acceptodds
Under review as a conference paper at ICLR 2027

Integrated PU Learning and Deep Generative VAEs for Indication-Driven Missingness to Improve MNAR Imputation and Prediction

Abstract

Recent advances in deep generative modeling have improved imputation for missing not at random (MNAR) data. However, jointly improving both imputation quality and downstream prediction remains challenging. In this paper, we propose to use the positive-unlabeled (PU) learning to deal with indication-driven missingness to leverage partially observed measurement-intent variable to improve MNAR imputation and downstream prediction. We first use a PU adversarial network to learn probabilities of unknown positives of intention to measure among all the missing cases, which then guide the training of variational autoencoders (VAE) that jointly model observed data and missingness indicators. Experiments on synthetic and real-world data show that our proposed framework has superior performance for both imputation and downstream prediction tasks. These results show that explicitly modeling missingness indicators integrated with additional PU measurement intent information can improve MNAR imputation and downstream predictions under various scenarios and assumptions.

Primary area:
generative models
Keywords:
Missing data imputation, Missing not at random (MNAR), Positive-unlabeled learning, Deep generative models, Variational autoencoders

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_38umOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

In some datasets, a value is often missing not because it is randomly missing, but because it was on purpose not recorded, maybe because the data was not considered useful. This means that a latent variable is responsible for its missing. The paper calls this indication-driven missingness. The authors consider that standard MNAR missing-not-at-random) imputers only models whether a value was observed, and thus cannot separate “measured and observed” from “intended to be measured but missing” from “never intended to be measured.” The proposed method, PU-VAE-MNAR, treats the latent measurement-intent indicator as a positive-unlabeled (PU) learning problem: observed entries are known positives, everything else is unlabeled. The authors use a per-variable adversarial PU classifier to turn fully observed covariates into a soft intent probability. This latter which conditions a VAE-based MNAR imputer. Imputed data are fed to a downstream logistic-regression predictor.

Critical strengths

The paper is novel in its framing and the treating of latent measurement intent as a PU-learning problem and using it to condition an MNAR VAE. The components (PAN, not-MIWAE-style mask decoding, conditional priors) are existing building blocks. The main novelty is in their combination and in the indication-driven formulation. The paper is well written and uses a large, realistic data (106k patients, 167 labs, 92% missingness), which is far bigger and more realistic than the small UCI tables common in deep imputation papers. The mix of continuous and binary labs is also handled explicitly. The paper evaluates together imputation and prediction., while many imputation papers report only RMSE. The authors of this also measures downstream prediction. The paper compares against 10 imputers across four families (statistical, MissForest, GAN, VAE/MNAR), including recent MNAR-specific models (not-MIWAE, GINA, PSMVAE, GNR), with links to the code for reproductibility. The paper is honest about its weaknesses (below mean imputation at extreme sparsity) and includes a detailed appendix (hyperparameter ranges, architectures, simulation equations). The authors do also provide an explicit AI-use statement.

Critical weaknesses

I could not really grasp from the paper how important the problem really is in practice. While missing data is a real problem and imputation techniques are really needed in practice. It's not clear to which extent using an imputation technique on healthcare related data is legally acceptable to use in medical sciences. The paper uses a non public real world dataset, which is frustrating. It should have used a publicly available dataset. There is no rationale given in section C.2 SYNTHETIC DATA GENERATION - on the choice of selected form, For instance X1 = 0.8Z1 + 0.4Z2 + 0.2Z3 + ϵ1. How does the approach work/scale in the presence of several latent variables (much more than 3) ?
The authors do no validate intent on a setting with ground truth or with review of high-q unlabeled entries done by health personnel. The paper misses an ablation that replaces PAN with (a) a plain classifier of S̃ on a, and (b) nnPU with an estimated prior, keeping everything else fixed. In general the choice of PAN is not justified and there is no evidence provided that it works in this setup. In the synthetic study, where the true intent S is known, the quality of PAN's scores (AUC or calibration against S) is never reported.

Questions for authors

Why do the results not include confidence intervals ( Tables 2–5 ) ? How does the approach work/scale in the presence of several latent variables (much more than 3) ? Can the authors demonstrate the solution on a public real world dataset ? Can the authors somehow discuss what happens if the underlying latent model is more complicated than the assumed one ? Can the authors discuss why PAN is used ? MaybeA a plain classifier would likely do also. Under the standard assumption that labeled positives are a random sample of all positives an ordinary classifier might do the job. That classifier needs no prior, no adversarial training, and costs 167 logistic regressions or GBMs instead of 167 GAN-style pairs. The paper might try to assess this alternative.

11 comments
AuthorsOpenReview ↗

Response to Reviewer 38um (Q5)

Q5: Why is PAN preferred over conventional classifiers?

Can the authors discuss why PAN is used ? MaybeA a plain classifier would likely do also. Under the standard assumption that labeled positives are a random sample of all positives an ordinary classifier might do the job. That classifier needs no prior, no adversarial training, and costs 167 logistic regressions or GBMs instead of 167 GAN-style pairs. The paper might try to assess this alternative.

We sincerely appreciate your insightful suggestion. For additional experimental comparisons with conventional classifiers and nnPU, please refer to our detailed response to W5.

Empirically and theoretically, we adopt Predictive Adversarial Networks (PAN) [1] for measurement-intent estimation because PAN does not require a known positive-class prior and demonstrates robust classification performance even when only a small fraction of positive examples are labeled. Its adversarial learning objective encourages the identification of latent positives (S=1,S~=0S=1, \widetilde{S}=0) within the unlabeled set U\mathcal{U} while distinguishing them from true negatives (S=0,S~=0S=0, \widetilde{S}=0), aligning with our partially observed measurement-intent setting.

To further validate intent estimation, we have added expected calibration error (ECE) results against the ground-truth SS in the right-hand table of Figure 2, providing an absolute assessment of probability calibration. The low ECE values indicate good calibration of the estimated measurement-intent probabilities. Furthermore, our real-world EHR ablation study (Table 6) demonstrates that PAN-based conditioning and PU supervision both improve imputation accuracy compared with nnPU. Specifically, RMSE decreases from 0.183 (nnPU) to 0.121 (PAN conditioning) and further to 0.110 (full PU-VAE-MNAR), representing a 39.9% reduction relative to nnPU. These findings support the empirical suitability of PAN for the proposed framework.

References:

[1] Wenpeng Hu, Ran Le, Bing Liu, Feng Ji, Jinwen Ma, Dongyan Zhao, and Rui Yan. Predictive Adversarial Learning from Positive and Unlabeled Data. Proceedings of the AAAI Conference on Artificial Intelligence, 35(9):7806–7814, 2021.

AuthorsOpenReview ↗

Response to Reviewer 38um (Q4)

Q4: Robustness to more complex latent models

Can the authors somehow discuss what happens if the underlying latent model is more complicated than the assumed one ?

We thank you for this interesting and constructive question.

If your concern is about the dimensionality of the latent variables, please refer to our responses to W3 and Q2, where we provide additional simulations with 16 latent variables and discuss the architecture used in our real-world experiments.

If you are referring to more complex generative models beyond VAE, our PU integration framework is not restricted to a specific generative architecture. The PU step estimates measurement-intent probabilities q∗\mathbf{q}^{\ast} independently of the generative model. These probabilities can potentially be integrated into other architectures, such as diffusion-based or Transformer-based models, through conditioning and auxiliary supervision. For example, a diffusion-based model could condition its denoising network on q∗\mathbf{q}^{\ast}:

ϵθ(xt,t,a,q∗).\boldsymbol{\epsilon}_{\theta}(\mathbf{x}_{t},t,\mathbf{a},\mathbf{q}^{\ast}).

Similarly, a Transformer-based model could integrate q∗\mathbf{q}^{\ast} into its input representations:

Transformer⁡(Embed⁡(xobs,m,a,q∗)).\operatorname{Transformer}\left(\operatorname{Embed}(\mathbf{x}^{\mathrm{obs}},\mathbf{m},\mathbf{a},\mathbf{q}^{\ast})\right).

We choose VAE because it allows us to jointly model observed values xobs\mathbf{x}^{\mathrm{obs}} and missingness masks m\mathbf{m} under MNAR. Its encoder, data decoder, and missingness decoder can be conditioned on q∗\mathbf{q}^{\ast}, while the variational training objective allows us to add weighted auxiliary PU supervision. This makes VAE a natural choice for evaluating the contribution of measurement-intent information to both imputation and downstream prediction.

However, such extensions would require architecture-specific designs and additional evaluation. Given recent advances in diffusion-based imputation for tabular data, we consider this a promising direction for future work.

We have also discussed this possibility in the revised Conclusion:

Future work may extend the proposed framework to accommodate more complex missingness mechanisms through additional conditioning strategies and alternative generative architectures, including integration into diffusion-based and Transformer-based models.

AuthorsOpenReview ↗

Response to Reviewer 38um (Q3)

Q3: Evaluation on a public real-world dataset

Can the authors demonstrate the solution on a public real world dataset?

We appreciate your suggestion. As discussed in our response to W2, we use the Cerner Health Facts dataset to illustrate the application of PU-VAE-MNAR to real-world EHR data. Our framework is not specific to Cerner and can be adapted to other datasets with partially observed measurement-intent information.

We agree that evaluation on a public real-world dataset would improve reproducibility and provide further evidence of generalizability. However, the current study does not include such an experiment. Instead, we evaluate the framework using three simulation scenarios with known data-generating mechanisms and a large real-world EHR cohort of 106,261 patients.

The simulation studies allow us to evaluate imputation accuracy against known ground truth and examine the proposed method under different MNAR mechanisms. The Cerner experiments further evaluate its application under naturally occurring EHR missingness, using fixed patient-level training, validation, and held-out test partitions, as described in Appendix D.3.

We acknowledge the lack of public real-world validation as a limitation. Extending the evaluation to suitable public EHR datasets with identifiable measurement-intent information is an important direction for future work.

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

The paper introduces indication-driven missingness, distinguishing the intention to measure a variable from the availability of its observed value. It formulates partially observed measurement intent as a positive–unlabeled (PU) learning problem and proposes PU-VAE-MNAR, a two-stage framework that employs an existing predictive adversarial network (PAN) to estimate measurement-intent probabilities and subsequently incorporates these probabilities into VAE-based MNAR imputation through conditional modeling and an auxiliary PU objective. Synthetic MNAR simulations, semi-synthetic experiments, and an EHR cohort are used to evaluate imputation and downstream HIV prediction. The results show promising improvements over the selected baselines.

Critical strengths

1.By distinguishing measurement intent from actual value availability and formulating partially observed measurement intent as a positive-unlabeled (PU) learning problem, the paper provides a meaningful new perspective on missing-data modeling. 2.The experimental results demonstrate that incorporating information derived from PU learning can improve downstream predictive performance.

Critical weaknesses

1.The methodological novelty is limited, and the contributions of the proposed PU integration mechanisms are not sufficiently validated through ablation studies. 2.The lack of comparisons with representative recent imputation methods limits the assessment of the proposed framework's competitiveness. 3.The missingness generation mechanism in the semi-synthetic experiments is not clearly specified, making it difficult to assess the relevance of the results to MNAR settings.

Questions for authors

1.The current ablation study does not sufficiently distinguish the contributions of PU-based conditioning and auxiliary PU supervision. I suggest evaluating these components separately and reporting both imputation accuracy and downstream prediction performance to better understand their individual contributions. 2.Although the paper includes several conventional and deep learning-based baselines, comparisons with representative recent tabular imputation methods are limited. I encourage the authors to include more recent approaches, such as ReMasker (Imputing Tabular Data with Masked Autoencoding) and PMAE (To Predict or Not to Predict? Proportionally Masked Autoencoders for Tabular Data Imputation), to provide a more comprehensive evaluation of the proposed framework's competitiveness. 3.The impact of the amount of labeled PU data on the proposed framework has not been sufficiently investigated. Could the authors provide a sensitivity analysis by varying the proportion of labeled positives to evaluate the robustness of the method under different levels of label availability? 4.The proposed framework is built upon a VAE-based imputation model, but the motivation for selecting this architecture is not sufficiently discussed. Could the authors clarify whether the proposed PU integration mechanisms can be extended to other generative architectures, such as diffusion-based or Transformer-based models? 5.The generation of artificial missingness masks in the semi-synthetic experiments is not clearly described. The authors should clarify whether the additional missingness follows an MCAR, MAR, or MNAR mechanism to improve the interpretability and reproducibility of the results.

8 comments
AuthorsOpenReview ↗

Response to Reviewer XsRG (Q4)

Q4: Why was a VAE-based architecture selected, and can PU integration be extended to diffusion-based or Transformer-based models?

The proposed framework is built upon a VAE-based imputation model, but the motivation for selecting this architecture is not sufficiently discussed. Could the authors clarify whether the proposed PU integration mechanisms can be extended to other generative architectures, such as diffusion-based or Transformer-based models?

We appreciate your insightful question. We selected the VAE architecture for its probabilistic formulation, flexible conditioning structure, and tractable training objective, which naturally support our proposed PU integration mechanisms. We have further clarified these motivations in the revised paper.

1. Motivation for selecting VAE.

First, VAE provides a probabilistic generative framework for MNAR imputation. Existing methods, including not-MIWAE [1], GINA [2], and GNR [3], demonstrate the use of latent-variable models to jointly model observed values and missingness masks.

Our framework follows this principle while integrating PU-learned information:

pθ,ϕ(xobs,m∣a,q∗)=∫pθ(xobs∣z,a,q∗)⋅pϕ(m∣z,a,q∗)⋅p(z∣a,q∗),dz.p_{\theta,\phi}(\mathbf{x}^{\mathrm{obs}},\mathbf{m}\mid\mathbf{a},\mathbf{q}^{\ast})=\int p_{\theta}(\mathbf{x}^{\mathrm{obs}}\mid\mathbf{z},\mathbf{a},\mathbf{q}^{\ast})\cdot p_{\phi}(\mathbf{m}\mid\mathbf{z},\mathbf{a},\mathbf{q}^{\ast})\cdot p(\mathbf{z}\mid\mathbf{a},\mathbf{q}^{\ast})\\,d\mathbf{z}.

Second, VAE allows PU-learned measurement-intent probabilities q∗\mathbf{q}^{\ast} to be integrated into multiple components, including the encoder, data decoder, and missingness decoder:

qγ(z∣xobs,m,a,q∗),pθ(xobs∣z,a,q∗),pϕ(m∣z,a,q∗).q_{\gamma}(\mathbf{z}\mid\mathbf{x}^{\mathrm{obs}},\mathbf{m},\mathbf{a},\mathbf{q}^{\ast}),\quad p_{\theta}(\mathbf{x}^{\mathrm{obs}}\mid\mathbf{z},\mathbf{a},\mathbf{q}^{\ast}),\quad p_{\phi}(\mathbf{m}\mid\mathbf{z},\mathbf{a},\mathbf{q}^{\ast}).

This flexible conditioning structure enables PU-learned information to guide both latent representation learning and missing-value reconstruction.

Third, VAE provides a tractable variational training objective that can be extended with auxiliary PU supervision. As shown in Equation (8), our objective is

J=λrecLrec+λMLM+λKLLKL+λPULPU.\mathcal{J}=\lambda_{\mathrm{rec}}\mathcal{L}_{\mathrm{rec}}+\lambda_{M}\mathcal{L}_{M}+\lambda_{\mathrm{KL}}\mathcal{L}_{\mathrm{KL}}+\lambda_{\mathrm{PU}}\mathcal{L}_{\mathrm{PU}}.

Thus, PU-learned probabilities not only serve as conditioning information but also provide auxiliary supervision during VAE training. These three properties make VAE a natural architecture for our framework.

2. Extension to diffusion-based and Transformer-based models.

The proposed PU integration mechanisms are conceptually modular and are not inherently restricted to VAEs. They can be extended to other generative architectures through conditioning and auxiliary supervision.

  • Diffusion-based models: PU-learned probabilities can condition the denoising network, for example, through ϵθ(xt,t,a,q∗)\epsilon_{\theta}(\mathbf{x}_{t},t,\mathbf{a},\mathbf{q}^{\ast}). An additional PU-consistency objective could supervise the reconstructed data representation during denoising.
  • Transformer-based models: PU-learned probabilities can be integrated as feature-level embeddings or additional conditioning tokens in the attention mechanism. An auxiliary prediction head could further encourage the learned representation to preserve measurement-intent information.

These extensions would require architecture-specific designs of the conditioning mechanisms and auxiliary objectives. Our current study focuses on VAE to establish the contribution of PU integration within a probabilistic MNAR imputation framework.

We have also added the following statement to Section 6 (Conclusions) to explicitly highlight this future direction:

Future work may extend the proposed framework to accommodate more complex missingness mechanisms through additional conditioning strategies and alternative generative architectures, including integration into diffusion-based and Transformer-based models.

References:

[1] Niels Bruun Ipsen, Pierre-Alexandre Mattei, and Jes Frellsen. not-MIWAE: Deep Generative Modelling with Missing not at Random Data. In International Conference on Learning Representations, 2021.

[2] Chao Ma and Cheng Zhang. Identifiable Generative Models for Missing Not at Random Data Imputation. In Advances in Neural Information Processing Systems, volume 34, 2021.

[3] Jialei Chen, Yuanbo Xu, Pengyang Wang, and Yongjian Yang. Deep Generative Imputation Model for Missing Not At Random Data. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pages 316–325, 2023.

AuthorsOpenReview ↗

Response to Reviewer XsRG (Q3)

Q3: How robust is PU-VAE-MNAR to different proportions of labeled positive data?

The impact of the amount of labeled PU data on the proposed framework has not been sufficiently investigated. Could the authors provide a sensitivity analysis by varying the proportion of labeled positives to evaluate the robustness of the method under different levels of label availability?

We sincerely appreciate your insightful suggestion to investigate the sensitivity of PU-VAE-MNAR to the availability of labeled positive data.

Following your recommendation, we have added sensitivity analyses with varying known-positive ratios in both simulation and real-world EHR experiments. The results are presented in Table 2 of Section 5.1, Appendix E.1 (Ablation Studies Using Simulation Data), and Appendix F.3 (Ablation Studies).

Specifically, we vary the known-positive ratio from 1% to 40% by relabeling observed positives (S~ij=1\widetilde{S}_{ij}=1) as unlabeled (S~ij=0\widetilde{S}_{ij}=0) under an MNAR mechanism, while keeping the underlying measurement intent SijS_{ij} unchanged. In simulations, we adjust the coefficient of (1−Mj)Sj(1-M_{j})S_{j} in the PU-label generation mechanism (Equation 15). In the real-world EHR data, the probability of relabeling a known positive depends on its corresponding laboratory value.

At extremely low known-positive ratios (1%–5%), imputation RMSE is higher and downstream AUC is lower. Performance improves as the ratio increases to 10%, with smaller gains between 25% and 40%.

For example, in simulation Scenario A, increasing the known-positive ratio from 1% to 40% reduces RMSE from 1.471 to 0.765 and improves AUC from 0.725 to 0.835. Similar trends are observed in Scenarios B and C. In real-world EHR data, RMSE decreases from approximately 0.144 to 0.122, while downstream AUC increases from approximately 0.902 to 0.922 as the known-positive ratio increases from 1% to 40%.

These findings are consistent with the sensitivity analyses in Tables 7 and 8 of the original PAN study [1], which report improved performance as the known-positive ratio increases, with smaller gains at higher ratios. Hu et al. [1] note that with sufficient positive labels, PU-learning performance can approach that of supervised binary classification and depend more on the underlying classifier.

Based on this observation and the diminishing improvements between 25% and 40% in our experiments, we focus the sensitivity analysis on the 1%–40% range, which captures both extreme label scarcity and the transition toward stable performance.

Overall, these experiments demonstrate that PU-VAE-MNAR remains effective when a substantial proportion of known-positive labels are converted into unlabeled instances. The sensitivity to limited positive supervision may also help explain the weaker imputation performance observed under extreme artificial masking rates (95%–99%) in Figure 3, where PU-VAE-MNAR can underperform simple mean imputation. Although these experiments examine different sources of information loss, both highlight the challenges of learning from extremely sparse observations. Nevertheless, PU-VAE-MNAR demonstrates considerable robustness to incomplete positive-label availability once a modest proportion of known positives is retained.

References:

[1] Wenpeng Hu, Ran Le, Bing Liu, Feng Ji, Jinwen Ma, Dongyan Zhao, and Rui Yan. Predictive Adversarial Learning from Positive and Unlabeled Data. Proceedings of the AAAI Conference on Artificial Intelligence, 35(9):7806–7814, 2021.

AuthorsOpenReview ↗

Response to Reviewer XsRG (W3 & Q5)

W3: Missingness generation mechanism in semi-synthetic experiments

3.The missingness generation mechanism in the semi-synthetic experiments is not clearly specified, making it difficult to assess the relevance of the results to MNAR settings.

Q5: Generation of artificial missingness masks

5.The generation of artificial missingness masks in the semi-synthetic experiments is not clearly described. The authors should clarify whether the additional missingness follows an MCAR, MAR, or MNAR mechanism to improve the interpretability and reproducibility of the results.

We thank you for pointing out the need to clarify the missingness generation mechanisms. We have revised Figure 3 to include MNAR masking and added a comparison of all three mechanisms in Appendix F.2 (Semi-synthetic Experiments on Real Data under MCAR, MAR, and MNAR).

In all three settings, we retain the original S~ij\widetilde{S}_{ij} when artificially masking observed laboratory values.

  • MCAR: Values are randomly masked independently of patient characteristics and laboratory values.
  • MAR: Masking probabilities depend on observed covariates, allowing missingness to vary across patients.
  • MNAR: Masking probabilities depend on the laboratory values themselves, with parameters adjusted to achieve the target artificial masking rate rr.

The revised Figure 3 indicates that absolute imputation and prediction performance is generally better under MCAR than MAR, and better under MAR than MNAR. However, the relative improvement of PU-VAE-MNAR over competing methods follows the opposite trend, with the largest improvement under MNAR, followed by MAR and MCAR.

This pattern is consistent with the motivation of our framework, as PU-learned measurement-intent information is particularly useful when missingness depends on the underlying laboratory values. We also recognize that performance under MAR may depend on the observed covariates used to generate the masking probabilities.

In summary, the revised Figure 3 and Appendix F.2 clarify how artificial missingness is generated under MCAR, MAR, and MNAR and provide additional results for evaluating PU-VAE-MNAR across these mechanisms.

Reviewer_DNQzOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

To develop a method that works under MNAR, the authors use positive-unlabeled (PU) learning to model intention-to-collect. The PU-learned probabilities of intention-to-collect are then used to guide the training of variational autoencoders (VAE) that jointly model observed data and missingness indicators. Synthetic and real-world data experiments were conducted to assess the performance of the proposed framework in terms of both imputation and downstream prediction tasks.

Critical strengths

In order to addressing MNAR, the authors attempted to model the intention to collect using PU learning in their proposed approach.

Critical weaknesses

The key claim that this method works under MNAR is not supported by the modeling framework. In particular, the PU learning model in Section 4.1 uses only fully observed features/variables, aia_i, (i.e., excluding any feature/variable with some missing values), while the authors argue that the intention to measure can depend on unobserved variables/features. In other words, the PU learning model should use all features in order for it to work under MNAR. Even under MAR, it is unclear whether the proposed model can use observed values in X.

Similarly, the experimental results do not offer strong support for this key claim. In the simulation study, the model for the measurement intention indicator depends on incomplete X2X_2 or X3X_3 (plus latent variables Z) and does not depend on X1X_1 (which is used as a), so the PU learning model in Section 4.1 would be mis-specified. Plus, imputation errors should be reported. Of note, comparison of the distributions of imputed data (in Figure 1) can be misleading, since it is possible for an imputation method to recover the distribution of incomplete feature while having larger imputation errors for individual missing values.

In the real data and semi-synthetic experiments, while mijm_{ij} is observed, it is unclear how S~ij\tilde{S}_{ij} is defined in EHRs data. Plus, in the semi-synthetic experiments, artificial masking does not depend on variables in the dataset at all, which would be MCAR.

Questions for authors

Q1: Intention-to-collect is only one contributing factor for missing data. There are other reasons for missing values under MNAR, particularly in EHRs. It is unclear whether the proposed framework can address these other factors for MNAR or to what extent the performance of the proposed method is robust to other reasons for MNAR.

Q2: When evaluating the downstream performance in prediction/classification, are there any reasons that the authors did not include end-to-end DL methods that can handle missing data and perform prediction jointly?

9 comments
AuthorsOpenReview ↗

Response to Reviewer DNQz (Q2)

Q2: Why are end-to-end deep learning methods not included in the downstream prediction comparisons?

When evaluating the downstream performance in prediction/classification, are there any reasons that the authors did not include end-to-end DL methods that can handle missing data and perform prediction jointly?

We thank you for highlighting end-to-end deep learning methods. We agree that they provide valuable alternatives for learning from incomplete data. Following your suggestion, we have included two representative methods, GRAPE [1] and NeuMiss [2], in Appendix F.

  • GRAPE: Uses graph neural networks for missing-value imputation through edge prediction and an MLP-based node prediction model for downstream outcomes.
  • NeuMiss: Directly predicts YY from incomplete xobs\mathbf{x}^{\mathrm{obs}} and missingness masks m\mathbf{m}, without explicitly imputing missing values.

Our reasons for adopting a separate imputation and prediction framework are as follows.

1. Distinct research objectives.

Our primary goal is to leverage the predictive utility of indication-driven missingness to improve both MNAR imputation and downstream prediction, rather than optimizing prediction alone.

As illustrated in Figure 1, PU-VAE-MNAR consists of three steps: PU learning, MNAR imputation, and downstream prediction. Specifically, PU-learned measurement-intent probabilities q∗\mathbf{q}^{\ast} guide VAE-MNAR to produce completed data x^\widehat{\mathbf{x}}, which are subsequently used for predicting YY.

This design allows us to evaluate imputation accuracy (RMSE) and predictive performance (AUC) separately. Moreover, x^\widehat{\mathbf{x}} can be retained for subsequent statistical analysis, clinical interpretation, and scientific discovery, rather than serving only as an intermediate representation for prediction.

In the revised paper, we have further clarified our contributions at the end of Section 1 (Introduction):

  • To the best of our knowledge, we are the first to formalize indication-driven missingness through partially observed measurement intent S~\widetilde{\mathbf{S}} as a PU learning problem.
  • We propose PU-VAE-MNAR, integrating PU learning with VAE-based MNAR imputation through weighted auxiliary supervision using q∗\mathbf{q}^{\ast}.
  • We achieve superior performance over competing imputation methods in both MNAR imputation and downstream prediction under an average laboratory missingness rate of 91.9%.

2. Controlled comparison and interpretability.

We use the same downstream LASSO logistic regression model across all imputation methods to isolate the contribution of imputation and ensure consistent comparisons:

\operatorname{logit}\left\\{P(Y_{i}=1\mid\mathbf{a}_{i},\widehat{\mathbf{x}}_{i})\right\\}=\beta_{0}+\boldsymbol{\beta}_{a}^{\top}\mathbf{a}_{i}+\boldsymbol{\beta}_{x}^{\top}\widehat{\mathbf{x}}_{i}.

In contrast, end-to-end methods use different prediction architectures, such as the MLP-based predictor in GRAPE, introducing additional architectural differences.

We select LASSO logistic regression as our primary downstream model for its interpretability and usefulness in statistical inference, particularly in clinical research.

We have also expanded Appendix F (Additional Real Data Results) to compare alternative prediction algorithms, demonstrating that predictive performance varies across model architectures even with identical input predictors.

3. Isolating the contribution of measurement intent.

A central objective of our study is to evaluate whether PU-learned measurement-intent information q∗\mathbf{q}^{\ast} improves MNAR imputation and downstream prediction beyond the observed PU labels S~\widetilde{\mathbf{S}}.

By fixing the downstream model fLASSOf_{\mathrm{LASSO}}, we compare predictions based on completed data from different imputation methods:

Y^i=fLASSO(ai,x^i).\widehat{Y}_{i}=f_{\mathrm{LASSO}}(\mathbf{a}_{i},\widehat{\mathbf{x}}_{i}).

This controlled design allows us to attribute performance differences more directly to the imputation framework and PU integration.

In contrast, end-to-end methods jointly optimize prediction using their own architectures and outcome supervision, making it more difficult to isolate the contribution of measurement intent. Moreover, NeuMiss does not explicitly generate x^\widehat{\mathbf{x}}, preventing direct evaluation of imputation RMSE.

Therefore, we retain the same downstream prediction model for controlled comparisons while including GRAPE and NeuMiss as additional end-to-end deep learning baselines in the revised Appendix F.

References:

[1] You et al. Handling Missing Data with Graph Representation Learning. NeurIPS, 2020.

[2] Le Morvan et al. NeuMiss Networks: Differentiable Programming for Supervised Learning with Missing Values. NeurIPS, 2020.

AuthorsOpenReview ↗

Response to Reviewer DNQz (Q1)

Q1: What is the role of the PU learning step in addressing MNAR missingness?

Intention-to-collect is only one contributing factor for missing data. There are other reasons for missing values under MNAR, particularly in EHRs. It is unclear whether the proposed framework can address these other factors for MNAR or to what extent the performance of the proposed method is robust to other reasons for MNAR.

We appreciate your point that EHR missingness can arise from multiple mechanisms. Our framework focuses on indication-driven missingness as an essentially and clinically meaningful source of missing laboratory data while retaining flexibility for other mechanisms.

  1. Measurement intent is a key driver of missingness in EHR data, with strong clinical motivation for modeling it.

In EHRs, laboratory measurements are often initiated based on symptoms, suspected diseases, and physician judgment [1,2]. Testing decisions may therefore depend on underlying health status and potentially unobserved measurement values, making MNAR plausible. Other sources, including administrative processes and data collection failures, may follow MCAR, MAR, or MNAR mechanisms [3].

Thus, we proposed this framework to leverages this informative measurement process through partially observed measurement-intent indicators.

  1. Empirical contribution of PU-learned information.

As illustrated in Figure 1, EHR data contain measurement-intent information (e.g., Lab 1 and Lab 2) beyond laboratory values and missingness masks, which conventional MNAR imputation methods generally do not explicitly model. Given the challenges of MNAR imputation with unobserved values, ignoring readily available measurement-intent information would be a missed opportunity.

Table 5 shows that using only observed covariates and measurement-intent indicators (a,S~)(\mathbf{a},\widetilde{\mathbf{S}}) achieves an AUC of 0.921 without laboratory-value imputation. Table 6 further demonstrates the contribution of PU-guided imputation:

ModelAUCAUCPR
Observed (a,S~)(\mathbf{a},\widetilde{\mathbf{S}})0.9210.835
VAE-MNAR0.9130.816
PU-VAE-MNAR0.9360.864

VAE-MNAR without PU information achieves an AUC of 0.913, even lower than the observed-input model (0.921). Integrating PU learning improves AUC to 0.936 and AUCPR to 0.864, demonstrating the complementary predictive value of PU-learned measurement intent.

  1. Extensibility to other missingness mechanisms.

Our framework does not assume that measurement intent is the sole cause of missingness. We adopt a VAE-based generative model for its flexible probabilistic factorization, allowing additional information and missingness mechanisms to be integrated. As discussed in Section 2 (Related Work — Deep Generative Models for MNAR Imputation):

VAE-based methods more directly model MNAR missingness through latent-variable formulations. not-MIWAE uses selection models (Little & Rubin, 2019) to jointly model observed values and missingness, GINA extends this framework with an identifiable latent-variable model, PSMVAE models ignorable and nonignorable missingness through pattern-set mixtures, and GNR jointly reconstructs data values and missingness masks.

PU-VAE-MNAR builds on this flexibility by integrating PU-learned information through conditioning and auxiliary supervision. Additional missingness-related factors can be integrated through extended conditioning variables or dedicated modeling components. GINA [4] illustrates related flexibility through different missingness structures in the graphical models of its Figure 1.

Thus, while our implementation focuses on indication-driven missingness, the framework provides a flexible foundation for modeling additional missingness mechanisms when relevant information is available.

References:

[1] Denis Agniel, Isaac S. Kohane, and Griffin M. Weber. Biases in electronic health record data due to processes within the healthcare system: retrospective observational study. BMJ, 361:k1479, 2018.

[2] Amelia L. M. Tan, Emily J. Getzen, Meghan R. Hutch, et al. Informative missingness: What can we learn from patterns in missing laboratory data in the electronic health record? Journal of Biomedical Informatics, 139:104306, 2023.

[3] Rose Sisk, Lijing Lin, Matthew Sperrin, Jessica K. Barrett, Brian Tom, Karla Diaz-Ordaz, Niels Peek, and Glen P. Martin. Informative presence and observation in routine health data: A review of methodology for clinical risk prediction. Journal of the American Medical Informatics Association, 28(1):155–166, 2021.

[4] Chao Ma and Cheng Zhang. Identifiable generative models for missing not at random data imputation. In Advances in Neural Information Processing Systems, volume 34, 2021.

AuthorsOpenReview ↗

Response to Reviewer DNQz (W3)

W3: What is the missingness mechanism for the artificial masking rate rr in the semi-synthetic experiments?

In the real data and semi-synthetic experiments, while mijm_{ij} is observed, it is unclear how S~ij\widetilde{S}_{ij} is defined in EHRs data. Plus, in the semi-synthetic experiments, artificial masking does not depend on variables in the dataset at all, which would be MCAR.

We thank you for pointing out the need to clarify the definition of S~ij\widetilde{S}_{ij} in real-world EHR data and semi-synthetic experiments.

1. Definition of S~ij\widetilde{S}_{ij} in EHR data.

In the revised paper, we have added Appendix B.1 (EHRs Data Structures) and Table 8 to illustrate how S~ij\widetilde{S}_{ij} is constructed from EHR laboratory records.

Table 8(a) presents a typical encounter-level EHR laboratory table containing patient IDs, encounter IDs, laboratory codes, and test results. For each patient ii and laboratory variable jj, we aggregate records over the 12-month observation window and define:

S~ij={1,if a laboratory record exists,0,otherwise.\widetilde{S}_{ij}=\begin{cases}1, & \text{if a laboratory record exists}, \\\\ 0, & \text{otherwise}.\end{cases}

The observation mask mijm_{ij} separately indicates whether xijx_{ij} is available. Therefore, a laboratory record without a result gives (S~ij,mij)=(1,0)(\widetilde{S}_{ij},m_{ij})=(1,0) and belongs to the known-positive set P\mathcal{P}. When no laboratory record exists, S~ij=0\widetilde{S}_{ij}=0, and the entry belongs to the unlabeled set U\mathcal{U}, where the underlying measurement intent SijS_{ij} may be either 0 or 1.

Table 8(b) provides patient-level examples, and Table 8(c) explains the corresponding interpretations.

Table 8. Measurement intent and value availability in EHR data.

(a) Encounter-level laboratory records

PatientEncounterLabResult
P01E11Lab 1138
P01E12Lab 2Pos
P02E21Lab 1NaN
P02E21Lab 2Neg
P03E31Lab 2Pos
P04E41Lab 2Neg

(b) Observed data (Lab 1)

PatientMeasurement intentxijobsx_{ij}^{\mathrm{obs}}SijS_{ij}S~ij\widetilde{S}_{ij}mijm_{ij}
P01Observed138111
P02ObservedNaN110
P03UnknownNaN100
P04UnknownNaN000

SijS_{ij} is latent. The values of SijS_{ij} for P03 and P04 illustrate the two unlabeled cases in Table 1 of the manuscript.

(c) Interpretation

PatientInterpretation
P01Laboratory record with a numerical result. Measurement intent and xijx_{ij} are observed.
P02Laboratory record without a numerical result. Measurement intent is observed but xijx_{ij} is missing.
P03Unrecorded positive intent (Sij=1S_{ij}=1). xijx_{ij} was intended to be measured but ultimately remained missing.
P04True absence of intent (Sij=0S_{ij}=0). xijx_{ij} was not intended to be measured.

2. Semi-synthetic missingness mechanisms.

We thank you for your comment. We have revised Figure 3 to include MNAR masking and added a comparison of all three missingness mechanisms in the section Appendix F2. Semi-synthetic Experiments on Real Data under MCAR, MAR, and MNAR.

In all three settings, we retain the original S~ij\widetilde{S}_{ij} when artificially masking observed laboratory values.

  • MCAR: Values are randomly masked independently of patient characteristics and laboratory values.
  • MAR: Masking probabilities depend on observed covariates, allowing missingness to vary across patients.
  • MNAR: Masking probabilities depend on the laboratory values themselves, with parameters adjusted to achieve the target artificial masking rate rr.

The results in Figure 3 indicate that absolute imputation and prediction performance is generally better under MCAR than MAR, and better under MAR than MNAR. However, the relative improvement of PU-VAE-MNAR over competing methods follows the opposite trend, with the largest improvement under MNAR, followed by MAR and MCAR. This pattern is consistent with the motivation of our framework, as PU-learned measurement-intent information is particularly useful when missingness depends on the underlying laboratory values. We also recognize that performance under MAR may vary depending on the observed covariates used to generate the masking probabilities.

In summary, Appendix B.1, Table 8, and the revised Figure 3, Appendix F. 2 clarify how S~ij\widetilde{S}_{ij} is defined in real-world EHR data and how the three artificial missingness mechanisms are generated in the semi-synthetic experiments.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers

We thank all reviewers for their professional review and valuable opinions and suggestions. Regarding the questions you raised, we have provided detailed answers one by one in our response for your review.

We also appreciate the reviewers’ recognition of the strengths of our work. Reviewer DNQz noted our use of PU learning to model intention-to-collect. Reviewer XsRG highlighted the meaningful perspective offered by distinguishing measurement intent from value availability, together with the predictive benefits of PU-derived information. Reviewer 38um highlighted the novelty of the indication-driven formulation, the scale and realism of our EHR cohort, the explicit handling of continuous and binary laboratory variables, and the evaluation of both imputation and downstream prediction against diverse baselines.

Furthermore, we have uploaded the revised paper, and here we summarize the updated content.

Reviewer DNQz

  • In Sections 1 and 4.1 and the new Figure 1, we have clarified how PU-learned measurement-intent information guides VAE-MNAR through conditioning and auxiliary supervision.
  • In Sections 5.1 and 5.2, we have clarified how observed covariates provide partial information about measurement intent through shared latent factors.
  • In Section 5.1 and Figure 2, we have added Wasserstein distance, imputation RMSE, and downstream AUC, with metric definitions in Appendix D.2.
  • In Section 3.2 and Appendix B.1, we have clarified the construction of measurement-intent labels from EHR records, with examples in the new Table 8.
  • In Section 5.2.2 and Figure 3, we have added MNAR masking results. Appendix F.2 provides definitions and comparisons of MCAR, MAR, and MNAR masking.
  • In Appendix F.6 and Table 17, we have added GRAPE and NeuMiss comparisons. Section 4.2 clarifies the choice of LASSO, with additional prediction-model comparisons in Appendices F.4–F.5.

Reviewer XsRG

  • In Section 1, we have explicitly stated our conceptual and methodological contributions. The new Figure 1 illustrates the complete framework.
  • In Tables 2 and 6, we have expanded the ablations to distinguish PU conditioning from auxiliary supervision and compare PAN with logistic regression, GBM, and nnPU using RMSE and AUC. Additional results appear in Appendix E.1.
  • In Tables 3–4 and Figure 3, we have added ReMasker, PMAE, and DiffPuter, increasing the imputation baselines from 10 to 13. Their implementation details are provided in Appendices D.5–D.6.
  • In Table 2 and Appendices E.1 and F.3, we have added sensitivity analyses across known-positive ratios of 1%–40%, evaluating both imputation and prediction.
  • In Figure 3 and Appendix F.2, we have clarified the artificial missingness mechanisms and compared performance under MCAR, MAR, and MNAR.
  • In Section 2, we have explained the rationale for selecting VAE. Section 6 discusses potential extensions to diffusion-based and Transformer-based models.

Reviewer 38um

  • In Section 2, we have strengthened the practical motivation for informative-missingness modeling, with supporting literature on selection bias, statistical power, privacy, and data governance.
  • In Section 5.1 and Figure 2, we have added ECE against ground-truth measurement intent, alongside Wasserstein distance, imputation RMSE, and downstream AUC.
  • In Tables 2 and 6, we have added classifier and PU-integration ablations to justify PAN. Appendix D.4 specifies training-set-only class-prior estimation for nnPU.
  • In Table 2 and Appendices E.1 and F.3, we have added sensitivity analyses across known-positive ratios of 1%–40%.
  • In Appendix F.1, Tables 11–14, we have added 95% confidence intervals for the imputation, prediction, and ablation results corresponding to Tables 3–6.
  • In Section 6, we have discussed future extensions to more complex missingness mechanisms and alternative generative architectures.

We hope that these revisions and the accompanied point-by-point responses could further address your concerns. Thank you again for your diligent effort in reviewing and constructive suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.