acceptodds
Under review as a conference paper at ICLR 2027

Learning Consistent Multimodal Representations under Missing-Not-at-Random Shifts

Abstract

Missing modalities are ubiquitous in real-world multimodal applications. Although substantial progress has been made in learning from incomplete multimodal data, most existing methods implicitly assume that modality missingness occurs at random and is independent of the prediction target. In practice, however, modality availability can be correlated with the label, giving rise to missing-not-at-random (MNAR) phenomenon. Under such conditions, multimodal models may exploit missingness patterns as predictive shortcuts rather than learning the underlying multimodal semantics, leading to severe performance degradation when missingness patterns shift at inference time. We observe that such shortcut learning induces representation inconsistency, where the representation of the same instance changes with the set of available modalities. We theoretically show that the performance degradation under missingness-pattern shift is bounded by two factors: the discrepancy between source and target missingness patterns and the model's representation inconsistency. This analysis suggests that reducing representation inconsistency can improve robustness to missingness-pattern shifts without requiring prior knowledge of the target missingness distribution. Motivated by this insight, we propose AURA, a multimodal learning framework that explicitly encourages representation consistency across different modality-availability patterns. To enable comprehensive evaluation under MNAR conditions, we introduce Controlled MNAR Shift Evaluation protocol, that systematically controls both the dependence between missingness and labels and the degree of source-target missingness-pattern shift. Extensive experiments on multiple large-scale multimodal datasets demonstrate that AURA consistently improves robustness under MNAR shifts while maintaining strong performance on standard missing-modality benchmarks. On stroke type prediction, AURA's relative improvement in balanced accuracy over the strongest competing method increases from 20.4% under milder MNAR shifts to 60.2% under more severe shifts. Our code is available at https://anonymous.4open.science/r/Learning-Consistent-Multimodal-Representations-under-Missing-Not-at-Random-Shifts

Primary area:
applications to neuroscience & cognitive science
Keywords:
AI for Healthcare, multimodal learning, deep learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RdHiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies label-correlated modality availability, which may encourage models to exploit missingness patterns as predictive shortcuts. Motivated by the argument that such shortcut learning induces representation inconsistency, it proposes AURA to encourage consistent representations across modality-availability patterns, together with a controlled evaluation protocol for missingness-mechanism shifts. Experiments on MIMIC-IV and ADNI assess predictive performance, while additional MNAR-shift experiments show improved robustness over the compared baselines.

Critical strengths

The paper addresses the important problem of model failures under shifts in modality missingness mechanisms and introduces a controlled evaluation protocol to assess robustness to such shifts.

Critical weaknesses
  1. The connection between missingness shortcuts and representation consistency is insufficiently explained. Although the bound motivates stability across missingness patterns, representation changes may also reflect the loss of task-relevant information caused by missing modalities. More direct evidence is needed to establish that the proposed objective reduces reliance on missingness shortcuts.
  2. The baseline comparison does not sufficiently cover recent methods specifically designed for missing-modality learning. MASAM and AMC primarily address modality imbalance rather than missing modalities. Comparisons with recent dedicated missing-modality methods are needed.
Questions for authors

The new controlled protocol has only been validated in the classification of strokes. Can the main conclusions be replicated in another task or dataset?

5 comments
AuthorsOpenReview ↗

Author Response to Reviewer RdHi (Part 1/3)

We thank the Reviewer for recognizing the importance of robustness to missingness-mechanism shifts and the novelty of our Controlled MNAR Shift Evaluation protocol.

We would like to emphasize that our work makes three closely connected contributions: (1) Representation Consistency, a principle supported by a theoretical bound on missingness-shift sensitivity; (2) AURA, a framework designed to promote this property while preserving predictive information; and (3) Controlled MNAR Shift Evaluation protocol, a protocol for systematically evaluating robustness to changes in the label–missingness relationship. The additional analyses and experiments below directly validate the first two contributions and extend the third to a new prediction task.

W1. Distinguishing missingness shortcuts from task-information loss

The connection between missingness shortcuts and representation consistency is insufficiently explained. Representation changes may also reflect the loss of task-relevant information caused by missing modalities. More direct evidence is needed to establish that AURA reduces reliance on missingness shortcuts.

Representation inconsistency δ\delta in Eq. (6) can arise from both reliance on missingness-pattern shortcuts and the loss of task-relevant information when modalities are removed. To distinguish these effects and verify that AURA reduces shortcut reliance while preserving predictive information, we conducted three analyses (detailed results in our responses to Reviewers PArf and 3en7).

1. Missingness-pattern probe (STP). We trained a probe to predict the modality-availability pattern RR from the final learned representation ZZ. AURA substantially reduced the recoverability of missingness patterns: probe F1 decreased from 88.95% without modality dropout and 80.02% without alignment to 58.54% with AURA. This provides direct representation-level evidence that AURA encodes less information about modality availability, supporting its intended effect of discouraging missingness-based shortcuts (see Response to Reviewer PArf, W1.1, Table 1).

2. Controlled MNAR Shift Evaluation protocol with dummy modalities. To isolate shortcut reliance from task-information loss, we constructed a controlled experiment with 7,446 admissions, retaining all three clinical modalities (clinical notes, ICD codes, and laboratory measurements) and adding two label-independent Gaussian dummy modalities sampled from N(0,1)\mathcal{N}(0,1). The Controlled MNAR Shift Evaluation protocol changes only the availability of these dummy modalities, leaving all clinical evidence intact.

With the same modality-dropout procedure, adding AURA's alignment reduced dummy-modality representation inconsistency from 0.6991 to 0.2223. Crucially, TEST-SHIFT AUROC improved from 66.19% to 74.56%, while AUPRC increased from 34.52% to 42.07%. Since no clinical information is removed, these results provide strong evidence that alignment reduces sensitivity to missingness-pattern shortcuts rather than merely compensating for lost modality information (see Response to Reviewer PArf, W1.1, Table 2).

3. Connecting final-representation consistency to robustness and information preservation. We further measured the worst-case Euclidean (ℓ2\ell_2) discrepancy of the final representation ZZ, as defined in Eq. (6), together with the TEST-SAME–TEST-SHIFT binary cross-entropy gap ΔBCE\Delta\mathrm{BCE}. Across w/o dropout → w/o alignment → AURA → full alignment, ℓ2\ell_2 inconsistency decreased monotonically (16.74 → 11.64 → 7.64 → 5.31), accompanied by a corresponding reduction in ΔBCE\Delta\mathrm{BCE} (1.3199 → 0.3743 → 0.2433 → 0.1733). This empirically supports the relationship between representation consistency and shift sensitivity established in Eq. (7) (see Response to Reviewer 3en7, W1, Table 1).

Importantly, AURA selectively aligns only artificially masked modality slots (m∈Dm \in \mathcal{D}), rather than forcing all modality representations to become identical. This helps preserve information in retained modalities that remains useful for prediction. On STP, selective alignment outperformed full alignment in all eight evaluated availability patterns containing at least three modalities, where richer cross-modal information is available (see Response to Reviewer 3en7, W4, Table 4). This explains how AURA balances representation consistency with the preservation of task-relevant information, rather than indiscriminately minimizing representation differences.

Together, these complementary analyses demonstrate that AURA reduces missingness-pattern information in learned representations, improves robustness even when clinical information loss is excluded, and preserves useful task information through selective alignment.

AuthorsOpenReview ↗

Author Response to Reviewer RdHi (Part 2/3)

(Continued from Part 1)

W2. Comparison with recent dedicated missing-modality methods

The baseline comparison does not sufficiently cover recent methods specifically designed for missing-modality learning. MASAM and AMC primarily address modality imbalance rather than missing modalities.

We added two recent dedicated missing-modality methods, DyMo (Du et al., ICLR 2026) and HAD (Liu et al., NeurIPS 2025), and evaluated them on all four prediction tasks under the Standard setting and stroke type prediction (STP) under Controlled MNAR Shift Evaluation protocol.

Unlike MASAM and AMC, which primarily address modality imbalance, these methods directly target incomplete multimodal learning. DyMo dynamically selects reliable recovered modalities at inference time based on task-relevant information, while HAD combines hierarchical feature extraction, Hilbert curve-based Mamba, and mutual-information learning to integrate heterogeneous incomplete modalities while preserving modality-specific information.

1. Standard setting: AURA outperforms recent dedicated missing-modality methods across all four tasks.

Table 1. Standard-setting results on stroke type prediction (STP) and length-of-stay (LOS) prediction. Mean ± standard deviation over five folds. Best results are shown in bold.

MethodSTP BAcc (%)STP F1 (%)STP κLOS MSELOS MAELOS R²
DyMo69.34 ± 1.1869.86 ± 0.980.5489 ± 0.01790.4276 ± 0.02870.4880 ± 0.01360.4015 ± 0.0402
HAD66.65 ± 1.1267.36 ± 1.070.5174 ± 0.01710.3911 ± 0.01680.4853 ± 0.00980.4527 ± 0.0219
AURA73.49 ± 0.3373.77 ± 0.750.6046 ± 0.00570.3216 ± 0.00600.4400 ± 0.00220.5498 ± 0.0105

Table 2. Standard-setting results on 100-day mortality prediction and ADNI disease-stage prediction. Best results are shown in bold.

MethodMort. BAcc (%)AUROC (%)AUPRC (%)ADNI BAcc (%)F1 (%)κ
DyMo69.39 ± 3.5085.90 ± 0.5967.03 ± 1.5158.01 ± 1.2954.84 ± 2.990.3353 ± 0.0488
HAD71.88 ± 2.2784.14 ± 0.7863.43 ± 3.0555.41 ± 2.0255.06 ± 2.240.3108 ± 0.0339
AURA74.69 ± 1.6786.86 ± 0.6969.90 ± 1.1961.65 ± 1.7060.84 ± 1.210.3901 ± 0.0204

AURA achieves the best performance across all reported metrics on all four tasks (Tables 1–2), The improvements are particularly pronounced on LOS regression, where AURA reduces MSE from 0.3911–0.4276 to 0.3216 and increases R2R^2 to 0.5498, and on ADNI, where Cohen's κ\kappa improves from 0.3108–0.3353 to 0.3901. These results support AURA's ability to exploit complementary cross-modal information while maintaining strong predictive performance under incomplete inputs.

2. Controlled MNAR Shift Evaluation protocol: AURA remains more robust than dedicated missing-modality methods.

We further compared AURA with DyMo and HAD under Controlled MNAR Shift Evaluation protocol on STP at τ=0.5\tau=0.5, where the label–missingness relationship changes between TEST-SAME and TEST-SHIFT.

Table 3. Robustness on STP under Controlled MNAR Shift Evaluation protocol (τ=0.5\tau=0.5). Parentheses indicate relative drops from TEST-SAME to TEST-SHIFT; bold denotes the best TEST-SHIFT performance.

MethodSettingBAcc (%)F1 (%)Cohen's κ
DyMoTEST-SAME67.23 ± 2.6867.79 ± 2.280.5159 ± 0.0239
TEST-SHIFT41.01 ± 1.08 (−39.0%)41.33 ± 0.89 (−39.0%)0.1613 ± 0.0185 (−68.7%)
HADTEST-SAME80.65 ± 1.2581.42 ± 0.990.7056 ± 0.0163
TEST-SHIFT45.08 ± 2.91 (−44.1%)45.32 ± 2.79 (−44.3%)0.2219 ± 0.0421 (−68.6%)
AURATEST-SAME80.83 ± 0.6981.80 ± 0.730.7130 ± 0.0111
TEST-SHIFT59.29 ± 1.89 (−26.6%)59.25 ± 1.86 (−27.6%)0.4236 ± 0.0125 (−40.6%)

Table 3 demonstrates robustness over recent dedicated missing-modality methods under Controlled MNAR Shift Evaluation protocol. HAD achieves nearly identical TEST-SAME BAcc to AURA (80.65% vs. 80.83%), its performance drops sharply to 45.08% under TEST-SHIFT, compared with 59.29% for AURA. The advantage is even more pronounced in Cohen's κ\kappa: HAD and DyMo achieve only 0.2219 and 0.1613, respectively, whereas AURA maintains 0.4236. These results demonstrate that AURA is substantially more robust than both DyMo and HAD to changes in the label–missingness relationship.

References

  • Siyi Du, Xinzhe Luo, Declan P. O'Regan, and Chen Qin. Inference-Time Dynamic Modality Selection for Incomplete Multimodal Classification. In ICLR, 2026.
  • Chengliang Liu, Yuanxi Que, Qihao Xu, Yabo Liu, Jie Wen, Jinghua Wang, and Xiaoling Luo. Hierarchical Information Aggregation for Incomplete Multimodal Alzheimer's Disease Diagnosis. In NeurIPS, 2025.
AuthorsOpenReview ↗

Author Response to Reviewer RdHi (Part 3/3)

(Continued from Part 2)

Q1. Generalization of Controlled MNAR Shift beyond stroke classification

The Controlled MNAR Shift protocol has only been validated on stroke classification. Can the main conclusions be replicated on another task or dataset?

To examine whether the robustness advantage observed on STP (Table 3) generalizes beyond stroke classification, we extended the Controlled MNAR Shift Evaluation protocol to 100-day mortality prediction. We evaluated AURA against DyMo, HAD, and the five original baselines under a stricter setting that matches both per-modality artificial-removal rates DmD_m and normalized label–missingness mutual information SMNARS _{\mathrm{MNAR}} between TEST-SAME and TEST-SHIFT at τ=0.5\tau=0.5.

The experiment includes 7,446 admissions, each with three available clinical modalities (clinical notes, ICD codes, and laboratory measurements). We added two label-independent Gaussian dummy modalities sampled from N(0,1)\mathcal{N}(0,1) and changed only their availability. Since all clinical modalities remain available, this setup isolates sensitivity to missingness patterns from the loss of clinical evidence.

Table 4. Robustness on 100-day mortality prediction under dummy-modality Controlled MNAR Shift Evaluation protocol (τ=0.5\tau=0.5), with matched removal rates and MNAR severity. Mean ± standard deviation over five folds. Parentheses indicate relative drops from TEST-SAME to TEST-SHIFT; bold denotes the best TEST-SHIFT performance.

MethodSettingBAcc (%)F1 (%)AUROC (%)AUPRC (%)
HEALNetTEST-SAME84.01 ± 1.5686.54 ± 1.1993.91 ± 0.4488.85 ± 0.78
TEST-SHIFT44.15 ± 0.57 (−47.4%)43.47 ± 0.65 (−49.8%)30.59 ± 2.72 (−67.4%)16.71 ± 0.59 (−81.2%)
Flex-MoETEST-SAME84.46 ± 1.0684.11 ± 1.9093.33 ± 0.8787.84 ± 1.25
TEST-SHIFT44.67 ± 1.56 (−47.1%)44.64 ± 1.43 (−46.9%)28.13 ± 2.22 (−69.9%)16.21 ± 0.58 (−81.5%)
MultiModNTEST-SAME83.38 ± 0.6878.48 ± 1.2291.87 ± 1.0377.89 ± 4.17
TEST-SHIFT49.95 ± 1.27 (−40.1%)49.83 ± 1.27 (−36.5%)33.51 ± 2.71 (−63.5%)25.52 ± 2.18 (−67.2%)
MASAMTEST-SAME83.86 ± 1.9786.83 ± 1.4694.13 ± 0.4989.02 ± 0.94
TEST-SHIFT44.12 ± 1.49 (−47.4%)43.23 ± 1.98 (−50.2%)31.42 ± 2.27 (−66.6%)16.85 ± 0.53 (−81.1%)
AMCTEST-SAME84.14 ± 1.6182.61 ± 1.9792.88 ± 1.2687.11 ± 1.83
TEST-SHIFT44.39 ± 0.92 (−47.2%)44.43 ± 0.94 (−46.2%)27.36 ± 1.95 (−70.5%)15.95 ± 0.31 (−81.7%)
DyMoTEST-SAME80.08 ± 2.7282.51 ± 2.5587.68 ± 1.6179.39 ± 3.66
TEST-SHIFT53.39 ± 3.48 (−33.3%)53.06 ± 4.22 (−35.7%)58.95 ± 5.10 (−32.8%)32.22 ± 6.48 (−59.4%)
HADTEST-SAME76.75 ± 0.9678.74 ± 0.3590.15 ± 0.7279.04 ± 1.42
TEST-SHIFT51.42 ± 1.28 (−33.0%)49.03 ± 0.69 (−37.7%)61.37 ± 5.61 (−31.9%)30.72 ± 2.34 (−61.1%)
AURATEST-SAME74.53 ± 3.6777.39 ± 3.4589.93 ± 1.1877.30 ± 3.82
TEST-SHIFT57.16 ± 2.70 (−23.3%)57.30 ± 3.95 (−26.0%)74.56 ± 1.57 (−17.1%)42.07 ± 3.27 (−45.6%)

The mortality results replicate AURA's robustness on a distinct binary prediction task (Table 4). AURA achieves the highest TEST-SHIFT performance across all four metrics, including 74.56% AUROC, compared with 61.37% for HAD and 58.95% for DyMo. Its relative AUROC decrease is only 17.1%, versus 31.9–32.8% for HAD/DyMo and 63.5–70.5% for the five original baselines.

Importantly, the two tasks provide complementary evidence. STP evaluates shifts that can remove clinically informative modalities, whereas mortality evaluates availability shifts while retaining all clinical evidence. The consistent findings across both tasks demonstrate that AURA's robustness advantage extends beyond the original stroke benchmark and further support the role of representation consistency in reducing missingness-shortcut reliance.

Overall, extending Controlled MNAR Shift Evaluation protocol to mortality prediction confirms that AURA's robustness advantage generalizes across distinct clinical prediction tasks. The results further demonstrate that representation consistency is essential for resisting changes in the label–missingness relationship, beyond simply handling incomplete modality inputs.

Reviewer_3en7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multimodal prediction when the relation between labels and modality availability changes at test time. AURA uses cross-modal refinement, modality dropout, and representation alignment to reduce sensitivity to missingness patterns. The paper also proposes a controlled MNAR-shift protocol and reports results on clinical prediction tasks.

Critical strengths
  1. Missingness-mechanism shift is an important problem when data collection depends on patient condition. The paper defines this setting clearly.

  2. The controlled protocol matches per-modality removal rates while changing the label-missingness relation. This is a useful step toward separating shift from the amount of missing information.

  3. The evaluation includes patient-level folds, several tasks, and component ablations. AURA shows clear gains in the reported shifted stroke and mortality settings.

Critical weaknesses
  1. The theory does not directly match the training objective. Eq 7 uses the Euclidean distance between final representations, while Eq 8 aligns selected modality tokens with Wasserstein distance. The paper should measure the quantity in the bound and show whether AURA reduces it.

  2. The controlled test does not hold MNAR severity fixed across environments. At target severity 0.5, Table 7 reports 0.5001 for training and 0.3905 for Test-Shift, although removal rates are matched. This makes the effect of changing the label-missingness relation harder to isolate.

  3. The controlled shift experiment is reported only for stroke type prediction. Mortality has a flipped-MNAR test, but not the same controlled protocol. The generality of the controlled-shift result remains unclear.

  4. Lower representation inconsistency does not always give better prediction. The full-alignment variant reduces the reported inconsistency from 0.3975 to 0.1923, but lowers Test-Shift kappa from 0.4236 to 0.3987. The paper should explain how selective alignment balances stability and task information.

Questions for authors

Please address the above weaknesses and answer the following questions.

  1. Can you compute the final-representation distance in Equation 6 for AURA and its ablations? Does it track the Test-Same to Test-Shift loss gap?

  2. Can you build a Test-Shift environment that matches both removal rates and normalized label-missingness mutual information? How does AURA perform under this stricter control?

  3. Can you apply the controlled shift protocol to mortality prediction or another task? Do the relative gains remain similar?

  4. Why does full alignment lower Test-Shift performance despite reducing representation inconsistency? Can you report which availability patterns or features lose useful label information?

7 comments
AuthorsOpenReview ↗

Author Response to Reviewer 3en7 (Part 1/3)

We thank the Reviewer for the feedback. We conducted additional experiments to address the four concerns below.

W1 & Q1. Does final-representation inconsistency track the TEST-SAME to TEST-SHIFT loss gap?

Can you compute the final-representation distance in Equation 6 for AURA and its ablations? Does it track the Test-Same to Test-Shift loss gap?

Lower final-representation inconsistency is generally associated with a smaller TEST-SAME–TEST-SHIFT loss gap, consistent with the sensitivity bound in Eq. (7). To examine this relationship, we measured the worst-case Euclidean (ℓ2\ell _2) discrepancy of the final concatenated representation ZZ, as defined in Eq. (6), together with binary cross-entropy (BCE) under TEST-SAME and TEST-SHIFT.

We evaluated 100-day mortality prediction at a target MNAR severity of τ=0.5\tau=0.5. We added two dummy modalities sampled independently of the target label from N(0,1)\mathcal{N}(0,1). All three clinical modalities (clinical notes, ICD codes, and laboratory results) remained available, while only the availability of the dummy modalities changed. This setting allows us to examine sensitivity to missingness-based shortcuts without removing clinical evidence. We define ΔBCE=BCESHIFT−BCESAME\Delta\mathrm{BCE}=\mathrm{BCE} _{\mathrm{SHIFT}}-\mathrm{BCE} _{\mathrm{SAME}}.

Table 1. Final-representation inconsistency and BCE loss under Controlled MNAR Shift Evalutation protocol (τ=0.5\tau=0.5). Scores are means over five folds.

Modelℓ2\ell _2 inconsistencyBCE TEST-SAMEBCE TEST-SHIFTΔ\DeltaBCE
AURA7.640.31750.56080.2433
w/o SCMA7.150.30670.55220.2455
w/o Alignment11.640.26470.63900.3743
w/o Modality Dropout16.740.22791.54781.3199
Full Alignment5.310.33050.50380.1733

Across w/o dropout → w/o alignment → AURA → full alignment, ℓ2\ell _2 inconsistency decreases from 16.74 → 11.64 → 7.64 → 5.31, while the BCE gap decreases from 1.3199 → 0.3743 → 0.2433 → 0.1733. The version without SCMA has similar inconsistency (7.15 vs. 7.64) and BCE gap (0.2455 vs. 0.2433) to AURA. Both variants retain alignment, while SCMA also helps the predictor use cross-modal information and reduce the influence of dummy inputs.

Table 2. Mortality ablation results under Controlled MNAR Shift Evaluation protocol (τ=0.5\tau=0.5). Scores for BAcc, F1, AUROC, and AUPRC are in %; parentheses indicate relative drops from TEST-SAME to TEST-SHIFT.

ModelSettingBAcc (%)F1 (%)AUROC (%)AUPRC (%)
AURATEST-SAME74.5377.3989.9377.30
TEST-SHIFT57.16 (−23.3%)57.30 (−26.0%)74.56 (−17.1%)42.07 (−45.6%)
w/o SCMATEST-SAME79.0780.9391.7781.32
TEST-SHIFT55.81 (−29.4%)55.91 (−30.9%)68.41 (−25.5%)35.04 (−56.9%)
w/o AlignmentTEST-SAME74.2777.1890.8986.45
TEST-SHIFT50.10 (−32.5%)46.06 (−40.3%)66.19 (−27.2%)34.52 (−60.1%)
w/o Modality DropoutTEST-SAME83.9886.7294.4089.24
TEST-SHIFT44.29 (−47.3%)43.53 (−49.8%)40.89 (−56.7%)18.92 (−78.8%)
Full AlignmentTEST-SAME71.9375.9688.5278.75
TEST-SHIFT54.53 (−24.2%)53.58 (−29.5%)72.58 (−18.0%)39.28 (−50.1%)

AURA achieves the highest TEST-SHIFT AUROC (74.56%) and AUPRC (42.07%). Across ablation variants, lower ℓ2\ell _2 representation inconsistency generally corresponds to a smaller TEST-SAME–TEST-SHIFT BCE gap. This relationship suggests that more consistent representations are associated with greater robustness to availability shifts.

Full alignment, which penalizes both hidden and retained modality slots, achieves the smallest ℓ2\ell _2 inconsistency and BCE gap but a lower TEST-SHIFT AUROC than AURA (72.58% vs. 74.56%). It also has a higher TEST-SAME BCE (0.3305 vs. 0.3175), indicating a trade-off between representation consistency and predictive performance. We examine this trade-off in W4.

AuthorsOpenReview ↗

Author Response to Reviewer 3en7 (Part 2/3)

(Continued from Part 1)

W2, W3 & Q2, Q3. Can TEST-SHIFT match removal rates and MNAR severity, and does the result extend to mortality prediction?

Can you build a Test-Shift environment that matches both removal rates and normalized label-missingness mutual information? How does AURA perform under this stricter control? Can you apply the controlled shift protocol to mortality prediction or another task? Do the relative gains remain similar?

We extended the Controlled MNAR Shift Evaluation protocol to 100-day mortality prediction and constructed a stricter TEST-SHIFT environment that matches both per-modality artificial-removal rates and normalized label–missingness mutual information. We evaluated this setting at a target severity, τ=0.5\tau=0.5:

\qquad S_{\mathrm{MNAR}}(T_{\mathrm{test}})=S_{\mathrm{MNAR}}(T_{\mathrm{train}})$$ The dataset contains **7,446 admissions**, each with clinical notes, ICD codes, and laboratory measurements. We added two label-independent Gaussian dummy modalities and the evaluation protocol changed only their availability. Because all three clinical modalities remain present, this setup helps distinguish sensitivity to modality availability from the loss of clinical evidence. We compared AURA against the five original baselines and two recent dedicated missing-modality methods, **DyMo** (Du et al., 2026) and **HAD** (Liu et al., 2025). **Table 3.** Mortality prediction under matched target MNAR severity ($\tau=0.5$). Scores for BAcc, F1, AUROC, and AUPRC are in %; parentheses indicate relative drops from TEST-SAME to TEST-SHIFT. | Model | Setting | BAcc (%) ↑ | F1 (%) ↑ | AUROC (%) ↑ | AUPRC (%) ↑ | |---|---|---:|---:|---:|---:| | **AURA** | TEST-SAME | 74.53 | 77.39 | 89.93 | 77.30 | | | TEST-SHIFT | **57.16** (−23.3%) | **57.30** (−26.0%) | **74.56** (−17.1%) | **42.07** (−45.6%) | | HEALNet | TEST-SAME | 84.01 | 86.54 | 93.91 | 88.85 | | | TEST-SHIFT | 44.15 (−47.4%) | 43.47 (−49.8%) | 30.59 (−67.4%) | 16.71 (−81.2%) | | Flex-MoE | TEST-SAME | 84.46 | 84.11 | 93.33 | 87.84 | | | TEST-SHIFT | 44.67 (−47.1%) | 44.64 (−46.9%) | 28.13 (−69.9%) | 16.21 (−81.5%) | | MultiModN | TEST-SAME | 83.38 | 78.48 | 91.87 | 77.89 | | | TEST-SHIFT | 49.95 (−40.1%) | 49.83 (−36.5%) | 33.51 (−63.5%) | 25.52 (−67.2%) | | MASAM | TEST-SAME | 83.86 | 86.83 | 94.13 | 89.02 | | | TEST-SHIFT | 44.12 (−47.4%) | 43.23 (−50.2%) | 31.42 (−66.6%) | 16.85 (−81.1%) | | AMC | TEST-SAME | 84.14 | 82.61 | 92.88 | 87.11 | | | TEST-SHIFT | 44.39 (−47.2%) | 44.43 (−46.2%) | 27.36 (−70.5%) | 15.95 (−81.7%) | | DyMo | TEST-SAME | 80.08 | 82.51 | 87.68 | 79.39 | | | TEST-SHIFT | 53.39 (−33.3%) | 53.06 (−35.7%) | 58.95 (−32.8%) | 32.22 (−59.4%) | | HAD | TEST-SAME | 76.75 | 78.74 | 90.15 | 79.04 | | | TEST-SHIFT | 51.42 (−33.0%) | 49.03 (−37.7%) | 61.37 (−31.9%) | 30.72 (−61.1%) | **AURA achieves the highest TEST-SHIFT performance across all four metrics:** BAcc (57.16%), F1 (57.30%), AUROC (74.56%), and AUPRC (42.07%). It also exhibits the smallest relative decreases across all four metrics (**23.3%** in BAcc, **26.0%** in F1, **17.1%** in AUROC, and **45.6%** in AUPRC). These results extend the controlled-shift evaluation beyond three-class stroke prediction. They show that AURA retains higher predictive performance when the label–missingness relationship changes under matched removal rates and target MNAR severity. Because clinical evidence remains unchanged, this experiment also provides a controlled setting for examining sensitivity to missingness-based shortcuts. **References** - Siyi Du, Xinzhe Luo, Declan P. O'Regan, and Chen Qin. Inference-Time Dynamic Modality Selection for Incomplete Multimodal Classification. In *The Fourteenth International Conference on Learning Representations (ICLR)*, 2026. URL https://openreview.net/forum?id=PWhDUWRVhM. - Chengliang Liu, Yuanxi Que, Qihao Xu, Yabo Liu, Jie Wen, Jinghua Wang, and Xiaoling Luo. Hierarchical Information Aggregation for Incomplete Multimodal Alzheimer's Disease Diagnosis. In *Advances in Neural Information Processing Systems (NeurIPS)*, 2025. URL https://openreview.net/forum?id=P4KjfMrVNU.
AuthorsOpenReview ↗

Author Response to Reviewer 3en7 (Part 3/3)

(Continued from Part 2)

W4 & Q4. Why can full alignment lower predictive performance, and which patterns are most affected?

Why does full alignment lower Test-Shift performance despite reducing representation inconsistency? Can you report which availability patterns or features lose useful label information?

Full alignment reduces representation inconsistency but can also suppress task-relevant information. For an artificially hidden modality (m∈Dm\in\mathcal D), alignment encourages its representation to recover information from the remaining observed modalities. For a retained modality (m∉Dm\notin\mathcal D), however, the teacher uses context from all originally observed modalities, including those in D\mathcal D, whereas the student lacks this context. Full alignment forces these representations toward each other despite their different available information, which can restrict useful cross-modal refinement. This explains why full alignment can reduce representation inconsistency and the loss gap while degrading predictive performance.

The mortality results illustrate this trade-off. Compared with AURA, full alignment achieves a smaller BCE gap (0.1733 vs. 0.2433) but a higher TEST-SAME BCE (0.3305 vs. 0.3175). Despite the smaller loss gap, its TEST-SHIFT AUROC is lower (0.7258 vs. 0.7456). The predictive trade-off is also evident under TEST-SAME, where full alignment achieves lower AUROC (0.8852 vs. 0.8993).

We compared our selective alignment and the full alignment across 16 availability patterns with sufficient sample sizes for stroke type prediction. Table 4 reports the results for the patterns.

Table 4. TEST-SHIFT Cohen's κ\kappa across all 16 evaluated availability patterns on stroke type prediction. Δκ=κselective−κfull\Delta\kappa=\kappa _{\mathrm{selective}}-\kappa _{\mathrm{full}}.

Availability patternSelective alignmentFull alignmentΔκ\Delta\kappa
clinical + mri + lab + vital0.45640.1999+0.2564
clinical + mri + lab0.39940.2726+0.1268
clinical + ct + mri + lab0.47730.3646+0.1127
clinical + lab + vital0.13390.0390+0.0949
clinical + ct + mri + lab + vital0.57150.4871+0.0845
clinical + ct + lab0.57030.5184+0.0519
ct + lab + vital0.08630.0695+0.0168
clinical + ct + lab + vital0.39910.3840+0.0151
lab + vital0.05210.0489+0.0032
vital0.00150.0004+0.0011
clinical0.49660.5078-0.0112
clinical + ct0.65950.6716-0.0121
clinical + lab0.21510.2405-0.0254
lab0.16800.2045-0.0364
clinical + mri0.64980.6919-0.0421
clinical + vital0.38650.4344-0.0479

Selective alignment achieves higher TEST-SHIFT Cohen's κ\kappa in 10 of the 16 evaluated availability patterns, including all eight patterns with at least three modalities. The largest gains occur for clinical + MRI + lab + vital (Δκ=+0.2564\Delta\kappa=+0.2564), clinical + MRI + lab (+0.1268+0.1268), and clinical + CT + MRI + lab (+0.1127+0.1127). These patterns provide richer cross-modal context, where full alignment can restrict useful refinement by forcing retained modality representations to remain consistent across views with different available information. In contrast, full alignment performs better in six of the eight patterns with only one or two modalities, with much smaller gains (0.01120.0112–0.04790.0479 in κ\kappa). With fewer modalities, there is less cross-modal context to preserve, making the additional alignment constraints less restrictive. These results help explain why selective alignment achieves higher overall TEST-SHIFT κ\kappa (0.4236 vs. 0.3987) despite full alignment producing more consistent representations.

Together, these findings suggest that selectively aligning hidden modality slots helps preserve predictive information while reducing sensitivity to missingness patterns. Full alignment can further reduce representation inconsistency, but its additional constraints on retained modalities may limit useful cross-modal information.

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

This paper investigates the problem of robust multimodal learning under MNAR shift conditions. The authors propose a theoretical upper bound for missingness shift from the perspective of representation consistency and design AURA.  Furthermore, the paper proposes a controlled MNAR shift evaluation protocol and validates the effectiveness of the method on the MIMIC-IV and ADNI datasets.

Critical strengths

Overall, this paper appears to be well-structured with clear figures.

The research problem is important and practical. This paper addresses a critical issue in multimodal learning: when modality availability is correlated with predicted labels, models may exploit modality missing patterns as prediction shortcuts (shortcut learning).

The evaluation protocol is novel. The proposed controlled MNAR shift evaluation enables an assessment of model robustness under different modal missingness mechanisms.

Critical weaknesses
  • The link between the paper’s design and its research motivation is insufficient

The core logic of this paper is:

MNAR → Shortcut learning → Representation inconsistency → Performance degradation

However, representation inconsistency may stem from two causes:

Shortcut learning → harmful representation inconsistency

Loss of modality-specific information → unavoidable representation inconsistency

The term δ (Eq. 6) fails to distinguish between these two scenarios. Also, it remains uncertain whether AURA reduces reliance on shortcuts. It is recommended that the authors either remove this statement or validate it through experiments such as a “probe prediction.”

When a missing modality contains unique information that cannot be recovered from other modalities, alignment may suppress valuable information. The ‘full alignment’ ablation experiments in the paper also illustrate this trade-off.

Furthermore, the test-shift transformation is also drop-only (Eq. (11)), which matches AURA’s training augmentation and may be good at treating such situations.

  • The novelty of this method and the necessity of its design require further justification

AURA primarily combines existing techniques such as selective cross-modal attention, modality dropout and teacher-student alignment.

Furthermore, the necessity of SCMA lacks a sufficient explanation. It selects only certain modalities rather than utilizing the complementary information from all available modalities

Questions for authors
  • Can the authors use a probe to verify whether the model relies on a missing data shortcut, and whether AURA reduces this reliance? For example, by using a probe to predict missing data patterns based on the learnt representations.

  • When a missing modality contains unique information that cannot be recovered by other modalities, how does AURA avoid over-alignment? Could the authors provide experimental analysis under different alignment strengths(λ_align)?

  • Could the authors explain conceptually why the alignment loss and SCMA are needed beyond standard modality dropout?

Minor Presentation Comments

  • Section 4 could be renamed “Method: AURA” to make the section structure clearer.
  • The placement of (3) in the caption of Figure 2 feels awkward; I recommend adjusting it.
  • The abstract is obscure and difficult to understand.
7 comments
AuthorsOpenReview ↗

Author Response to Reviewer PArf (Part 1/4)

We sincerely thank the Reviewer for the thoughtful comments and constructive suggestions. We appreciate the Reviewer's recognition of the importance of the problem and the value of our Controlled MNAR Shift Evaluation protocol. We address the Reviewer's concerns and questions below.

W1.1 & Q1. Distinguishing the causes of representation inconsistency and verifying shortcut reliance

The representation inconsistency δ\delta in Eq. (6) can arise from two sources: (1) reliance on missingness patterns as predictive shortcuts and (2) loss of useful information when a modality is removed. To distinguish these effects and verify whether AURA reduces shortcut reliance, we conducted two experiments: (i) a missingness-pattern probe to measure how much availability information is encoded in the learned representations, and (ii) a controlled experiment with dummy modalities to evaluate missingness shifts without removing task-relevant clinical information.

Experiment 1: Missingness-pattern probe. On the stroke type prediction task, we trained a probe to predict the modality-availability pattern from the final concatenated representation. A lower probe F1 indicates that the missingness pattern is harder to recover from the representation.

As shown in Table 1, AURA achieved a probe F1 of 58.54%, compared with 80.02% without alignment and 88.95% without modality dropout. These results show that AURA reduces the missingness-pattern information recoverable from the learned representation.

Table 1. Missingness-pattern probe on stroke type prediction (STP). Lower F1 indicates that availability patterns are harder to predict from the final representation.

ModelProbe F1 (%)
AURA58.54 ± 5.00
w/o SCMA74.14 ± 7.87
w/o Alignment80.02 ± 2.99
w/o Modality Dropout88.95 ± 3.57
Full Alignment54.94 ± 7.55

Experiment 2: Controlled MNAR Shift Evaluation protocol with dummy modalities. To examine shortcut reliance without the effect of losing clinical information, we conducted an additional experiment on 100-day mortality prediction.

All 7,446 admissions contain three clinical modalities: clinical notes, ICD codes, and laboratory results. We added two dummy modalities sampled from N(0,1)\mathcal{N}(0,1) independently of the target label. We then applied the Controlled MNAR Shift Evaluation protocol to change only the availability of these two dummy modalities. In this experiment, TEST-SAME and TEST-SHIFT have matched per-modality removal rates and normalized MNAR severity. All three clinical modalities remain available in both environments.

We distinguish two representation inconsistency measures: RIall\mathrm{RI} _{\mathrm{all}}, computed over feasible modality-availability subsets in STP, and RIdummy\mathrm{RI} _{\mathrm{dummy}}, computed over patterns that retain all clinical modalities and vary only the availability of the two dummy modalities in mortality prediction.

Since the dummy modalities contain no information, this setting allows us to examine sensitivity to missingness-based shortcuts without the effect of removing clinical evidence.

As shown in Table 2, with the same modality-dropout procedure, adding alignment reduced RIdummy\mathrm{RI}_{\mathrm{dummy}} from 0.6991 to 0.2223. It also improved TEST-SHIFT AUROC from 66.19% to 74.56% and AUPRC from 34.52% to 42.07%. The corresponding relative performance drops decreased from 27.2% to 17.1% for AUROC and from 60.1% to 45.6% for AUPRC.

Together, the probe and dummy-modality experiments show that AURA reduces the representation's dependence on missingness patterns and improves robustness when the label–missingness relationship changes.

Table 2. Ablation results on 100-day mortality prediction under Controlled MNAR Shift Evaluation protocol with dummy modalities. RIdummy\mathrm{RI}_{\mathrm{dummy}} measures representation inconsistency when only dummy-modality availability varies. Percentages indicate relative drops from TEST-SAME to TEST-SHIFT.

ModelSettingBAcc (%)F1 (%)AUROC (%)AUPRC (%)RIdummy\mathrm{RI}_{\mathrm{dummy}}
AURATEST-SAME74.5377.3989.9377.300.2223
TEST-SHIFT57.16 (-23.3%)57.30 (-26.0%)74.56 (-17.1%)42.07 (-45.6%)
w/o SCMATEST-SAME79.0780.9391.7781.320.2649
TEST-SHIFT55.81 (-29.4%)55.91 (-30.9%)68.41 (-25.5%)35.04 (-56.9%)
w/o AlignmentTEST-SAME74.2777.1890.8986.450.6991
TEST-SHIFT50.10 (-32.5%)46.06 (-40.3%)66.19 (-27.2%)34.52 (-60.1%)
w/o Modality DropoutTEST-SAME83.9886.7294.4089.240.8778
TEST-SHIFT44.29 (-47.3%)43.53 (-49.8%)40.89 (-56.7%)18.92 (-78.8%)
Full AlignmentTEST-SAME71.9375.9688.5278.750.1405
TEST-SHIFT54.53 (-24.2%)53.58 (-29.5%)72.58 (-18.0%)39.28 (-50.1%)
AuthorsOpenReview ↗

Author Response to Reviewer PArf (Part 2/4)

(Continued from Part 1)

W1.2 & Q2. Preventing over-alignment and sensitivity to alignment strength

When a missing modality contains unique information that cannot be recovered by other modalities, how does AURA avoid over-alignment? Could the authors analyze different alignment strengths?

AURA limits over-alignment through selective alignment and a moderate alignment weight. Instead of forcing all modality representations to become similar, AURA applies alignment only to artificially masked modalities. This helps maintain representation consistency while preserving useful modality-specific information.

Selective alignment. As defined in Eq. (8), the alignment loss is applied only to the slots of artificially dropped modalities (m∈Dm \in \mathcal{D}). It encourages the model to recover information that can be inferred from the remaining observed modalities. The representations of retained modalities (m∉Dm \notin \mathcal{D}) are not directly penalized by the alignment loss, allowing them greater flexibility to preserve modality-specific information for prediction.

To examine this design, we compared selective alignment with full alignment, which applies the loss to all modality slots. We considered two settings: stroke prediction, where masking can remove useful clinical information, and mortality prediction with dummy modalities, where the controlled missingness shift does not remove clinical information.

On stroke prediction, where RIall\mathrm{RI} _{\mathrm{all}} can reflect both shortcut reliance and loss of modality-specific information, full alignment reduced RIall\mathrm{RI} _{\mathrm{all}} from 0.3975 to 0.1923 (Table 23 in the paper), an absolute difference of 0.2052 (51.6%). However, it also decreased TEST-SHIFT Cohen's κ\kappa from 0.4236 to 0.3987.

On mortality prediction with dummy modalities, where no clinical information is removed, the difference in RIdummy\mathrm{RI} _{\mathrm{dummy}} was smaller: 0.2223 vs. 0.1405, an absolute difference of 0.0818 (36.8%). Selective alignment also achieved higher TEST-SHIFT AUROC (74.56% vs. 72.58%).

Thus, full alignment produced a larger reduction in representation inconsistency on STP, where masking can remove useful information, than in the dummy-modality setting. These results support our interpretation that selective alignment reduces sensitivity to missingness-based shortcuts while preserving useful modality-specific information.

Sensitivity to alignment strength (Table 3). We further evaluated five alignment weights, λalign∈0,0.01,0.1,1,10\lambda _{\text{align}} \in \\{0, 0.01, 0.1, 1, 10\\}, on stroke prediction. Increasing λalign\lambda _{\text{align}} consistently reduced RIall\mathrm{RI} _{\mathrm{all}}, from 0.8890 at λ=0\lambda=0 to 0.3292 at λ=10\lambda=10. However, TEST-SHIFT performance was highest at λ=0.1\lambda=0.1, reaching F1 = 59.25% and Cohen's κ\kappa = 0.4236. At λ=10\lambda=10, F1 decreased to 54.05% and κ\kappa to 0.3626.

These findings show that stronger alignment does not necessarily lead to better prediction. When a modality contains useful information, strong alignment may force representations with and without that modality to become too similar, reducing the model's ability to use this information for prediction. AURA therefore uses selective alignment with a moderate weight to balance representation consistency and predictive performance.

Table 3. Sensitivity to alignment weight on stroke type prediction. RIall\mathrm{RI} _{\mathrm{all}} measures representation inconsistency over feasible modality-availability subsets. Scores are mean ± SD. Drop (%) indicates the relative decrease from TEST-SAME to TEST-SHIFT.

λalign\lambda _{\text{align}}Standard F1 (%)TEST-SAME F1 (%)TEST-SHIFT F1 (%)F1 Drop (%)TEST-SHIFT κ\kappaRIall\mathrm{RI} _{\mathrm{all}}
072.88 ± 0.4484.54 ± 0.6646.42 ± 1.0845.10.2642 ± 0.01600.8890 ± 0.0178
0.0173.66 ± 0.6084.13 ± 0.4754.16 ± 1.2835.60.3575 ± 0.01550.7281 ± 0.0261
0.173.77 ± 0.7581.80 ± 0.7359.25 ± 1.8627.60.4236 ± 0.01250.3975 ± 0.0385
172.54 ± 0.7278.18 ± 1.2358.19 ± 2.7325.60.4040 ± 0.03600.3792 ± 0.0285
1070.23 ± 0.6672.01 ± 0.6154.05 ± 1.1624.90.3626 ± 0.01700.3292 ± 0.0138
AuthorsOpenReview ↗

Author Response to Reviewer PArf (Part 3/4)

(Continued from Part 2)

W2 & Q3. Why are alignment loss and SCMA needed beyond standard modality dropout?

AURA combines existing techniques such as selective cross-modal attention, modality dropout, and teacher-student alignment. Why are alignment loss and SCMA necessary beyond standard modality dropout?

These components serve different purposes. Modality dropout trains the model to handle incomplete inputs. Alignment loss encourages consistent representations across availability patterns. SCMA helps each modality select useful information from other modalities during refinement. Our ablations show that alignment and SCMA provide benefits beyond standard modality dropout.

Alignment loss beyond modality dropout. Modality dropout exposes the model to different combinations of available modalities, but it does not directly require their representations to be consistent. Alignment loss adds this constraint by encouraging the representations of artificially masked modalities to approximate information available from the remaining modalities.

The mortality ablation in Table 2 isolates this contribution by keeping the same modality-dropout procedure and removing only the alignment loss. Without alignment, RIdummy\mathrm{RI}_{\mathrm{dummy}} increased from 0.2223 to 0.6991, and the stroke probe F1 increased from 58.54% to 80.02% (Table 1). On mortality, the absolute TEST-SAME–TEST-SHIFT gap increased from 15.37 to 24.70 percentage points for AUROC and from 35.23 to 51.93 percentage points for AUPRC.

These results show that alignment provides additional robustness beyond standard modality dropout.

SCMA beyond standard attention. The usefulness of a source modality depends on which target modality is being refined. Standard dense attention uses all source modalities, including those that may provide little useful information. SCMA instead uses Sparsemax to select source modalities and allows less useful sources to receive zero weight.

Importantly, SCMA does not remove any modality from the final prediction. It only selects sources during cross-modal refinement. All refined modality representations are concatenated into the final representation ZZ before being passed to the predictor.

To examine this design, we compared SCMA with standard dense self-attention on mortality prediction with two dummy modalities (Tables 4 and 5). In both tables, each row represents a target modality and each column represents a source modality. Entry (i,j)(i,j) reports the average percentage of attention that target modality ii assigns to source modality jj during refinement.

When refining ICD representations, standard self-attention assigned 46.7% of its attention to the two dummy modalities (Table 5). SCMA reduced this to 10.6%, while assigning 40.1% to clinical notes and 49.4% to laboratory results (Table 4). Across the three clinical target modalities, the average attention assigned to dummy modalities decreased from 31.6% to 12.6%.

The prediction results support the benefit of this source selection. Compared with the model without SCMA, AURA achieved higher TEST-SHIFT AUROC (74.56% vs. 68.41%) and AUPRC (42.07% vs. 35.04%) on mortality prediction. Stroke probe F1 also decreased from 74.14% to 58.54%.

Together, these results show that alignment reduces representation dependence on missingness patterns, while SCMA reduces attention to less informative sources. Both provide benefits beyond modality dropout.

Table 4. Mean cross-modal attention (%) with SCMA on mortality prediction. Rows represent target modalities, columns represent source modalities, and entry (i,j)(i,j) is the average attention percentage assigned by target ii to source jj.

Target / SourceClinicalICDLabDummy 1Dummy 2
Clinical—70.813.37.68.3
ICD40.1—49.45.25.4
Lab72.216.6—5.45.8
Dummy 156.97.926.5—8.6
Dummy 257.112.915.514.5—

Table 5. Mean cross-modal attention (%) with standard self-attention on mortality prediction. Rows, columns, and entries follow the same definitions as Table 4.

Target / SourceClinicalICDLabDummy 1Dummy 2
Clinical—58.514.613.213.8
ICD25.1—28.225.820.9
Lab62.216.6—10.410.8
Dummy 161.019.014.9—5.1
Dummy 259.619.810.410.2—

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We sincerely thank the three Reviewers for your careful evaluation and constructive feedback. We appreciate your recognition of the practical importance of robust learning under Missing-Not-at-Random (MNAR) shifts and the value of the Controlled MNAR Shift Evaluation protocol.

In response, we conducted additional experiments and analyses addressing the major concerns raised by all three Reviewers regarding our three main contributions: Representation Consistency, the AURA architecture, and Controlled MNAR Shift Evaluation. We summarize the new evidence below and provide an index to all detailed responses.

1. Key Additional Evidence

1.1. Representation Consistency: Missingness Shortcuts and Shift Robustness

Concern (PArf, 3en7, RdHi): Does representation inconsistency reflect missingness shortcuts rather than information loss, and how does it relate to shift robustness?

Response: We conducted three complementary analyses:

  • Missingness-pattern probe: AURA makes missingness patterns less recoverable from learned representations.
  • Controlled dummy-modality experiment: Alignment reduces inconsistency and improves shift robustness when all clinical information is preserved.
  • Final-representation analysis: Lower inconsistency is generally associated with smaller TEST-SAME–TEST-SHIFT loss gaps, consistent with the motivation of Eq. (7).

Together, these results provide empirical evidence linking representation consistency, reduced missingness-pattern dependence, and shift robustness.

1.2. AURA: Selective Alignment and SCMA Beyond Dropout

Concern (PArf, 3en7): Are alignment and SCMA necessary beyond modality dropout, and how does AURA avoid over-alignment?

Response: With modality dropout held fixed, adding alignment improves shift robustness. Further ablations examine the roles of selective alignment, alignment strength, and SCMA:

  • Selective alignment outperforms full alignment across availability patterns with richer cross-modal information, supporting its role in mitigating over-alignment.
  • Alignment-weight analysis shows that strong alignment weights further reduce representation inconsistency but hurt predictive performance, supporting our choice of a moderate weight.
  • SCMA reduces attention to uninformative modalities, allowing more useful cross-modal information to contribute to the refined representations.

These findings support the architectural choices beyond standard modality dropout.

1.3. Controlled MNAR Shift: Stricter Evaluation and Generalization

Concern (3en7, RdHi): Do AURA's advantages persist under stricter MNAR controls, across tasks, and against recent missing-modality methods?

Response: We extended Controlled MNAR Shift to 100-day mortality prediction, matching both per-modality removal rates and normalized MNAR severity. We also added recent dedicated missing-modality baselines, DyMo and HAD, across standard benchmarks and controlled-shift evaluations.

AURA outperforms these baselines across the standard benchmarks and retains stronger TEST-SHIFT performance on both stroke and mortality prediction. These results provide additional evidence of robustness under stricter controls and beyond the original stroke task.

1.4. Drop-Only Protocol: Evaluation Design and Training Augmentation

Concern (PArf): Does the drop-only TEST-SHIFT transformation favor AURA because it resembles the training augmentation?

Response: The drop-only constraint is necessary because naturally missing modalities have no observed values. Both transformations apply this constraint to all methods.

In contrast, AURA uses random modality dropout during training, without conditioning on labels or the TEST-SHIFT mechanism. Moreover, ablations with identical dropout settings show that alignment further improves TEST-SHIFT performance, supporting its contribution beyond modality dropout.

2. Per-Reviewer Response Index

ReviewerConcern → detailed response
PArfW1.1/Q1: Shortcut reliance vs. information loss → Part 1/4 (Tables 1–2)
W1.2/Q2: Over-alignment and alignment sensitivity → Part 2/4 (Table 3)
W2/Q3: Alignment and SCMA beyond dropout → Part 3/4 (Tables 4–5)
W1.3/Minor: Drop-only constraint and presentation → Part 4/4
3en7W1/Q1: Final-representation inconsistency and loss gap → Part 1/3 (Tables 1–2)
W2–W3/Q2–Q3: Matched MNAR controls and mortality → Part 2/3 (Table 3)
W4/Q4: Full-alignment limitations → Part 3/3 (Table 4)
RdHiW1: Missingness shortcut evidence → Part 1/3
W2: Recent missing-modality baselines → Part 2/3 (Tables 1–3)
Q1: Mortality generalization → Part 3/3 (Table 4)

Thank you again for the thoughtful reviews. We hope these responses address the Reviewers' concerns and would be glad to continue the discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.