitudinal Data in Early Neurodevelopment.

[2] Longitudinal Developmental Trajectories of Functional Connectivity Reveal Regional Distribution of Distinct Age Effects in Infancy.

[3] Neural Ordinary Differential Equations.

[4] Latent Ordinary Differential Equations for Irregularly-Sampled Time Series.

[5] BrainWorld: A Structural-Prior-Conditioned Generative Model for Whole-Brain 4D fMRI Dynamics.

[6] Scalable Diffusion Transformer for Conditional 4D fMRI Synthesis.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

Thank you for your thoughtful feedback and for the time you devoted to our paper. Your questions helped us clarify the contribution of DevelopmentODE and test how reliably it forecasts a child’s future functional connectivity from earlier scans. Below, we summarize our contribution, the strengths highlighted by the reviewers, and what the additional analyses established.

Task and methodological contribution

  • DevelopmentODE forecasts within-child changes across irregular intervals of up to a decade.
  • The tangent to a learned population trajectory defines a local developmental direction that constrains individual deviation velocities.
  • Ordered age-dependent deformations adapt a shared vector field across developmental stages. Absolute age shapes the dynamics, while elapsed time determines how long they act.

Strengths recognized by the reviewers

  • W4WY and rWeX highlighted the importance and difficulty of forecasting development from sparse, irregular scans.
  • W4WY recognized the smooth stage-specific corrections as a strength.
  • Bxwh found the population–individual formulation well motivated and noted the model’s robustness to hyperparameter choices.

What the rebuttal established

  • The forecasting gains are statistically reliable. Paired tests across folds and subjects, subject-level permutation tests, and bootstrap intervals consistently supported improvements over competitive baselines.
  • DevelopmentODE effectively uses subject-specific information. With age held fixed, replacing the child’s own scan with an age-matched peer’s scan or the same-age population mean modestly reduced forecasting accuracy.
  • Ablations support the proposed developmental structure. Removing the canal formulation, shared deviation field, or age-dependent adaptation each increased error.
  • Additional comparisons and clarifications. DevelopmentODE also outperformed interpolation of population means. We clarified the ODE derivation, the baseline evaluation protocol, and how multi-year forecasting differs from short-timescale fMRI generation.

Thank you for your consideration.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.