acceptodds
Under review as a conference paper at ICLR 2027

CAMOS: Coupled Oscillatory State-Space Model for Multimodal Clinical Time-Series

Abstract

Longitudinal clinical cohorts are multimodal, irregularly sampled and pervasively incomplete: in ADNI, positron emission tomography and cerebrospinal fluid assays are absent from roughly half of all visits. Linear state-space models handle irregular sampling gracefully but treat a missing modality by masking the input, leaving the transition operator untouched. We prove that this is a representational limitation: the latent state of any linear state-space layer whose transition operator does not depend on the availability pattern is an additive function of the availability indicators, so no such layer can represent an interaction between two modalities being jointly present or jointly absent. We propose CAMOS, which gives each modality a bank of second-order oscillators coupled through a matrix that sits inside the differential equation and is gated by availability, so the transition operator itself becomes a function of which measurements were taken. Coupling invalidates the analysis of uncoupled oscillatory models, and we restore it: a per-channel Gershgorin budget makes the effective stiffness positive definite uniformly over all availability patterns and all gaps, an energy argument charges amplification to availability transitions rather than sequence length, and a channel factorization preserves exact associative parallel scans. On ADNI, CAMOS outperforms uncoupled oscillatory state-space models and clinical fusion models on same-visit staging, landmark prediction and longitudinal forecasting, and under zero-shot transfer to OASIS-3 it is the only model that avoids collapse to the majority class.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multimodal Learning, Clinical Time-Series, State-Space Models

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_Ry6NOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper deals with irregular sampling and missing modalities in longitudinal multimodal data. It identifies and proves a representational limitation of linear state-space models whose transition operators are independent of modality availability. The paper then proposes CAMOS, which makes the transition operator a function of the modality availability pattern. The method is evaluated on two public longitudinal clinical datasets, ADNI and OASIS-3, using Alzheimer’s disease-related prediction tasks.

Critical strengths
  1. The paper identifies and proves the modality availability-additivity limitation of linear state-space layers whose transition operators do not depend on the modality availability pattern: their latent states cannot directly represent higher-order interactions among availability indicators.
  2. multi-facet experiments: CAMOS is evaluated on three complementary tasks, same-visit staging, landmark prediction, and longitudinal forecasting on ADNI dataset. And it demostrates stronger zero-shot transfer to the OASIS-3 dataset.
Critical weaknesses
  1. The motivation for incorporating availability into the transition operator is not fully convincing. As the paper pointed in (line 255-260) Theorem 5 establishes availability-additivity for an individual linear state-space layer; it does not show that deep networks with nonlinear readouts cannot model availability interactions, nor does it distinguish CAMOS from representation-level fusion methods such as MedFuse or DrFuse, whose encoders and fusion modules are nonlinear with respect to availability. If such interactions can already be modeled by adding nonlinear layers or fusion modules, why is it necessary to encode them in the transition operator, particularly given the additional structural constraints imposed to obtain the stability guarantees?
  2. Regarding the research gap, the paper claims that ''Multimodal fusion methods fuse after the temporal modelling has been applied, so nothing lets one modality shape another’s trajectory across a long gap.'' (line 053-056) This is too strong a claim without sufficient references or empirical evidence to support it. Why cannot the trajectory of one modality influence another in existing longitudinal multimodal fusion methods?
  3. The presentation could be improved for clarity. Several notations are undefined. In particular, uˉqn\bar{u}_q^n in Figure 1, a0a_0, P0P_0, C(r)C^{(r)}, and α(i),r\alpha{(i),r} are not explicitly defined. Theorem 7 depends on the undefined boundary quantities a0a_0 and P0P_0. Moreover, hh, ss, and γ\gamma are each overloaded for several unrelated objects. The framework figure does not clearly specify the inputs and outputs of each module, making the data flow less clear.
  4. About the experiments, several commonly used longitudinal multimodal baselines are missing, including FlexMoE[1], FuseMoE[2], UMSE[3], and Zhang et al., 2023's work[4]. Baselines such as MedFuse and DrFuse were not originally designed for longitudinal multimodal modeling; in their original settings, they use a single or last available CXR as input rather than modeling longitudinal observations for every modality. Therefore, they do not include dedicated temporal modeling for each modality. Further justification and implementation details for ''replacing their modality encoders with per-modality gated recurrent encoders over the visit sequence'' would help establish a clearer and more informative comparison.

References:

[1] Yun, Sukwon, et al. "Flex-MoE: Modeling arbitrary modality combination via the flexible mixture-of-experts."Advances in Neural Information Processing Systems 37 (2024): 98782-98805.

[2] Han, Xing, et al. "FuseMoE: Mixture-of-experts transformers for fleximodal fusion." Advances in Neural Information Processing Systems 37 (2024): 67850-67900.

[3] Lee, Kwanhyung, et al. "Learning missing modal electronic health records with unified multi-modal data embedding and modality-aware attention." Machine learning for healthcare conference. PMLR, 2023.

[4] Zhang, Xinlu, et al. "Improving medical predictions by irregular multimodal electronic health records modeling." International conference on machine learning. PMLR, 2023.

Questions for authors
  1. Do all baselines use the same task heads and jointly optimize the same multi-task losses (Equation 10) as CAMOS?
  2. Can the authors report performance under controlled missing rates and modality combinations, and analyze results by availability pattern to show when transition gating helps?
  3. MedFuse and DrFuse were originally evaluated on MIMIC-based datasets[5], and the MIMIC ecosystem provides longitudinal multimodal data, including CXRs, clinical notes, ECGs, and electronic health records. Could the authors explain why MIMIC was not included in the experiments? Evaluation on this dataset would help assess CAMOS in a broader longitudinal multimodal setting.

Reference: [1] Johnson, Alistair EW, et al. "MIMIC-IV, a freely accessible electronic health record dataset." Scientific data 10.1 (2023): 1.

6 comments

We thank the reviewer for the valuable feedback. We have addressed all comments and updated the paper accordingly wherever possible.

Questions For Authors:

Q1: Yes.

Protocol elementShared by all models
Task headsSame three heads (staging, landmark, forecasting)
LossesSame masked losses with weights (λst,λcv,λfc)=(1,0.5,1)(\lambda_{\mathrm{st}}, \lambda_{\mathrm{cv}}, \lambda_{\mathrm{fc}}) = (1, 0.5, 1)
Selection criterionMean of staging macro-F1, landmark AUROC and one minus the normalized forecasting MAE
Hyper-parameter grid and trial budgetIdentical (Appendix G)

MedFuse and DrFuse use per-modality gated recurrent encoders over the visit sequence, so they see the same temporal information as the state-space models; the oscillatory baselines receive the availability vector and the gaps as extra channels (Appendix G, baseline adaptation). The ℓ1\ell_1 penalty on the couplings in Eq. 10 applies only to CAMOS, since no baseline has a coupling matrix.

Q2: The only systematic case in the original submission is OASIS-3, where CSF is never observed. We now add a controlled analysis on the trained ADNI checkpoints (full table in our response to Reviewer yCSE, Question 6). With CSF and imaging both removed at test time, CAMOS keeps a macro-F1 of 0.765 (all observed: 0.779) and a landmark AUROC of 0.949 (0.961), whereas the strongest baseline, DrFuse, falls from 0.735 to 0.690 in macro-F1 and from 0.939 to 0.903 in AUROC. The advantage of CAMOS therefore increases as more biomarkers are missing.

The ablation shows that the availability gate contributes to this. The table gives the ungated variant minus the full model, with 95% paired bootstrap intervals over test subjects:

PatternMacro-F1Landmark AUROCLandmark AUPRC
All observed−0.010-0.010 (−0.022,+0.002)(-0.022, +0.002)−0.013-0.013 (−0.027,+0.001)(-0.027, +0.001)−0.077-0.077 (−0.146,−0.008)(-0.146, -0.008)
CSF removed−0.014-0.014 (−0.027,−0.001)(-0.027, -0.001)−0.018-0.018 (−0.033,−0.003)(-0.033, -0.003)−0.093-0.093 (−0.164,−0.022)(-0.164, -0.022)
CSF + imaging removed−0.021-0.021 (−0.036,−0.006)(-0.036, -0.006)−0.027-0.027 (−0.045,−0.009)(-0.045, -0.009)−0.118-0.118 (−0.197,−0.039)(-0.197, -0.039)
CSF/imaging dropped 75%−0.025-0.025 (−0.042,−0.008)(-0.042, -0.008)−0.031-0.031 (−0.050,−0.012)(-0.050, -0.012)−0.132-0.132 (−0.213,−0.051)(-0.213, -0.051)

With everything observed, the ungated variant is already worse than the full model in Table 3 (landmark AUPRC 0.7916 vs. 0.8687, forecasting MAE about 11% higher). Its macro-F1 and AUROC gaps (0.010 and 0.013) are small and, with 162 test subjects, not individually distinguishable from zero, whereas the AUPRC gap is. The gaps grow when several biomarker modalities are missing together: with CSF and imaging both removed, or three quarters of their blocks dropped, the ungated variant loses 0.021 to 0.025 macro-F1, 0.027 to 0.031 AUROC and 0.118 to 0.132 AUPRC, all with intervals excluding zero, roughly 1.5 to 2.5 times the all-observed gaps. This is consistent with the explanation given with Table 3: without the gate, unobserved or unreliable states propagate across modalities, whereas with the gate an unobserved oscillator is isolated from the others (Proposition 3).

Q3:

AspectTarget regime of CAMOS (ADNI, OASIS-3)MIMIC
Time scaleMulti-year gaps (0.5 to 4 years in ADNI)Dense hourly physiology, short stays
MissingnessProcedure-level (lumbar puncture, PET session)A single chest X-ray per stay in the MedFuse and DrFuse configurations
Disease courseSlowly progressing, mechanistically coupled modalitiesAcute care

Our target regime is continuous-time dynamics over multi-year gaps with procedure-level missingness in slowly progressing disease, and MIMIC’s regime lies outside it. It would be a worthwhile test of generality, and we list it as future work.

Critical Weaknesses:

W1: Remark 8 states that Theorem 5 concerns a single linear layer and does not separate CAMOS from MedFuse or DrFuse. We do not claim that transition gating is necessary. It is a sufficient mechanism with properties that readout-level reconstruction does not provide:

AspectReadout-level or fusion-level reconstructionTransition gating (CAMOS)
Isolation of an unobserved modalityLearned from the availability patterns seen in trainingExact for all 2M2^M patterns, including unseen ones (Proposition 3)
Where the interaction is formedBy a nonlinear readout acting on an additive summary of the historyInside the recurrence, at the visit where it occurs
Stability under changing availabilityNo statementUniform in pattern and gap (Lemma 2, Proposition 6, Theorem 7)
CostAdditional layers or fusion modules0.14% of the parameters per layer; constraints hold by construction (Eq. 5); parallel scan preserved (Lemma 14)

The empirical evidence is consistent with this: CAMOS improves on MedFuse and DrFuse on all three tasks; on OASIS-3, where CSF is never observed, DrFuse falls to the single-class solution and CAMOS does not; ungated coupling is worse than no coupling on landmark prediction and forecasting (Table 3); and the effect of the gate grows with missingness (Q2).

W2: The statement holds for late-fusion designs such as MedFuse and DrFuse, where each modality has its own temporal encoder and the temporal state of one modality is not an input to the update of another. It does not hold for methods that interleave cross-modal attention with temporal modelling, such as Zhang et al. (2023), UMSE and Coupled Mamba.

FamilyCross-modal influence over timeTreatment of the gapGuarantee under a missing modality
Late fusion (MedFuse, DrFuse)None inside the temporal encodersPer-modality encoder inputLearned representation of the missing modality
Interleaved attention or coupled states (Zhang et al., 2023; UMSE; Coupled Mamba)Learned, in a discrete token or state spaceEmbedding or featureNone stated
CAMOSCoupling inside the differential equationIntegration step Δtn\Delta t_nExact isolation and stability bounds

W3:

SymbolDefinition
uˉqn\bar u^n_q, bqb_q (Figure 1)The projected input u~kn\tilde u^n_k and input matrix BkB_k of Eqs. 1 and 6, with qq the modality index kk
a0a^0, P0P_0 (Theorem 7)Convention a0:=a1a^0 := a^1, hence P0:=P1P_0 := P_1, so the first visit is not counted as a transition; E0E_0 is the energy of the initial state x0x^0 from the static encoder
α(i),r\alpha_{(i),r}The ii-th smallest of α1,r,…,αM,r\alpha_{1,r}, \dots, \alpha_{M,r}, matched to the ii-th smallest eigenvalue νi(r)\nu_i^{(r)} in Weyl’s inequality
hhProjection width; also the forecast horizon index and the modality token hknh^n_k
ssStatic covariates; also Δtn2\Delta t_n^2 in Proposition 11 and the Laplace variable in Proposition 12
γ\gammaCoupling strength max⁡r∥C(r)∥2\max_r \lVert C^{(r)} \rVert_2; also the horizon discount γh\gamma_h

For hh, ss and γ\gamma the meaning is fixed by context in each case. The data flow in Figure 1 is:

ModuleInputOutput
Encodersuknu^n_k, ana^n, Δtn\Delta t_n, ssu~n\tilde u^n, x0x^0
GCO stack (LL blocks)u~n\tilde u^n, x0x^0, ana^n, Δtn\Delta t_ny(L),ny^{(L),n}, split into modality tokens hknh^n_k
Within-visit attentionhknh^n_k, ana^nhˉn\bar h^n
Task headshˉn\bar h^np^n\hat p^n, q^\hat q, u^kn+h\hat u^{n+h}_k

W4: Appendix I compares against twelve baselines from four families, and CAMOS exceeds all twelve on every ADNI staging and landmark metric:

FamilyBaselines
Uncoupled oscillatory SSMsLinOSS-IM, LinOSS-IMEX, D-LinOSS
General and coupled SSMsS5, Mamba, Coupled Mamba
Irregular-time and missing-value modelsGRU-D, mTAN, Raindrop, ODE-RNN
Clinical fusion modelsMedFuse, DrFuse

mTAN is the temporal module underlying the irregularity-aware embeddings of FuseMoE and Zhang et al. (2023), so that family is represented. Flex-MoE is formulated for a single observation per subject and needs the same longitudinal adaptation as MedFuse and DrFuse.

In their original form, MedFuse and DrFuse use a single or last image and would be compared on unequal information. We therefore replace each modality encoder by a gated recurrent encoder over that modality’s visit sequence and keep the published fusion module of each method, so both receive more temporal information than in their original form, with the same heads and losses as CAMOS.

Thank you for the clarifications and the experiments.

  • Regarding Q2, in the experiment result table, compared to full model, the "CSF removed" "CSF + imaging removed" "CSF/imaging dropped 75%" variants actually has two changes: the input data, and the transition gating, which makes the drop of the performance hard to be attributed. Further control of the experiment would be helpful.
  • Regarding Q3, a follow up question is why CAMOS can not fit MIMIC dataset (with the same irregularity and missingness problem)?
Reviewer_P4D8OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+2)
Read review
Paper summary

This paper addresses irregular visit intervals and missing modalities in longitudinal clinical datasets like ADNI, where expensive or invasive tests such as PET scans or spinal taps for CSF are missing at many visits. Standard linear state-space models (SSMs) handle a missing test by zeroing out its input while leaving the state transition matrix unchanged. The authors prove that under this approach, the hidden state is an additive function of the availability indicators, meaning a single layer cannot capture interactions between which tests were jointly observed or jointly missing (Theorem 5).

To fix this, the authors propose CAMOS, which models each modality as a bank of second-order oscillators coupled together inside the differential equation. When a modality is missing at a visit, its coupling connections to other modalities are gated off in the stiffness matrix, isolating its state so unobserved modalities do not influence observed ones. Because coupling oscillators across modalities breaks both the stability guarantees and the fast parallel computation of uncoupled oscillatory SSMs, the authors add a per-channel Gershgorin row-sum budget that keeps the coupled stiffness matrix positive definite across all 2M2^M missingness patterns, an implicit backward-Euler discretization whose energy bound grows with the number of availability changes rather than with sequence length, and a channel-wise factorization that, after reordering channels, splits the system into dd independent M×MM \times M blocks so the recurrence can still be computed with an O(log⁡N)\mathcal{O}(\log N)-depth parallel associative scan. The model is evaluated on ADNI and zero-shot on OASIS-3 for disease staging, three-year MCI-to-AD conversion and next-visit forecasting.

Critical strengths

The paper identifies a specific limitation of input masking in linear state-space models. Theorem 5 shows that when a missing modality is handled by zeroing its input, the hidden state is additive in the availability indicators, so a single layer cannot represent interactions between which tests were jointly observed, for example an effect of having both CSF and imaging at the same visit that differs from the sum of their separate effects. Remark 8 limits this claim to a single linear layer. This motivates gating the transition rather than the input, which is the central idea of the paper and is relevant to clinical data with structured missingness.

The supporting design restores the stability guarantee and the efficient parallel scan that uncoupled oscillatory SSMs rely on, both of which coupling removes by making the stiffness matrix non-diagonal. The symmetric product gate removes unobserved modalities from the coupling graph (Proposition 3), the Gershgorin budget keeps the stiffness matrix positive definite for all 2M2^M availability patterns with a single condition (Lemma 2), the implicit discretization stays stable over multi-year gaps (Proposition 6, Theorem 7), and the channel factorization keeps the associative scan within block-structured matrices (Lemma 14). The proofs in Appendices C and D cite the assumptions and earlier results they use, which makes them easier to follow.

Critical weaknesses

The supplementary code raises questions about how the reported results were produced. The conversion head in the code reads the hidden state at each subject’s last valid visit (model.py lines 138 to 143), whereas Section 4.5 pools over visits with learned weights, and neither the code nor the paper restricts inputs to visits up to the index MCI visit defined in Appendix H. If either version produced Table 2, the landmark results use visits after the prediction time, including the visit where AD was diagnosed, and the reported AUROC of 0.9611 and AUPRC of 0.8687 would not measure three-year conversion. The code differs from Section 4 in its other components as well. It uses a dense low-rank Laplacian coupling without the Gershgorin budget, computes transitions with a matrix exponential and a learned damping term rather than backward Euler, runs a sequential loop rather than the associative scan, and has no within-visit attention. The ablations labeled IMEX and asymmetric gate implement forward Euler and a symmetric coupling. The code may be an earlier version, but as submitted the experiments do not test the model the theory describes, and resolving this would most affect my assessment.

In the paper’s evaluation, the genetics features (APOE ε4 allele count, genotype indicators and a polygenic risk summary) do not change over time, yet they are forecasting targets and account for the largest reported gains (−45.8% MAE on ADNI, −63.9% on OASIS-3). The ablation with C ≡ 0 reaches an AUPRC of 0.8193, above the best baseline (DrFuse, 0.7689), which suggests most of the margin comes from components other than the coupling that Theorem 5 motivates. In addition, the text around lines 454 to 458 reports different numbers for the ungated variant than Table 3, and Table 5 reports 7,432 visits for 2,749 ADNI subjects, a mean of 2.7 visits per subject, although Section 5 and Appendix H (lines 1076 to 1077) require at least three diagnosed visits for inclusion. Together, these mean the empirical section does not yet support the claim that coupling drives the improvements.

Questions for authors

Clarifications:

  1. Which implementation produced Tables 1 to 3? Does it differ from the supplementary code?
  2. What are the conversion results, for CAMOS and the baselines, with inputs truncated at the index MCI visit?
  3. Figure 1 lists APOE4 as a static covariate, while Section 5 uses genetics as a longitudinal modality. Which setup produced the results, and how do the forecasting gains change when genetics is treated as a static covariate rather than a forecasting target?
  4. Which numbers for the ungated variant are correct, and how is the 2.7 mean visit count reconciled with the three-visit inclusion rule?

Suggestion for future work:

A synthetic experiment in which the target depends on which modalities are jointly observed would test whether the coupling captures the availability interactions that Theorem 5 shows input masking cannot, which the clinical benchmarks do not isolate.

3 comments

We thank the reviewer for the valuable feedback. We have addressed all comments and updated the paper accordingly wherever possible.

Questions For Authors:

Q1: The reported results were produced by the final implementation of Section 4. The code we uploaded was an earlier development snapshot, used for smoke tests on synthetic data because ADNI and OASIS-3 cannot be redistributed. It omitted components of the final model. The behaviours you list describe that snapshot, not the model of the paper.

The complete implementation is uploaded in the supplementary material. It contains:

ComponentReference in the paper
Budgeted couplingEq. 5, Assumptions 1 and 2
Backward-Euler discretization via per-channel CholeskyEq. 8
Associative scanAppendix D
Within-visit attention, three task heads and the joint objectiveSection 4.5, Eq. 10
ADNI preparation script with the inclusion rule and landmark truncationAppendix H
Test suite for the stability and isolation propertiesLemma 2, Propositions 3 and 6

Q2: For landmark prediction, the index visit is the first MCI diagnosis, and the inputs to the landmark head are truncated at that visit: the model sees only visits up to and including the index visit, and the pooling of Eq. 17 runs over those visits only,

q^=σ ⁣(wcv⊤∑n≤nidxωnhˉn+bcv).\hat q = \sigma\!\Big(w_{\mathrm{cv}}^\top \sum_{n \le n_{\mathrm{idx}}} \omega_n \bar h^n + b_{\mathrm{cv}}\Big).

Later visits, including any AD-diagnosed visit, are never visible, so the label cannot be read from the input. Because the recurrence is causal and attention acts within a single visit, this is equivalent to feeding the truncated sequence, and the released code includes a test that verifies it (later visits cannot change the landmark prediction).

The baselines received the same truncated inputs: their landmark predictions also use only visits up to and including the index visit, so CAMOS and every baseline in Tables 2 and 7 are evaluated under the same protocol. All twelve baselines reach AUROC between 0.88 and 0.94 under this protocol, so the level of AUROC is a property of the task and the features.

Q3: The results use the setup of Section 5: genetics is a longitudinal modality (APOE ε4 allele count, genotype indicator, polygenic risk summary; p3=3p_3 = 3), and the static covariates are baseline age, sex and years of education (ps=3p_s = 3). The static APOE4 label in Figure 1 is a diagram simplification; the implementation and the text follow Section 5.

These genetic features are time-invariant per subject, so the gains on that modality reflect retention of subject-specific constants rather than trajectory prediction, and we do not use them as evidence for coupling. Forecasting metrics are computed per modality, so the gains on the other modalities are unaffected. Restricting Table 8 to the time-varying modalities:

ADNI next-visit MAECAMOSBest of the 12 baselinesChange
Cognition0.35880.4176 (DrFuse)−14.1%-14.1\%
CSF0.59060.7056 (Mamba)−16.3%-16.3\%
Imaging0.41460.4468 (LinOSS-IM)−7.2%-7.2\%
Mean, three time-varying modalities0.45470.5537 (MedFuse)−17.9%-17.9\%
Mean, all four modalities0.44320.5899 (Mamba)−24.9%-24.9\%

CAMOS remains the best model on each time-varying modality.

Q4: The values in Table 3 are correct. The sentence at lines 457 to 459 of the submitted version reports ungated values (AUROC 0.9531, AUPRC 0.8296, macro-F1 0.7804, MAE 0.5016 sparse / 0.4014 dense) that come from an earlier run. The text was not updated when Table 3 was finalized. This is an error in the text, not in the experiments.

Ablation (Table 3)Macro-F1AUROCAUPRCForecasting MAE vs. CAMOS (sparse / dense)
CAMOS0.77900.96110.8687reference (0.5097 / 0.4029)
Coupling ungated0.76940.94780.7916+11.2%+11.2\% / +11.4%+11.4\%
Without coupling0.75610.95300.8193+7.7%+7.7\% / +8.1%+8.1\%
IMEX discretization0.65300.95460.8063+39.8%+39.8\% / +25.5%+25.5\%

On landmark prediction and forecasting the ungated variant is therefore worse than having no coupling at all, as the ablation paragraph (lines 465 to 468) states. In the uploaded revised paper we have replaced this sentence (now lines 457 to 461) with the Table 3 values and the corrected conclusion, and revised the Conclusion (now lines 486 to 487) to attribute the gains to the implicit discretization, the coupling and the availability gate.

The inclusion rule in Appendix H is correct: a subject is retained if it has at least three visits with a recorded diagnosis and, at each retained visit, at least one fully observed modality. The statistics in Table 5 were, however, computed on the table before this filter was applied, which is why a mean of 2.7 visits per subject is incompatible with the rule.

ADNI cohortSubjectsVisitsMean visits/subjectMax visits/subject
Table 5, submitted (before the filter)2,7497,4322.715
Table 5, revised (cohort given to the model)1,0725,4755.115

We have corrected Table 5 in the uploaded revised paper. No result in the paper changes, since all experiments were run on the filtered cohort.

Q5 Suggestion: We agree that the clinical benchmarks do not isolate the interaction of Theorem 5. The controlled missingness analysis (Reviewer yCSE, Q6; Reviewer Ry6N, Q2) is the closest evidence in the present data, and a synthetic task whose target depends on joint availability is the direct test.

Critical Weaknesses:

The weaknesses in code implementation and landmark prediction are addressed under Q1 and Q2.

W1: The genetics features are addressed under Q3 and the two reporting errors under Q4. On the contribution of coupling: the variant without coupling already exceeds the strongest baseline, so we report how the ADNI margin divides between the gated coupling and the rest of the architecture.

MetricStrongest baselineC≡0C \equiv 0CAMOSTotal marginDue to gated coupling
Staging macro-F10.7354 (DrFuse)0.75610.77900.04360.0229 (53%)
Landmark AUROC0.9386 (DrFuse)0.95300.96110.02250.0081 (36%)
Landmark AUPRC0.7750 (mTAN)0.81930.86870.09370.0494 (53%)

About half of the margin is due to the gated coupling. The remainder comes from the per-modality oscillator banks with gated forcing, the implicit discretization, the static-conditioned initial state and the within-visit attention, and the revised Conclusion attributes the gains to these components jointly.

Two results concern the availability-dependent transition specifically:

EvidenceResult
Coupling helps only when it is gatedUngated coupling gives AUPRC 0.7916, below the uncoupled variant (0.8193; Welch p<0.001p < 0.001)
The effect of the gate grows with missingnessThe deficit of the ungated variant rises from 0.010 to 0.021 in macro-F1 and from 0.077 to 0.118 in AUPRC when CSF and imaging are removed, with intervals that exclude zero (table in our response to Reviewer Ry6N, Q2)

Thank you for answering my questions and for updating the supplementary files. My initial score was based on inconsistencies between the paper and the original code, including data leakage in the landmark task. I have verified that the revised code fixes these issues and now uses the proposed architecture, including the backward-Euler discretization. I also appreciate the clarifications on the genetic features, and because the empirical results now support the paper’s claims, I have raised my score.

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

This paper introduces CAMOS, a coupled oscillatory state-space model for irregularly sampled multimodal clinical time series with missing modalities. Unlike conventional state-space models that handle missingness primarily through input masking, CAMOS incorporates modality availability into the transition dynamics through availability-gated cross-modal coupling. The authors develop a stability-preserving coupling parameterization, establish an availability-additivity limitation for a restricted class of linear state-space layers, and derive stability and computational properties that preserve associative parallel scans. Experiments on ADNI and OASIS-3 evaluate disease staging, three-year MCI-to-AD conversion prediction, and longitudinal forecasting. The results are promising, but inconsistencies in cohort statistics, the precise scope of the theoretical claims, and the interpretation of the zero-shot transfer protocol require further clarification.

Critical strengths
  1. The paper introduces a principled mechanism for incorporating modality availability into state-transition dynamics rather than treating missingness solely as an input-masking problem.
  2. The Gershgorin-based coupling constraint provides a mathematically motivated way to maintain positive definiteness across availability patterns and irregular time gaps.
  3. The channel-wise factorization preserves an associative scan structure while enabling cross-modal interactions, addressing an important computational challenge in coupled state-space models.
  4. The theoretical analysis covers representational limitations, stability under changing availability patterns, and computational complexity, providing a relatively comprehensive justification for the design.
  5. Experiments evaluate three complementary clinical prediction tasks and report improvements over oscillatory state-space models and clinical fusion baselines on both ADNI and OASIS-3.
Critical weaknesses
  1. The cohort statistics appear inconsistent: Appendix H states that every retained subject has at least three diagnosed visits, yet Table 5 reports 2,749 ADNI subjects and 7,432 visits, corresponding to only 2.7 visits per subject on average.
  2. The OASIS-3 evaluation is described as zero-shot transfer, but the normalization statistics are refitted on the target cohort, making the precise transfer protocol and its reliance on target-domain data unclear.
  3. The availability-additivity theorem applies only to linear state-space layers whose transition operators are independent of availability, so it does not establish a general representational limitation for nonlinear fusion architectures or availability-conditioned selective state-space models.
  4. The empirical evaluation is restricted to two Alzheimer's disease cohorts, and the extent to which the proposed dynamics generalize to other clinical populations, diseases, and missingness mechanisms remains uncertain.
  5. The reported statistical significance is assessed using five random seeds, but the paper does not sufficiently clarify whether the tests account for uncertainty at the subject level or dependence among repeated runs on the same data split.
  6. The improvement on OASIS-3 is substantial for some metrics, but several baseline results exhibit high variability, making it important to establish how robust the reported transfer advantage is across different cohort splits and evaluation settings.
Questions for authors
  1. Could the authors reconcile the ADNI inclusion criterion of at least three diagnosed visits per subject with the reported total of 7,432 visits across 2,749 subjects, which implies an average of only 2.7 visits per subject?
  2. Which OASIS-3 samples are used to estimate the refitted normalization statistics, and would the results remain similar if normalization were performed using only an explicitly designated unlabeled calibration subset or fixed ADNI statistics?
  3. Can the authors clarify the intended scope of Theorem 5 and explain how CAMOS differs from models whose transition dynamics explicitly depend on availability indicators, including selective state-space models?
  4. Can the authors provide additional evaluations on other clinical cohorts or under controlled synthetic missingness patterns to assess robustness to different modality availability distributions?
  5. Were the reported significance tests conducted across independent subject-level evaluations or only across five random-seed runs on fixed splits, and can the authors provide confidence intervals based on subject-level uncertainty?
  6. Can the authors report how the performance changes across different availability patterns, particularly when key modalities such as CSF or PET are systematically missing, to establish that the learned coupling remains reliable under clinically relevant missingness?
3 comments

We thank the reviewer for the valuable feedback. We have addressed all comments and updated the paper accordingly wherever possible.

Questions For Authors:

Q1: Thank you for catching this inconsistency. The inclusion rule in Appendix H is correct: a subject is retained if it has at least three visits with a recorded diagnosis and, at each retained visit, at least one fully observed modality. The statistics in Table 5 were, however, computed on the table before this filter was applied, which is why a mean of 2.7 visits per subject is incompatible with the rule.

ADNI cohortSubjectsVisitsMean visits/subjectMax visits/subject
Table 5, submitted (before the filter)2,7497,4322.715
Table 5, revised (cohort given to the model)1,0725,4755.115

We have corrected Table 5 in the uploaded revised paper. No result in the paper changes, since all experiments were run on the filtered cohort.

Q2: The only quantities estimated on OASIS-3 are the feature normalization statistics. They are computed from feature values only, with no labels, on the 963-subject training partition of OASIS-3, which is disjoint from the 207 test subjects (the partition is 963 / 207 / 207 subjects). No model parameter is updated, and every model in the transfer tables received the same refitted statistics, so the comparison between models does not depend on this choice.

The protocol therefore already normalizes on a designated unlabeled calibration subset that is disjoint from the evaluation subjects. Fixed ADNI statistics are not well defined for the full feature set, because several OASIS-3 features are the nearest available measure of the same construct (Appendix H) and differ in scale from their ADNI counterparts. Here, “zero-shot” means no retraining and no label access, with input standardization refitted on the OASIS-3 training partition.

Q3: Theorem 5 is deliberately limited to the class A\mathcal{A} of Definition 4: single linear layers whose TnT_n, BknB_k^n and RR do not depend on the availability pattern. Definition 4 already excludes selective models such as Mamba with the availability vector appended, since selection makes TnT_n depend on aa. Remark 8 states that the theorem does not show that deep or nonlinear networks cannot represent availability interactions, and that it does not separate CAMOS from MedFuse or DrFuse.

What distinguishes CAMOS is where availability enters: through a symmetric product gate on the coupling inside

Pn=A−Dan C Dan.P_n = A - D_{a^n}\, C\, D_{a^n}.

Property of the gated transitionResult
Exact isolation of unobserved modalitiesProposition 3
Positive definiteness for all 2M2^M availability patterns and all gapsLemma 2, Proposition 6
Switching bound charged to availability transitionsTheorem 7

We do not claim that selective state-space models cannot model such interactions; this comparison is empirical. Mamba and Coupled Mamba are among the baselines of Appendix I:

ADNIMambaCoupled MambaCAMOS
Staging macro-F10.73040.70520.7790
Landmark AUPRC0.61870.64630.8687

Q4: We agree that the evidence covers two Alzheimer’s cohorts. Additional cohorts are beyond the scope of what can be done within the rebuttal period. The two cohorts do differ in their missingness mechanism: CSF is entirely unobserved in OASIS-3, where CAMOS retains macro F1 0.5557 while every baseline stays below 0.44. In addition, we now report performance under controlled modality masking on trained checkpoints (see our response to Q6).

Q5: The significance markers are two-sided Welch’s t-tests over five random-seed runs of each model. Therefore, they reflect training stochasticity only, not the sampling uncertainty of the test subjects.

We now report subject-level uncertainty. We resample the 162 test subjects with replacement (2,000 resamples, the same resamples for all models, so differences are paired; a subject brings all of its visits) and give 95% percentile intervals for every metric and for the difference between CAMOS and the strongest baseline, computed on the mean over the five runs. The intervals describe test-subject sampling only. All five paired intervals exclude zero.

MetricCAMOSStrongest baselineDiff (CAMOS −- baseline)
Accuracy0.782 (0.760, 0.802)DrFuse 0.732 (0.708, 0.756)+0.050+0.050 (+0.021,+0.079)(+0.021, +0.079)
Staging macro-F10.779 (0.757, 0.799)DrFuse 0.735 (0.712, 0.757)+0.044+0.044 (+0.013,+0.075)(+0.013, +0.075)
Landmark AUROC0.961 (0.934, 0.981)DrFuse 0.939 (0.906, 0.966)+0.023+0.023 (+0.002,+0.044)(+0.002, +0.044)
Landmark AUPRC0.869 (0.769, 0.947)mTAN 0.775 (0.654, 0.879)+0.094+0.094 (+0.006,+0.182)(+0.006, +0.182)
Forecasting mean MAE ↓\downarrow0.443 (0.421, 0.465)Mamba 0.590 (0.562, 0.618)−0.147-0.147 (−0.175,−0.119)(-0.175, -0.119)

Q6: OASIS-3 is a naturally occurring case of systematic missingness, i.e., the CSF modality is never observed. There CAMOS retains three-class discrimination (macro F1 0.5557, macro recall 0.5261) while all baselines are below 0.44. Proposition 3 gives the mechanism that an unobserved oscillator is removed from the coupling graph, so its state cannot corrupt the others.

To test this under controlled conditions, we mask CSF, imaging, or both at test time on the trained ADNI checkpoints and compare CAMOS with the strongest baseline, analysing results by availability pattern.

PatternMacro-F1, CAMOSMacro-F1, DrFuseAUROC, CAMOSAUROC, DrFuse
All observed0.7790.7350.9610.939
CSF removed0.7750.7220.9580.928
Imaging removed0.7720.7150.9550.921
CSF + imaging removed0.7650.6900.9490.903
CSF/imaging dropped 50%0.7700.7080.9530.915
CSF/imaging dropped 75%0.7610.6800.9450.895

From all-observed to CSF and imaging both removed, CAMOS loses 0.014 macro-F1 (95% interval 0.004 to 0.024), DrFuse loses 0.045 (0.023 to 0.067); the difference between the two losses is 0.031 (0.007, 0.055), excluding zero. The same holds for landmark AUROC (−0.012-0.012 vs. −0.036-0.036). The lead of CAMOS over the strongest baseline in macro-F1 therefore grows from +0.044+0.044 with everything observed to +0.075+0.075 with CSF and imaging removed and to +0.081+0.081 when 75% of the CSF/imaging blocks are dropped. This is the ADNI counterpart of the OASIS-3 transfer result, where CSF is never observed. Cognition remains observed in all patterns, so this analysis probes the loss of biomarker modalities.

Critical Weaknesses:

W1 to W5 correspond to Q1 to Q5 and are answered there.

W6: The staging advantage under transfer is resolved across seeds, and the landmark AUPRC advantage is not:

OASIS-3 metricCAMOSStrongest of the 12 baselinesWelch pp (5 seeds)
Staging macro-F10.5557±0.00280.5557 \pm 0.0028Mamba, 0.4312±0.04920.4312 \pm 0.04920.0050.005
Staging macro recall0.5261±0.00300.5261 \pm 0.0030Mamba, 0.4389±0.04060.4389 \pm 0.04060.0090.009
Landmark AUPRC0.7894±0.00930.7894 \pm 0.0093S5, 0.5982±0.18890.5982 \pm 0.18890.090.09

The marker on OASIS-3 AUPRC in Table 2 refers to the four baselines shown in that table. The robust transfer result is therefore the staging result, where CAMOS attains the highest macro-F1 and macro recall of all thirteen models.

Thank you for the detailed responses and additional experiments. The authors have addressed my major concerns and strengthened the empirical evidence. I therefore raise my rating from 2 (Weak Reject) to 3 (Weak Accept).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.