acceptodds
Under review as a conference paper at ICLR 2027

ClockBridge: Partial Identification of Cross-Modal Disease Timing across Unpaired Cohorts

Abstract

Biomedical cohorts often measure different modalities in different individuals and at different disease stages, so stage labels that appear comparable can still admit multiple incompatible cross-modal timelines. We formulate cross-modal disease timing across unpaired cohorts as a partial identification problem. ClockBridge places each cohort on its own disease clock, retains the synchronizations compatible with the observed stage-resolved evidence in a synchronization equivalence set (SES), and reports a temporal direction only when it is invariant over the resulting identified interval. Otherwise the claim remains unresolved. On hidden-pair synchronization, ClockBridge substantially lowers selective false-order risk relative to shared-protocol forced-map references and a same-loss point estimator. FO@40 is 12.3% on ADNI versus 52.2–53.5% for forced maps and 17.7% for Point-L, and 3.3% on PD versus 20.8–49.6% and 7.2%, respectively, while decision coverage at low error remains higher. Under degraded stage evidence, unresolved claims increase and unstable directions are withheld. On fully unpaired molecular and imaging cohorts, ClockBridge yields lower aggregate contradiction under cohort replacement and abstains more often on unstable claims. These results support set-valued identification as a principled alternative to forcing a single cross-modal disease timeline when the available stage evidence does not uniquely determine one.

Primary area:
applications to neuroscience & cognitive science
Keywords:
unpaired multimodal learning, disease progression, temporal synchronization, partial identification, selective prediction

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Zn3XOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The authors formulate cross-modal disease timing across unpaired cohorts as a partial identification problem and introduce ClockBridge, a set-valued alignment framework that reports temporal orderings only when invariant across stage-compatible alignments (SES) and abstains otherwise. On ADNI, PD, and unpaired molecular–imaging cohorts, ClockBridge lowers false-order risk and cross-cohort claim contradictions compared to point estimators and forced single-timeline baselines.

Critical strengths
  1. Motivation: Rightly challenges the assumption of a shared cross-cohort disease timeline.
  2. Uncertainty Handling: Abstaining under ambiguous stage evidence is methodologically sound.
Critical weaknesses
  1. Unclear presentation and logical flow. The manuscript does not clearly connect the biomedical question, modeling assumptions, method, and evaluation. Several concepts are introduced without sufficiently clear definitions or explanations. For example, the manuscript does not clearly specify what constitutes a "witness," how "bidirectional capacity" is calculated, or what criteria distinguish "not-evaluable" from "unresolved" outcomes. These make it difficult to follow the method and understand how the available evidence supports the conclusions.

  2. Insufficient justification for temporal identification. The paper does not adequately explain why its alignment loss captures meaningful temporal correspondence between modalities, or why the SES should contain the true correspondence. Thus, alignment consistency does not guarantee biological correctness.

  3. Insufficient and unclear validation. Under the translation model h(u)=u+sh(u)=u+s (Sec. 3.2, Eq. 1), the common shift cancels from the predicted pairwise differences used for ordering (Appendix A.3). It is therefore unclear how the ordering metrics in Table 1 assess synchronization accuracy. Appendix A.3 also introduces "true local coordinates" without clearly specifying how they are obtained, while Sec. 4.1 describes PD pairing through "subject co-enrollment" without establishing temporal correspondence between scans. The external experiments assess cohort-replacement stability rather than temporal correctness (Sec. 4.4). Finally, Table A6 reports top-20/25/30 decided claims per replacement block, although Table A5 lists only 19 decided ClockBridge claims for molecular replacement and 28 for imaging replacement. This discrepancy requires clarification.

  4. Unclear source of improvement. The experiments do not show whether the gains come from the SES itself or from confidence scoring and abstaining on difficult cases. Comparisons with uncertainty-aware alignment methods are also insufficient. In the acquisition study, it remains unclear what new data are collected and how they update the timing conclusions.

Questions for authors
  1. How does Equation (3) compare distinct feature spaces without absorbing true biological latency?
  2. What estimand and assumptions define the identified set? When does the empirical ε\varepsilon-sublevel set contain the true synchronization, and how is ε\varepsilon calibrated?
  3. How can pairwise ordering assess a common translation when the offset cancels?
  4. How are 20, 25, or 30 decided claims selected per block when Table A5 reports only 19 or 28?
  5. How are "true local coordinates" obtained? What temporal reference does PD co-enrollment provide when scan dates differ?
  6. Do improvements persist with comparable uncertainty estimation and abstention for baselines, evaluated on a common test population?
  7. What does each acquisition action observe, and how are its outcomes, costs, and evidence updates defined?
Reviewer_kRWQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes ClockBridge, a framework for identifying temporal relationships across unpaired biomedical cohorts. Instead of forcing a single alignment between molecular and imaging data, it considers multiple plausible synchronizations and reports temporal ordering only when the evidence consistently supports a direction. Experiments on Alzheimer's and Parkinson's disease datasets demonstrate reduced false-order risk and improved robustness under uncertain stage information.

Critical strengths

• The paper addresses an underexplored problem of temporal alignment across unpaired biomedical cohorts.

• The set-valued formulation provides an interpretable way to represent uncertainty in cross-modal disease timing. Rather than forcing a single temporal alignment, the method considers multiple plausible synchronizations and reports directional claims only when they are consistently supported.

• The experiments include multiple datasets, robustness analyses, and comparisons under matched decision coverage.

Critical weaknesses

• The methodological novelty is somewhat limited, as the framework largely builds on existing partial-identification and selective-prediction principles. The proposed synchronization equivalence set is constructed using a loss-based threshold, which resembles conventional set-valued estimation. The paper could better clarify what fundamentally distinguishes its approach from existing uncertainty-aware prediction frameworks.

• Much of the reported improvement appears to arise from abstention rather than improved synchronization accuracy. In particular, ClockBridge and Point-L* achieve identical MAE values, suggesting that the underlying point estimation is not improved.

• The translation-based synchronization model imposes restrictive assumptions about disease progression across modalities. In practice, molecular and imaging biomarkers may evolve at different rates, making a simple temporal shift insufficient to capture their relationships.

• The external experiments evaluate consistency under cohort replacement rather than biological correctness of the inferred temporal ordering. Although consistent predictions across cohorts are desirable, they do not necessarily indicate that the estimated temporal relationships reflect actual disease progression. Additional validation using biologically established event sequences or longitudinal multimodal data would strengthen the conclusions.

• The statistical validity of the synchronization equivalence set is not theoretically established. Its construction depends on a predefined loss threshold and synchronization family, but there is no formal guarantee that the resulting set contains the true synchronization.

• The active acquisition component is evaluated using a simplified surrogate rather than realistic biomedical data.

Questions for authors

• How does the proposed SES framework fundamentally differ from existing set-valued estimation and selective prediction methods?

• Would ClockBridge still outperform Point-L* equipped with conventional uncertainty-based abstention under matched decision coverage?

• How realistic is the translation-only synchronization assumption when molecular and imaging biomarkers may evolve at different rates?

• Given the high abstention rates, especially under more flexible synchronization families, how useful are the resulting temporal claims for real biomedical research?

• How can the inferred temporal ordering be validated biologically when the external unpaired cohorts lack ground-truth cross-modal event timing?

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

The paper examines how to infer disease-event order across biomedical data modalities such as images and molecular data, when the same disease-stage labels allow different ways of aligning the cohorts. The authors propose using stage labels to construct a disease timeline for each cohort and consider different time shifts between these timelines, retaining all shifts whose stage-matching loss is at most ε above the minimum. These retained alignments form the synchronization equivalence set. The method reports an event order only when all alignments in this set agree, otherwise the claim remains unresolved. An optional module selects further analyses from a predefined set to help resolve these claims. Evaluation includes imaging tests with pairings hidden from the method (ADNI fMRI and structural MRI; PD FDG and CFT PET) and ablations on synthetic data. It also examines whether molecular–imaging conclusions remain consistent when cohorts are replaced (ROSMAP/SEA-AD with ADNI2/ADNI3). The authors report a lower risk of incorrect ordering compared to methods that select a single alignment and fewer contradictions under cohort replacement.

Critical strengths

Relevant problem and good framing: Alzheimer’s research increasingly draws conclusions about event timing from cohorts that share disease-stage labels but include different participants. Methods that force a single alignment can report an event order that the data do not support. ClockBridge reports which of two events occurs first only when all well-fitting alignments agree, and the authors explicitly avoid claiming to recover a unique biological or causal timeline.

Critical weaknesses
  1. Validation that tests neither the target setting nor the published methods. Accuracy is measured only on imaging–imaging pairs from the same participants (ADNI fMRI vs structural MRI; PD FDG vs CFT PET; Sec. 4.1, App. B.1). The molecular–imaging setting the paper targets is tested only for consistency under cohort replacement (Sec. 4.4). This is not a proxy for accuracy, since a method can be consistently wrong and consistency under cohort replacement alone does not establish that ClockBridge’s event orders are correct.. The validation is further weakened by the fact that the baselines are “intentionally simplified protocol adaptations” (App. A.3) of the published methods. On ADNI they are wrong on more than half of their most confident predictions (FO@40 52–54%), yet they supply the headline comparison in the abstract. The paper therefore does not show that ClockBridge’s event orders are accurate in the setting it is designed for, even though such a test appears feasible, for example with simulated data or with ADNI’s CSF biomarkers measured in the same participants as the scans.
  2. Neither a new method nor a new finding. The paper does not propose a methodological advance. Its core method is a direct application of partial identification and selective prediction, both of which the authors cite. Nor does it add theoretical support: no result shows that the retained set contains the true alignment with any guaranteed probability, which the authors also acknowledge (Sec. 3.4). Finally, the paper provides no new insight into the disease it studies. Although it is motivated by cross-modal timing claims in Alzheimer’s disease, it reports no specific event order that ClockBridge supports or refutes, only aggregate shares of supported, contradicted and abstained claims.
Questions for authors
  1. Can you report accuracy in the molecular–imaging setting the paper targets?
  2. Why were the baselines implemented as “intentionally simplified protocol adaptations” rather than as the published methods, and why were components such as SUE’s RID/MMD-net removed? Please report results for at least one baseline run from its original implementation so that the headline comparison reflects the published methods.
  3. Which specific timing claims does ClockBridge support or refute in the external analysis, and how do they compare with published findings ?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.