acceptodds
Under review as a conference paper at ICLR 2027

PI-NAIM: Path-Integrated Neural Adaptive Imputation Model

Abstract

Medical imaging and multi-modal clinical settings often face the challange of missing modality in their diagnostic pipelines. Existing imputation methods either lack representational capacity or are computationally expensive. We propose PI-NAIM, a novel dual-path architecture that dynamically routes samples to optimized imputation approaches based on missingness complexity. Our framework integrates: (1) intelligent path routing that directs low missingness samples to efficient statistical imputation (MICE) and complex patterns to powerful neural networks (GAIN with temporal analysis); (2) cross-path attention fusion that leverages missingness-aware embeddings to intelligently combine both branches; and (3) end-to-end joint optimization of imputation accuracy and downstream task performance. Extensive experiments on MIMIC-III and multimodal benchmarks demonstrate state-of-the-art performance, achieving RMSE of 0.108 (vs. baselines’ 0.119-0.152) and substantial gains in downstream tasks with an AUROC of 0.812 for mortality prediction. PI-NAIM’s modular design enables seamless integration into vision pipelines handling incomplete sensor measurements, missing modalities, or corrupted inputs, providing a unified solution for real-world scenario. Code and data will be made publicly available

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
3D Scene Completion, Deep Learning, Geometric Modeling, Point Cloud Completion

Reviews

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

This paper proposes PI-NAIM, a hybrid framework for missing data imputation that combines statistical and neural approaches. The method integrates MICE for relatively simple missingness patterns and a temporally enhanced GAIN model for more complex cases. A dynamic routing mechanism is intended to select the appropriate imputation pathway, while cross-path attention combines the imputation representations. The framework further incorporates curriculum learning across MCAR, MAR, and MNAR missingness mechanisms and jointly optimizes imputation accuracy and downstream prediction.

The authors evaluate PI-NAIM on MIMIC-III and CIFAR image datasets. The reported results show improvements in imputation RMSE and downstream AUROC over several baselines. However, important inconsistencies in the methodological description and experimental results make it difficult to assess the reliability of these improvements.

Critical strengths

1. Practical motivation and potentially useful hybrid design.

Combining efficient statistical imputation with more expressive neural models is a reasonable approach to handling heterogeneous missingness patterns. The proposed routing and fusion mechanisms could offer a useful accuracy-efficiency trade-off if their effectiveness were demonstrated convincingly.

2. Consideration of both imputation and downstream prediction.

The paper evaluates reconstruction quality alongside downstream predictive performance and considers multiple missingness mechanisms. This is valuable because improved reconstruction error does not necessarily translate into better task performance.

Critical weaknesses

1. Major inconsistencies and ambiguities in the proposed methodology.

The dynamic routing mechanism is not consistently defined. Section 2.2.1 introduces a learned gating network that predicts routing probabilities from missingness embeddings and missingness rates. However, Equation (5) specifies deterministic routing using a fixed missingness threshold of 0.2. It is therefore unclear whether the routing mechanism is actually learned or rule-based.

Furthermore, the paper describes routing each sample to one imputation pathway, while the cross-path attention module requires outputs from both MICE and GAIN. The interaction between selective routing and dual-path fusion is not explained. This ambiguity also affects the claimed computational efficiency.

The end-to-end optimization claim is insufficiently justified because MICE involves iterative statistical estimation rather than a conventional differentiable neural module. The paper does not explain how gradients propagate through MICE or how the routing and fusion components are jointly optimized.

2. Experimental evidence is insufficient and, in several places, inconsistent with the claims.

The ablation results in Appendix Table 5 contradict the accompanying conclusions. Full PI-NAIM achieves an AUROC of 0.8725, whereas removing adaptive fusion increases AUROC to 0.8764, and removing the imputation path increases it to 0.8775. Despite these results, the paper claims that the complete architecture achieves the highest AUROC and that every component contributes positively.

Additionally, the ablation RMSE values differ substantially in scale from the main results (Table 2), without a clear explanation of the evaluation settings. Important experimental details, including data splitting, missingness simulation, hyperparameters, and variability across runs, are insufficiently documented.

Finally, the paper claims computational efficiency but provides no quantitative comparison of runtime, memory usage, or computational cost. Without these measurements, a central motivation for the dynamic routing architecture remains unvalidated.

Questions for authors

Q1. Routing mechanism: Is the dynamic path selection based on the learned gating network in Equation (4) or the fixed missingness threshold in Equation (5)? If learned routing is used, how is the gating network optimized, and how are routing decisions made during inference?

Q2. End-to-end differentiability and cross-path fusion: How are gradients propagated through the MICE branch during joint training? If each sample is routed to only one branch, how does the cross-path attention module access both imputation outputs? Are both branches executed during inference?

Q3. Ablation inconsistencies: In Appendix Table 5, removing adaptive fusion or the imputation path improves AUROC compared with the full model. How do the authors reconcile these findings with the claim that every component improves downstream performance? Additionally, why are the RMSE values in Table 5 substantially different from those in Table 2?

Q4. Experimental reproducibility and efficiency: Could the authors provide complete experimental configurations, including data split procedures, missingness generation protocols, hyperparameters, and results across multiple random seeds? Since computational efficiency is a central motivation, could the authors also report training and inference costs compared with MICE, GAIN, and NAIM?

1 comment

We thank the reviewer for the clear summary and the identification of the main inconsistencies. We agree the current version does not convincingly demonstrate the claimed improvements.

Method. We will replace the conflicting routing descriptions (learned gate in Section 2.2.1 vs. fixed 0.2 threshold in Eq. 5) with one mechanism: per-sample MR_i = 1 - (1/d)Σ_j M_ij, learned gate G(E,MR_i), training with Gumbel-softmax/straight-through or REINFORCE plus an auxiliary routing loss, hard thresholding only at inference. Routing one sample to one branch is incompatible with cross-path attention requiring both MICE and GAIN outputs; both branches currently run for every sample, so the efficiency claim is unsupported. We will either report the two-branch model, remove the efficiency claim, and position the contribution as missingness-aware fusion, or add a separate hard-routing variant without fusion with runtime, memory, FLOPs, and per-branch counts. MICE is not differentiable; we will state that its outputs are detached and that the gate/fusion are trained jointly with the neural path and task loss, and replace "end-to-end" with "jointly trained with a detached statistical branch" unless a differentiable surrogate is used. The fusion weight uses X_true and y, unavailable at test time; we will use test-time proxies (discriminator/ensemble confidence, reconstruction consistency, predictive entropy), learn lambda_t on validation data, rerun all downstream results, and certify no test labels or held-out true values enter the forward pass.

Experiments. Table 5 shows removing adaptive fusion or the imputation path improves AUROC; we will revise the claims, report and analyze any case where the full model is not best, and correct the parameter-count row. We will recompute the ablation RMSE under the Table 2 protocol and explain any scale difference. We will provide splits, missingness simulation, hyperparameters, seeds, mean +/- std, and significance tests. We will report training/inference costs versus MICE, GAIN, and NAIM, including runtime, memory, FLOPs, and per-branch counts; if the two-branch model does not save compute, we will remove that claim. We will resolve Table 2/5/Fig. 2d by recomputation, define or remove "Default"/"Event," recompute PSNR/SSIM/MSE on the same masked pixels and range for Table 3, remove the nonexistent "w/o NAIM Path" row, and delete Sections 2.7-2.10, which duplicate 2.3-2.6.

Q1. Learned gate (Eq. 4); Eq. 5 removed; gate trained with Gumbel/REINFORCE or an auxiliary routing loss; hard thresholding at inference; MR per sample. Q2. MICE detached; both branches run for every sample in the current implementation, so fusion has both outputs and routing does not decide the output; we will remove the efficiency claim or add a separate hard-routing variant without fusion, and state which variant produces each result. Q3. Claims revised to match the ablations; any component whose removal improves AUROC will be reported and analyzed; RMSE recomputed under the Table 2 protocol. Q4. Complete configurations and variability provided; costs versus MICE/GAIN/NAIM reported, including runtime, memory, FLOPs, and per-branch counts.

We apologize for the errors. The revision will include a reproducibility checklist, corrected references, one consistent method specification, and a complete protocol, and will drop any claim (including SOTA or efficiency) that the corrected experiments do not support.

Reviewer_1GkhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies how an imputation model can decide, record by record, if a fast statistical imputer is accurate enough or a slower but more flexible neural imputer is needed, and whether training imputation together with a downstream classifier improves prediction. The authors propose PI-NAIM, which sends each record to either MICE (a statistical imputer) or GAIN (a GAN-based imputer) depending on how much of it is missing, combines the two outputs, and trains the imputer and a classifier together. On ICU records from MIMIC-III they report lower imputation error and better mortality prediction than four baselines, including MICE and GAIN on their own, and the best image reconstruction by PSNR on CIFAR images with masked pixels. The authors attribute these gains to matching each record to the imputer suited to its amount of missing data and to joint training, and they present the design as a general imputation component for clinical, multimodal and vision pipelines.

Critical strengths

The strength of the paper is its idea of using the expensive neural imputer only on records where the statistical imputer is not accurate enough, which could save compute, and of training the imputer jointly with the classifier, which optimizes it for prediction accuracy rather than reconstruction error alone.

Critical weaknesses
  1. The method description contradicts itself, so I cannot tell which model produced the reported numbers (Section 2).
  • Routing rule. Eq. 5 and Figure 1 use the fixed threshold that Section 2.2 says the learned gate replaces, and Sections 2.3 and 2.4 use yet other thresholds. Eq. 3 gives one missing rate for the whole dataset, which would send every record down the same branch.

  • The paper does not explain how gradients reach the gate through a discrete routing choice, or through MICE, which cannot be differentiated. The statement of end-to-end training is therefore unsupported.

  • Both branches run on every record (Section 2.5, Appendix E). Routing then saves no compute and does not decide the output, which removes the efficiency argument that motivates the paper.

  • The fusion weight in Section 2.6 depends on the true missing values and the true label, neither of which is available at test time. If it is computed this way during evaluation, the reported AUROC uses the test labels; if not, the paper needs to say how it is computed. Appendix F.0.5 gives a different formula.

  • Eq. 6 lacks the penalty on large noise scales from [1], which the paper does not cite, so the loss is minimized by driving every loss weight to zero.

  • The labels used by the linear discriminant analysis step in the MICE branch, the parallel variant of MICE, and three symbols in Section 2.1.2 and Eqs. 2 and 3 are never defined.

  1. The experiments do not support the claims, and the paper's own numbers contradict one another (Section 3 and Appendices B to D). Since the contribution rests on these results, this is the main reason for my rating.
  • The paper does not say how RMSE is computed on MIMIC-III, where naturally missing values have no ground truth, how values were removed for evaluation and at what rate, or which features, time windows, data split, hyperparameters, and number of runs were used, and it reports no standard deviations or significance tests. The results therefore cannot be reproduced or compared with other work. The "Default" and "Event" tasks in Table 2 are never defined (default prediction is not a MIMIC-III task), and the "multimodal benchmarks" in the abstract do not appear.

  • On features standardized to unit variance (Appendix C), mean imputation should give an RMSE near 1, not 0.152. Table 2, Table 5, and Figure 2d give three different results for the full model. Figure 2c shows RMSE near zero throughout training, and the ROC curve in Figure 2d has fewer than ten steps, as if computed on a few positive cases rather than the cohort in Table 6. I cannot tell which of these numbers, if any, describes the model.

  • In Table 5, removing the imputation path improves both AUROC and RMSE, and removing adaptive fusion improves AUROC, contrary to Appendix D. By the paper's own numbers, parts of the model can be removed without loss. The variants discussed in the text also do not match the rows of the table, and the parameter count rises when a component is removed.

  • In Table 3, PI-NAIM has worse SSIM and MSE than three baselines but the highest PSNR. For a fixed pixel range a higher PSNR always means a lower MSE, so these columns cannot come from the same computation. No downstream vision experiment supports the claim in Section 3.4, and without results per mask type, the claim that routing helps with structured occlusions is untested.

  • Cited methods (BRITS [2], SAITS [3], GRU-D [4], MIWAE [5], kNN [6]) are missing, as is a plain average of the MICE and GAIN outputs, the comparison that would show whether routing adds anything. The efficiency claim has no runtime, memory, or FLOP measurement and no count of records sent to each branch, and the uncertainty estimates (Figure 1, Table 4) are never evaluated.

  • Table 2 and Appendix A attribute NAIM to different papers, so I cannot tell which method the headline comparison is against or why it shares the proposed method's name.

References

[1] A. Kendall, Y. Gal, and R. Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. CVPR 2018.

[2] W. Cao, D. Wang, J. Li, H. Zhou, L. Li, and Y. Li. BRITS: Bidirectional recurrent imputation for time series. NeurIPS 2018.

[3] W. Du, D. Cote, and Y. Liu. SAITS: Self-attention-based imputation for time series. Expert Systems with Applications, 219:119619, 2023.

[4] Z. Che, S. Purushotham, K. Cho, D. Sontag, and Y. Liu. Recurrent neural networks for multivariate time series with missing values. Scientific Reports, 8:6085, 2018.

[5] P.-A. Mattei and J. Frellsen. MIWAE: Deep generative modelling and imputation of incomplete data sets. ICML 2019.

[6] O. Troyanskaya, M. Cantor, G. Sherlock, P. Brown, T. Hastie, R. Tibshirani, D. Botstein, and R. B. Altman. Missing value estimation methods for DNA microarrays. Bioinformatics, 17(6):520-525, 2001.

Minor issues

  1. Citations in Section 1 are attached to statements they do not support. For example, a system that protects elderly people from bad habits (Borghesi et al., 2020) is cited for autonomous systems, hyperspectral image classification (Li et al., 2021) for healthcare, and multimodal sentiment analysis (Yan et al., 2020) for hardware malfunctions.

  2. The reference list has duplicate entries (Chen et al. 2023a/b, Cini et al. 2022a/b, Du et al. 2023a/b, and three entries for van Buuren 2018). MWAE is cited as Mattei et al. (2019), whose listed title is an information-extraction system, while MIWAE (Mattei and Frellsen, 2019) is listed separately. Appendix A describes BRITS as Transformer-based, but it is a recurrent model.

  3. The caption of Figure 2 does not match its panels. Panels a and b are bar charts of the Table 2 results, not an "architectural comparison", and panel d is an ROC curve, not an illustration of the training schedule over missingness types. Section 3.1 cites Figure 2 for capabilities that appear only in Table 4.

  4. Sections 2.7 to 2.10 repeat Sections 2.3 to 2.6 almost word for word, and equation numbers are reused and out of order.

Questions for authors
  1. Which routing rule produced the reported results: the fixed 20% threshold in Eq. 5 or the learned gate? Is the missing rate computed per record or over the whole dataset? What share of MIMIC-III records went to each branch?

  2. Is every record run through both branches, as Section 2.5 and Appendix E suggest? If so, what does the routing change?

  3. At test time, how is the fusion weight in Section 2.6 computed, given that it depends on the true values of missing entries and on the label? Were test labels or held-out true values used in any way to produce the reported test AUROC?

  4. How is the gate trained when each record is sent to only one branch, and how do gradients pass through MICE?

  5. On MIMIC-III, how were true values obtained for computing RMSE, how were values removed for evaluation and at what rate, and on what scale is RMSE reported? Please reconcile the RMSE in Table 2 (0.108) and Table 5 (1.952), and the AUROC in Table 2 (0.812), Table 5 (0.8725), and Figure 2d (0.936).

  6. What are the "Default" and "Event" tasks, and which dataset do they come from?

  7. Which paper and which code does the NAIM baseline correspond to?

  8. In Table 3, what pixel range is used, and are PSNR, SSIM, and MSE computed on the masked pixels only or on the whole image? How can PI-NAIM have the highest PSNR but a higher MSE than three of the baselines?

  9. Does the loss in Eq. 6 include the penalty on large noise scales from Kendall et al. (2018)? If not, what stops the noise scales from growing without bound?

1 comment

We thank the reviewer for the detailed and correct identification of inconsistencies. We agree the current version does not support its claims. The revision will define one method and rerun experiments under one protocol.

Method. We will replace the conflicting routing descriptions (Eq. 5 fixed threshold, Fig. 1, Eq. 3 global rate) with a single per-sample MR_i = 1 - (1/d)Σ_j M_ij and the learned gate of Eq. 4, trained with Gumbel-softmax/straight-through or REINFORCE plus an auxiliary routing loss; hard thresholding only at inference. MICE is not differentiable, so its outputs are detached; we will replace "end-to-end" with "jointly trained with a detached statistical branch" unless a differentiable surrogate is used. Cross-path attention requires both branches, so both currently run for every record and routing does not save compute or decide the output; we will remove the efficiency claim and position the contribution as missingness-aware fusion, or provide a separate hard-routing variant without fusion with measured runtime and branch counts. The fusion weight uses X_true and y, unavailable at test time; we will use test-time proxies (discriminator/ensemble confidence, reconstruction consistency, predictive entropy), learn lambda_t on validation data, rerun all AUROCs, and certify no test labels or held-out true values enter the forward pass; Appendix F.0.5 will be reconciled. We will add the Kendall et al. (CVPR 2018) log-variance/penalty term to Eq. 6, and define Psi, e, tau_mask, the LDA labels, and the parallel chained equations.

Experiments. We will specify how MIMIC-III RMSE is computed, which entries were masked, the masking mechanism and rate, features, windows, splits, hyperparameters, and number of runs, with mean +/- std and significance tests. "Default"/"Event" are undefined and will be removed unless identified; "multimodal benchmarks" will be removed unless real. On standardized features, mean imputation should give RMSE near 1, not 0.152; we will recompute the scale. We will recompute Table 2/5/Fig. 2 under one protocol, replace Fig. 2 with matching values, and correct or remove Fig. 2c/2d. Table 5 shows removing the imputation path or adaptive fusion improves AUROC, contrary to Appendix D; we will revise the claims and make the text variants match the table rows. Table 3 is impossible at fixed pixel range; we will recompute PSNR/SSIM/MSE on the same masked pixels. We will add BRITS, SAITS, GRU-D, MIWAE, kNN, and a plain MICE/GAIN average, plus runtime, memory, FLOPs, branch counts, and either evaluation or removal of the uncertainty estimates. The Kim et al. NAIM citation will be removed; we will identify the correct baseline and explain the shared name.

Q1. Learned gate (Eq. 4) with per-sample MR; Eq. 5 removed. Both branches run in the current implementation, so the efficiency claim will be dropped unless a hard-routing variant is provided; branch shares will be reported for any such variant. Q2. Yes, both run because fusion needs both outputs; we will remove the efficiency claim or add a separate hard-routing variant. Q3. No test labels or true missing values in the forward pass; proxies only; AUROCs rerun and affected numbers withdrawn if they were used. Q4. MICE detached; gate trained with Gumbel/REINFORCE or an auxiliary routing loss. Q5. Masking protocol, features, splits, runs, and scale provided; Table 2/5/Fig. 2 conflicts resolved by recomputation. Q6. "Default"/"Event" defined or removed. Q7. Correct NAIM identified with exact code/version; no RMSE/PSNR if it does not impute in the same space. Q8. Pixel range and masked-vs-whole-image stated; metrics recomputed consistently. Q9. Kendall et al. term added and cited. Ethics. We will provide DOIs/URLs for all references, remove any we cannot verify at its stated venue, and ask the AC to treat the current list as unreliable.

Reviewer_NJrHOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes PI-NAIM, an imputation framework that routes each sample to either a statistical branch (MICE) or a neural branch (a WGAN-GP variant of GAIN with temporal attention). Routing is determined by the missingness rate and a learned gate operating on LSTM-based missingness embeddings. The branch outputs are combined through cross-path attention, then integrated with a task representation using a “task-supervised” weight. Training follows an MCAR → MAR → MNAR masking curriculum and uses a multi-task loss with learned uncertainty weights. The experiments evaluate imputation RMSE and downstream AUROC on MIMIC-III against Mean, MICE, GAIN, and NAIM, alongside pixel imputation on CIFAR-10 and an ablation study in Appendix D. The paper claims state-of-the-art performance, reporting an RMSE of 0.108 and a mortality AUROC of 0.812.

Critical strengths

None that would support acceptance. The motivating question, allocating imputation capacity according to how difficult the missingness pattern is, is relevant for clinical data. However, the paper does not turn it into a well-defined method or support it with verifiable evidence.

Critical weaknesses

W1. Numerous references are untraceable or misattributed, including the main baseline.

NAIM, the strongest baseline in Table 2, is cited inconsistently. Table 2 attributes it to Kim et al. (2023), citing a single-cell RNA-seq paper whose title begins with “PI-NAIM”, the proposed method’s name. Appendix A instead cites Luo et al. (2023, IEEE TKDE) and describes NAIM as autoregressive. I could not locate either work. The identifiable NAIM paper is Caruso et al. “Not Another Imputation Method: A Transformer-based Model for Missing Values in Tabular Datasets” (AI Open 7, 2026; arXiv:2407.11540). This model learns from observed values through masked self-attention without imputing missing values, leaving unclear how its RMSE (Table 2) and PSNR (Table 3) were obtained. Similar problems affect the related work. NAOMI (NeurIPS 2019) is attributed to Luo et al., although its authors are Liu, Yu, Zheng, Zhan, and Yue. “MWAE (Mattei et al., 2019)” is described as a hybrid imputation method but references “Miwa: an information-extraction system for the MICHA summit”. I could not locate the cited T-GAIN (Cini et al., NeurIPS 2022), E-GAIN (You et al., AAAI 2021), SPINN (Jarrett and van der Schaar, AISTATS 2022), HINT (Awan et al., KDD 2023), or ST-Transformer (Chen et al., ICLR 2023). Several introductory claims also cite unrelated work, including a hyperspectral remote-sensing survey for healthcare claims and a sentiment-analysis paper for hardware malfunctions. These problems undermine both the main comparison and the positioning of the contribution in Appendix A and Table 4. Meanwhile, HyperImpute (Jarrett et al., ICML 2022), the closest existing approach that adaptively combines classical and neural imputers, is not discussed.

W2. Internal inconsistencies obscure the implemented method and the interpretation of its results. Routing is described through three incompatible formulations: a learned gate with threshold τgate\tau_{\mathrm{gate}} (Equation 4), a fixed MR<0.2\mathrm{MR}<0.2 rule (Equation 5 and Figure 1), and a missingness rate averaged across all nn samples (Equation 3), which cannot support sample-specific routing. The fixed rule is also introduced immediately after fixed thresholds are identified as a limitation. Moreover, Section 2.5 applies attention to the concatenated MICE and GAIN outputs, requiring both branches to run for every sample, as the overhead listed in Appendix E confirms. This undermines the claimed computational benefit of routing, and no runtime results are provided. The fusion weight λt\lambda_t in Section 2.6 depends on Cimp=MSE(X^∣Xtrue)C_{\mathrm{imp}}=\mathrm{MSE}(\hat{X}\mid X_{\mathrm{true}}) and Ctask=CrossEntropy(y,y^task)C_{\mathrm{task}}=\mathrm{CrossEntropy}(y,\hat{y}_{\mathrm{task}}), requiring both the true missing values and the label. As written, this introduces the label into the forward pass and would invalidate the reported AUROCs. The numerical results are also inconsistent. GAIN’s RMSE is 0.119 in the contributions and 0.126 in Table 2. The full model’s MIMIC-III RMSE/AUROC changes from 0.108/0.812 in Table 2 to 1.952/0.8725 in Table 5. Contrary to the accompanying text, removing the imputation path improves both metrics to 1.9453/0.8775. Table 5 also includes a “w/o NAIM Path” variant, although no such path appears in the architecture. In Table 3, PI-NAIM achieves the best PSNR despite worse MSE and SSIM than Mean, MICE, and GAIN. Since PSNR decreases monotonically with MSE, these results imply differences in metric computation that are not explained. Finally, the “Default” and “Event” tasks in Table 2 are undefined, with no corresponding datasets described; Appendix B lists only MIMIC-III and CIFAR. The MIMIC-III RMSE protocol, including the evaluated entries, masking mechanism, and missingness rate, as well as the data splits and number of runs are unspecified. Sections 2.7–2.10 also repeat Sections 2.3–2.6. Together, these discrepancies raise concerns about whether the manuscript was checked against the implementation.

Questions for authors

Q1. Which NAIM implementation was used, and how were imputed values (for RMSE and PSNR) obtained from a model that does not impute? Please also give a DOI or URL for each reference listed in W1. Q2. How are C_imp and C_task computed at test time, when X_true and y are unknown? Were labels used in the forward pass for the AUROCs in Tables 2 and 5? Q3. Which routing rule produced the results, Eq. 4 or Eq. 5? Is MR computed per sample? Are both paths run for every sample, and how do gradients pass through MICE and the hard routing? Q4. On MIMIC-III, which entries were masked to compute RMSE, what are the splits, and how many runs were performed? What are the “Default” and “Event” tasks? Why do Tables 2 and 5 report different values for the full model, and how were PSNR and MSE computed in Table 3?

3 comments

We thank the reviewer for identifying serious inconsistencies. We agree the current manuscript does not adequately specify the method and that several references and numbers are unverifiable. We will revise and will not retain unsupported claims.

W1. References/NAIM. The Kim et al. (2023) NAIM citation was erroneous and will be removed. The identifiable NAIM is Caruso et al. (AI Open 7, 2026; arXiv:2407.11540), which does not impute in the same output space; we will not report RMSE/PSNR for it and will either use it only for downstream prediction or drop it. We will correct NAOMI (Liu et al., NeurIPS 2019) and MWAE (Mattei & Frellsen, ICML 2019), remove the unrelated "Miwa" entry, and remove or replace unverifiable references (T-GAIN, E-GAIN, SPINN, HINT, ST-Transformer) unless a verified DOI/URL exists. We will provide DOIs/URLs for all references, add HyperImpute (Jarrett et al., ICML 2022) as the closest adaptive classical/neural baseline, and fix unrelated citations and duplicates.

W2. Method/experiments. We will replace the three conflicting routing descriptions (learned gate Eq. 4; fixed 0.2 threshold Eq. 5/Fig. 1; global MR Eq. 3) with one mechanism: per-sample MR_i = 1 - (1/d)Σ_j M_ij, learned gate G(E,MR_i), Gumbel/REINFORCE with an auxiliary routing loss, hard thresholding only at inference. Both branches currently run for every sample because cross-path attention needs both outputs, so the efficiency claim is unsupported; we will remove it and position the contribution as missingness-aware fusion, or add a separate hard-routing variant without fusion with runtime/FLOPs/branch counts. Fusion: C_imp and C_task currently use X_true and y, which would invalidate the AUROCs; we will use test-time proxies (discriminator/ensemble confidence, reconstruction consistency, predictive entropy), learn lambda_t on validation data, rerun all results, and certify no test labels or true missing values enter the forward pass (affected numbers withdrawn if they did). We will recompute Tables 2/5 and Fig. 2 under one protocol, remove the nonexistent "w/o NAIM Path" row, and revise ablation claims if removing a component improves AUROC. Table 3: higher PSNR cannot coexist with worse MSE at fixed pixel range; we will recompute PSNR/SSIM/MSE on the same masked pixels/range. "Default"/"Event" will be defined or removed; "multimodal benchmarks" removed unless real. We will specify MIMIC-III masking, features, windows, preprocessing, RMSE scale, splits, and runs. Sections 2.7-2.10 duplicate 2.3-2.6 and will be deleted.

Q1. Kim et al. removed; DOIs/URLs provided; unverifiable refs removed. Q2. No test labels or X_true in the forward pass; proxies only; AUROCs rerun. Q3. Learned gate with per-sample MR; Eq. 5 removed; both paths run for fusion in the current implementation, so the efficiency claim will be dropped; MICE is detached and the gate is trained with Gumbel/REINFORCE or a routing auxiliary loss. Q4. We will provide the masking protocol, features, splits, runs, and scale, and resolve Table 2/5 by recomputation under one protocol; PSNR/SSIM/MSE will be recomputed consistently.

After carefully rereading the revised manuscript, I remain unable to establish a coherent connection between the proposed architecture, the experimental protocols, and the reported results. The manuscript has changed substantially since the initial submission: much of the previous main text now appears in the supplementary material, while the main article has been extensively rewritten. These changes have not resolved my concerns and have introduced further inconsistencies.

  1. End-to-end optimization remains unclear. The MICE branch appears to involve non-differentiable operations. Consequently, I do not understand how the proposed LSTM → MICE → cross-path pipeline can be optimized end to end. The authors should explicitly identify which components receive gradients, how gradients propagate through or around MICE, and whether any components are fitted separately or kept fixed. This information is essential to assess whether the implementation supports the methodological claims.
  2. The main article lacks sufficient quantitative evidence and analysis. The results section provides little substantive interpretation, while the quantitative tables appear in the supplementary material. This organization makes it difficult to assess the contribution from the main article alone. I would move Figure 3 to the supplementary material and bring the central quantitative evidence, particularly Tables 4–6, into the main text, together with a clear analysis of the findings.
  3. The ablation components do not map clearly to the proposed architecture. The main article describes two principal imputation branches, GAIN and MICE, whereas Table 6 introduces a variant labelled “Without NAIM path.” Section B.4 then discusses relationships among the “NAIM path,” “GAIN path,” “MICE path,” and “imputation path” without consistently defining these terms. Its explanations do not establish which modules each ablation removes, which remain active, or why the parameter counts change. Each ablation should correspond to a clearly identified architectural component and a reproducible modification of the full model.
  4. The experimental inconsistencies undermine the interpretation of the results.* Section B.4 explicitly acknowledges several problems that require correction rather than discussion as limitations:
  • The meaning of the “10% sample” comprising 62,111 rows is unclear, including its relationship to the 90%/5%/5% split reported elsewhere.
  • The ablation changes the normalization procedure and prediction target, using min–max scaling and 30-day mortality. It therefore does not directly establish the contribution of each component under the main benchmark protocol.
  • The explanation involving mean absolute error, inverse scaling, and RMSE does not clearly specify the training objective, evaluation metric, or scale on which reconstruction errors were calculated.
  • The row-wise split may place observations from the same patient in both training and testing, creating a risk of patient-level leakage. Different metric values alone do not demonstrate an invalid experiment. However, the acknowledged changes in data partitioning, preprocessing, outcome definition, and evaluation scale prevent a clear comparison between the benchmark and ablation results. The appropriate response is to repeat the ablations under a consistent, patient-disjoint protocol. More broadly, several passages read as an unresolved internal critique of the manuscript rather than a precise account of completed experiments. Acknowledging limitations is appropriate, but it does not resolve contradictions in the experimental design or establish the reliability of the reported findings. My concerns extend to whether the reported results faithfully reflect experiments conducted as described. The manuscript does not provide sufficiently consistent and traceable information for me to establish this. I cannot substantiate a claim that the results were fabricated, but I also cannot confidently attribute them to the stated experimental procedures. I therefore urge the Area Chair to independently examine this aspect of the submission and, where appropriate, request supporting experimental records, configurations, and outputs. In its current form, the manuscript does not provide a sufficiently reliable methodological and empirical basis for acceptance. I maintain my score and my recommendation of clear rejection.

We sincerely thank Reviewer NJrH for taking the time to reread our revised manuscript and provide further detailed feedback. We appreciate the reviewer’s careful examination of the methodological and experimental aspects of our work.

We acknowledge that the previous revision did not adequately resolve several fundamental concerns, particularly regarding the optimization procedure, the relationship between the architectural components and the ablation experiments, and the consistency of the experimental protocols.

We would like to address the main concerns as follows:

1. End-to-end optimization and gradient propagation

We acknowledge that MICE is a non-differentiable statistical procedure and should not be presented as a component optimized through conventional backpropagation. The revised methodological description will explicitly distinguish between the statistical MICE branch, whose parameters are fitted separately, and the trainable neural components, including the GAIN network, missingness encoder, gating mechanism, attention layers, and prediction head. We will clarify the gradient flow and remove any unsupported claims of fully end-to-end optimization.

2. Presentation of quantitative results

We agree that the main manuscript should provide sufficient quantitative evidence to allow readers to evaluate the contribution without relying heavily on supplementary material. We will reorganize the results section to include the central quantitative comparisons, particularly the relevant results currently presented in Tables 4–6, together with a more detailed interpretation of the findings. We also agree that Figure 3 would be more appropriate for the supplementary material.

3. Architectural consistency and ablation studies

We acknowledge the ambiguity surrounding the terminology used for the NAIM path, GAIN path, MICE path, and imputation path. We will revise the terminology and explicitly define each architectural component. Every ablation will be mapped to a specific module, explaining what is removed, what remains active, and how the corresponding parameter count changes. Unsupported or incorrectly described ablation configurations will be removed.

4. Experimental protocol and reproducibility

We recognize that the inconsistencies in data partitioning, normalization, prediction targets, and evaluation metrics prevent a reliable comparison between the main experiments and the ablation study. We will re-examine the experimental pipeline and establish a consistent evaluation protocol using patient-disjoint data partitions. The experiments will be repeated under the same preprocessing, prediction targets, normalization procedures, and evaluation metrics. We will also clarify the relationship between the reported 10% sample and the training, validation, and test partitions.

We take the concern regarding potential patient-level data leakage particularly seriously. We will verify the patient-level separation and ensure that the reported results are supported by traceable experimental configurations and outputs.

Finally, we appreciate the reviewer’s recommendation that the Area Chair independently examine the experimental records. We respect this recommendation and are prepared to provide the available implementation details, experimental configurations, and supporting results for independent examination.

We sincerely apologize that our previous revision did not sufficiently address these issues. We understand that methodological clarity and reproducible experimental evidence are essential, particularly for a contribution claiming improvements in clinical data imputation.

Thank you again for your constructive criticism. Your comments have helped us identify important weaknesses that require substantive correction rather than merely editorial changes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.