acceptodds
Under review as a conference paper at ICLR 2027

PAMF: Prior-Aware Multimodal Fusion for Incomplete Time Series Data

Abstract

Multimodal time-series tasks often operate on incomplete observations in practice. For example, in healthcare, ECG segments are lost because electrodes detach or an entire respiratory channel is unavailable during overnight monitoring. Such missingness typically appears in two structurally distinct patterns: within-modality missing, where values are absent within an otherwise observed modality, and modality-level missing, where an entire modality is unavailable. Existing methods typically represent unobserved data implicitly through masks or missing embeddings, without learning instance-specific missing information, and most are designed for only one missingness pattern. A natural approach is to explicitly estimate the missing data; however, existing imputation methods treat missingness uniformly despite their different structural priors, and the imputation process is often isolated from downstream tasks, preventing downstream tasks from guiding imputation toward more informative representations. To address these limitations, we present PAMF, a multimodal time-series framework that explicitly handles different missingness patterns while coupling imputation with downstream prediction through prior-aware flow matching and task-informed representation transfer. Specifically, the method initializes the flow-matching source state with type-specific priors to distinguish the two missing types. It further connects imputation and classification through task-informed representation transfer between architecturally matched encoders. Experiments on multiple multimodal healthcare time-series benchmarks, real native-missingness clinical datasets, and explicit reconstruction comparisons show that the proposed method achieves the strongest overall downstream performance across the evaluated datasets and missingness settings compared with existing baselines.

Primary area:
learning on time series and dynamical systems
Keywords:
Multimodal Time Series, Missing Data, Flow Matching, Imputation, Foundation Models, Incomplete Multimodal Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_D3rUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PAMF, a method that first imputes missing signals in multimodal time series before performing classification. For localized missing values, it leverages nearby observations; for cases where an entire modality is missing, it begins with the channel mean and incorporates existing signals, completing the imputation via conditional flow matching. To harness information from the classification task, the authors first train a classification encoder and copy its weights to the imputation encoder. Once the imputation model is trained, it is frozen, and the classifier is then trained to adapt to the imputed data.

Critical strengths
  1. The experiments have a broad scope. The paper is evaluated on four benchmarks with controlled missing data and two clinical datasets containing naturally occurring missing data, covering scenarios such as partial missingness, complete modality absence, mixed missingness, and high rates of missing data.
Critical weaknesses
  1. The training and model-selection protocols for the main classification baselines are insufficiently documented, and the additional computational cost remains unclear. The paper provides a detailed account of the three-stage training configuration and checkpoint selection criteria for PAMF, but it fails to specify the corresponding training epochs, hyperparameter selection processes, and model selection criteria for key baselines such as Maestro, FuseMoE, Flex-MoE, and MIRA. Consequently, it is currently difficult to verify whether these methods were appropriately trained and tuned, which also hinders the reproducibility of the experiments. In addition, PAMF introduces an additional completion encoder, flow-matching module, and multi-stage training, but the computational efficiency comparison only includes the diffusion completion method and does not report its total training cost relative to the direct classification baseline and end-to-end inference latency. The absence of this information undermines the credibility of the primary performance comparisons.

  2. Lacks comparison with closely related flow-based completion methods. The paper discusses flow-based imputation methods such as CLWF, MIRI, and CFMI in related work, but the direct completion experiment only compares CSDI and SSSD. Although the "Related Work" section outlines how these methods differ from PAMF in terms of their conceptual positioning, it fails to explicitly explain why certain methods were deemed unsuitable for the current evaluation framework or why they were excluded from the comparison. Given that CFMI and CLWF have already been evaluated for imputation tasks using multivariate clinical time-series data from PhysioNet, they should be considered relevant, representative baselines.

Questions for authors
  1. What training and model selection protocols were used for the primary baselines? Please specify the number of training epochs (or stopping criteria), the hyperparameter selection process, and the checkpoint selection criteria for Maestro, FuseMoE, Flex-MoE, and MIRA. This information is essential for determining whether the reported gains are built upon baselines that have been sufficiently trained and reasonably tuned.

  2. Why were flow-based completion methods with clinical time-series experience not included? The paper discusses CFMI and CLWF, both of which have been evaluated in experiments involving the imputation of multivariate clinical time series. Could you clarify why these methods were not included in the direct imputation comparison? Could you incorporate a suitable representative method into the matching data partitioning, preprocessing, missing masking, and downstream classification training protocol to compare reconstruction quality, classification performance, and inference cost?

  3. How much additional end-to-end computational cost does PAMF incur compared to direct classification methods? Please report the total training time and end-to-end inference latency for PAMF versus direct-classification baseline, ensuring that both hardware and batch size are matched. These results will help determine the additional computational cost incurred for any gains in classification performance, rather than simply assessing whether PAMF is faster than the chosen diffusion-based completer.

  4. What specific factors are altered by "No Prior" and "No Transfer" ablations? Please clarify the specific implementation of the two ablation studies shown in Figure 3. For "No Prior," what source state is used in place of the proposed initialization? For "No Transfer," are the pre-training of the prediction branch, the fusion encoder, the classification head initialization, and the subsequent training pipeline all kept the same, with only the copying of weights to the completion encoder omitted?

  5. Do the normalization and source-state statistics rely solely on information available at inference time? For datasets that employ in-sample or in-window normalization, are the mean and standard deviation calculated before or after the application of the missing-value mask? Are the artificially hidden values ​​included in the calculation of these statistics?

3 comments

Dear D3rU,

We thank you for your careful review and for recognizing the broad range of missingness settings in our experiments. We address your concerns and questions below.

W1/Q1: Baseline training and model selection

We retain the baselines’ core fusion mechanisms, report their training configurations, and select checkpoints by validation Macro-F1. All methods use an 80% / 10% / 10% training / validation / test split. The baseline configurations are listed in our response to Reviewer icex (W1/W4/Q1). We have clarified the baseline hyperparameter settings in the Experiments section of the revised manuscript.

For PAMF, as stated in the appendix, we select the imputation checkpoint by the lowest validation RMSE and the final prediction checkpoint by the highest validation F1.

W1/Q3: Training and inference cost

Compared with direct classification using the same backbone, PAMF raises Macro-F1 on PTB-XL from 0.605 to 0.641, at about 3.3 times the training time and 4.2 times the inference latency. On PTB-XL under 20% mixed missingness, No Imputation uses the same classification backbone to classify incomplete inputs directly. It achieves Macro-F1 of 0.605 ± 0.022, while PAMF achieves 0.641 ± 0.011.

MethodTraining time (min) ↓Inference latency (ms/batch) ↓
No Imputation7.03 ± 0.3817.87 ± 1.56
PAMF23.54 ± 0.4775.07 ± 7.72

Training times are summarized over seeds 42, 52, and 62. Inference latency was measured on an NVIDIA A6000 GPU with a batch size of 64.

W2/Q2: Comparison with a flow-based imputer

PAMF achieves lower reconstruction MSE (0.477 vs. 0.690) and higher Macro-F1 (0.641 vs. 0.634) than CFMI, a flow-matching imputer, under the same PTB-XL protocol. We include CFMI as an additional flow-matching baseline on PTB-XL under 20% mixed missingness. The table also includes the CSDI, SSSD, and PAMF results reported in Table 4.

MethodReconstruction MSE ↓Macro-F1 ↑NFE ↓Latency (ms/batch) ↓
CSDI1.148 ± 0.0040.626 ± 0.01650137.77 ± 6.28
SSSD0.782 ± 0.0140.616 ± 0.00950526.98 ± 23.75
CFMI0.690 ± 0.0620.634 ± 0.00620996.81 ± 10.48
PAMF0.477 ± 0.0040.641 ± 0.0112075.07 ± 7.72

Inference latency was measured on an NVIDIA A6000 GPU with a batch size of 64.

Although CFMI uses 20 function evaluations, its latency is higher in the reported measurements. One possible contributing factor is the cost of each network evaluation: the CFMI implementation uses four Transformer layers following Tashiro et al., each computing attention over both the temporal and lead dimensions.

Q4: Definitions of No Prior and No Transfer

No Prior uses a zero source, while No Transfer omits only the copying of pretrained prediction-encoder weights to the imputation encoder.

  • No Prior: We replace the missingness-specific prior initialization with a zero source.
  • No Transfer: We omit copying the pretrained prediction encoder weights to the imputation encoder. Prediction-branch pretraining, the fusion encoder, classification-head initialization, and subsequent training remain the same; weight transfer is the only change.

We have clarified these definitions in Section 4.5 of the revised manuscript.

Q5: Normalization and source-state statistics

On PTB-XL, each complete ECG window is normalized per lead before synthetic masking; this preprocessing is identical for PAMF and all baselines. PAMF's source prior itself uses only values that remain visible after masking.

After masking, PAMF constructs its source prior using visible values: the temporal prior uses visible neighboring observations, and the modality-level prior uses the corresponding channel statistics from samples in the current batch where that modality is observed.

Thank you for the detailed rebuttal and additional experiments. I still have a concern about normalization. In practice, the missing signal values would not have been measured, but normalizing the complete test signal before masking uses those values. Could this make the test results overly optimistic? Could you retrain and evaluate PAMF and a strong baseline on PTB-XL under mixed missingness, using normalization that does not rely on hidden values, to see whether PAMF’s advantage remains?

Thank you for raising this concern.

PAMF's advantage over MAESTRO remains when normalization uses only training-set statistics: PAMF achieves higher Macro-F1 for all three seeds, with an average gain of 3.17 percentage points.

We retrained and evaluated PAMF and MAESTRO on PTB-XL under 20% mixed missingness, applying the missingness mask before normalization and using statistics computed only from the training set. Thus, masked test values do not contribute to the normalization of test inputs.

MethodMacro-F1 ↑AUROC ↑
PAMF0.637 ± 0.0050.869 ± 0.009
MAESTRO0.605 ± 0.0240.866 ± 0.005

Results are mean ± sample standard deviation over seeds 42, 52, and 62. These results show that PAMF's Macro-F1 advantage persists under this normalization protocol.

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

PAMF is a flow-matching imputer for multimodal time series, covering both gaps inside an observed channel and whole modalities absent. Since conditional flow matching does not need a Gaussian source, the source state is initialized from a prior chosen by missingness type: interpolation from temporal neighbours where a channel still has observations, a channel statistic where the modality is gone. Imputation is then coupled to the task in three stages. The claim is that this beats missing-aware fusion methods on four masked benchmarks and two clinical cohorts with native gaps, and reconstructs PTB-XL better than the diffusion imputers at fewer function evaluations.

Critical strengths

The source initialization distinguishes two kinds of missingness that prior imputers treat alike, and it is a concrete design choice: the two priors are specified, and the paper ablates them against learned and retrieval-based alternatives. The clinical cohorts carry the strongest evidence here: Challenge2012 and MIMIC-III are naturally incomplete rather than synthetically masked, native gaps are kept out of imputation supervision, and the MIMIC-III margins are large relative to the three-seed deviations.

Critical weaknesses

What the paper offers is a set of margins over baselines, and only one side of those comparisons has a reported configuration. Table 9 is captioned as the configuration used for PAMF, and Figure 4 sweeps its expert count and width. For the four fusion baselines there is no learning rate, no capacity, no training budget and no statement that published settings were used, and Appendix F describes each of them in prose without a number. Appendix K does state a matched split, normalization and protocol for the reconstruction comparison against CSDI and SSSD. The sweep also settles the final configuration: Appendix H.2 reports performance strongest around four experts and best at width 128, and says those values were chosen on that basis. The sweep runs under mixed 20% missingness, which is itself a reported evaluation condition. So one side had its capacity picked at a peak on a benchmark the paper scores, and the other side has no reported configuration at all, which is what makes a margin such as 0.033 on PTB-XL hard to read as evidence for the design.

On both ECG benchmarks, part of what is masked as a missing modality is determined by the leads that remain. PTB-XL groups the twelve leads into Limb, Augmented and Precordial; Chapman-Shaoxing treats each lead as its own modality. The augmented leads are fixed combinations of I and II, so with the Limb group retained the augmented signals are recoverable by algebra, and per-lead normalization rescales without removing that dependence. Masking the Augmented group therefore tests recovery of redundant measurements rather than loss of an independent sensor. This limits how far these two datasets support the missing-sensor motivation.

Equation 6 also closely resembles the decay imputation of GRU-D, a convex combination of a neighbouring observed value and that channel's mean weighted by an exponential decay in the gap, and GRU-D is not cited anywhere. The differences I can see are the nearest observed entry in place of the last one, a per-channel batch mean in place of a training-set mean, and a fixed decay in place of a learned one. Since the type-conditioned source is half of what is new here, where the within-modality half comes from matters.

A baseline configuration table, or a statement that published settings were used, would let me judge whether these comparisons are controlled.

Questions for authors

Q1. How did you choose the baseline hyperparameters and training budgets, and is there a configuration table for them that I missed? Relatedly, how were these four fusion baselines selected? ShaSpec and CLIMB are named twice, only as methods the clean-input table excludes, and I could not find them compared anywhere else.

Q2. With Limb retained, recovering the Augmented group is an algebraic consequence of the observed leads rather than a sensor-loss problem, so how much of the modality-level gain depends on that redundancy? A non-redundant missing group would help assess it.

1 comment

Dear Reviewer icex,

We thank you for your careful review and for recognizing the type-conditioned source design and the clinical evaluations. We address your concerns and questions below.

W1/Q1: Baseline selection and configurations

We retain the baselines’ core fusion mechanisms and report their training settings. For PAMF, the number of experts and the expert width were chosen once on PTB-XL and then kept fixed, without retuning, for every other dataset and missingness setting. For MAESTRO, FuseMoE, Flex-MoE, and MIRA, we followed the key hyperparameter settings reported in their respective original papers. We have clarified these baseline settings in the Experiments section of the revised manuscript.

For PAMF, Figure 5 examines the number of experts and expert width under 20% mixed missingness on PTB-XL. We selected four experts and a width of 128 using validation AUROC, then used these values across the other datasets and missingness settings without retuning. For PAMF and all baselines, the training, validation, and test split is 80% / 10% / 10%; baseline checkpoints are selected by validation Macro-F1.

We use the key model-specific design choices from the original methods. The baselines are evaluated using shared data splits, preprocessing, missingness settings, and evaluation metrics. Training settings are adapted where needed; for example, batch sizes may differ from those in the original papers to improve training efficiency. The table below reports the learning rate, training epochs, and batch size used for each baseline. We select the baseline checkpoints using validation Macro-F1. These training batch sizes are separate from the batch size used for imputation timing.

BaselineLearning rateTraining epochsTrain batch size
MAESTRO1e-3200128
FuseMoE1e-33032
Flex-MoE1e-45064
MIRA3e-43016

Our baseline set covers different approaches to incomplete multimodal inputs: expert-based fusion in FuseMoE and Flex-MoE, modality-aware attention and dropout in MAESTRO, and channel-independent time-series modeling in MIRA. MAESTRO has previously been evaluated against ShaSpec in the MAESTRO paper: It reports DaliaHAR F1 of 0.84 for MAESTRO and 0.78 for ShaSpec under complete modalities, and Figure 5 compares them under modality dropout. MIRA provides a clinical time-series comparator relevant to CLIMB, which is a benchmark encompassing multiple encoders.

W2/Q2: ECG lead redundancy

PAMF's modality-level gains do not rely on algebraic redundancy alone: the precordial group (V1–V6) cannot be derived from the limb leads, and PAMF also achieves the highest F1 and AUROC under modality-level missingness on Sleep-EDF and PPG-DaLiA, which have no such redundancy (Table 2). As discussed in our response to Reviewer KWDV (Q4–Q5), PAMF learns relationships among observed ECG leads without explicit lead-derivation formulas, and Q5 shows reconstruction examples with the entire precordial group masked. Our evaluation also includes two naturally incomplete clinical datasets (Table 5). The aggregate ECG results do not separately quantify the contribution of algebraically related leads.

W3: Relation to GRU-D

We agree that Equation 6 follows GRU-D's decay-to-mean form. We now cite GRU-D and do not claim this interpolation format as a contribution; it supplies the within-modality source state for conditional flow matching [1]. Equation 6 shares the decay-to-mean idea with GRU-D: both combine an observed value with a channel mean using a decay weight. GRU-D uses the last observation, a training-set mean, and a learned decay, whereas our default rule uses the nearest observed value, a batch mean, and a fixed decay.

PAMF uses missingness-specific structural priors to initialize conditional flow matching: the temporal rule provides the source for within-modality gaps, while a channel statistic initializes entirely missing modalities. The resulting flow is conditioned on observed inputs and combined with task-informed representation transfer for downstream prediction. We have cited GRU-D and clarified this distinction in the Method section of the revised manuscript.

Reference

[1] Che et al. Recurrent Neural Networks for Multivariate Time Series with Missing Values. Scientific Reports, 2018.

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

The paper addresses multimodal physiological time-series classification with missing inputs, distinguishing between two scenarios: gaps within individual modalities and entirely absent modalities. The proposed PAMF method first fills missing entries using rule-based initialisations, a blend of nearest-neighbour and mean estimates for within-modality gaps, and the batch mean for absent modalities, and then refines these estimates using a conditional flow-matching network conditioned on the observed modalities. The conditioning encoder is initialised from a classifier pre-trained on incomplete inputs, and the classifier is subsequently fine-tuned on the imputed signals. Experiments on four benchmarks with synthetic masking (Sleep-EDF, PTB-XL, PPG-DaLiA, and Chapman) and two cohorts with native missingness (PhysioNet 2012 and MIMIC-III) report higher F1 scores and AUROC than missing-aware fusion baselines, alongside lower reconstruction MSE and latency than CSDI and SSSD on PTB-XL.

Critical strengths

S1. Broad and consistent empirical evaluation. The method is evaluated on six datasets, including two with native missingness, across within-modality, modality-level, mixed, high-rate, MAR, and MNAR missingness protocols, using three random seeds throughout. Consistent improvements over the baselines across all settings establish the pipeline as a strong practical reference.

S2. Simple and efficient design. The pipeline is straightforward to reproduce and faster at inference than diffusion-based imputers (Table 4). The results also show that staged training outperforms joint optimisation (Table 6), a useful finding in its own right.

Critical weaknesses

W1. The experiments do not isolate the contribution of the proposed imputation method from that of the classification backbone. PAMF achieves the best overall results, and the effectiveness of the full pipeline is clear. However, several findings suggest that a substantial share of the improvement may stem from the backbone, per-modality Informer encoders coupled with a cross-modal MoE, and staged training, rather than prior-aware flow matching. When paired with the same classifier, simpler imputation methods, including the rule-based initialisation alone and CSDI, already outperform all baselines (Figure 3 and Table 4). The model also performs best on complete inputs, where no imputation is required (Table 11). Furthermore, the gains attributable to the proposed components are small relative to the variability across seeds (Figure 3). The component ablations are limited to a single dataset and setting (PTB-XL, mixed missingness, F1 only), while the key “No Prior” variant is undefined and appears only as a bar in Figure 3. Overall, the experiments establish the effectiveness of the pipeline but do not sufficiently demonstrate that its claimed contributions drive the observed improvements.

W2. The flow-matching formulation does not support generative imputation, which may explain its limited contribution. The source state x0x_0 is a deterministic function of the observed data, with no added noise. Consequently, the source distribution in Equation 1 is a point mass, and the ODE produces a single output for each input. At the optimum of the CFM objective, the learned field transports x0x_0 along a straight path to E[x1∣xobs]\mathbb{E}[x_1 \mid x_{\mathrm{obs}}], regardless of the chosen prior, so a single Euler step would suffice. The module thus acts as a residual regressor targeting the conditional mean. It cannot represent uncertainty and instead averages over plausible signals when the posterior is broad, as may occur with entirely missing modalities. Moreover, all datasets are z-score normalised, making the modality-level batch-mean prior close to zero and devoid of sample-specific information. This is consistent with the identical performance of the batch-mean and batch-matched priors in Table 10. Given that prior-aware initialisation is presented as the main contribution, these limitations warrant explicit discussion. They also affect the interpretation of Table 4, since a conditional-mean estimator is expected to achieve lower MSE than sampling-based imputers.

Questions for authors

Q1. Attribution. Could you compare No Imp., Prior Only, CSDI paired with your classifier, and PAMF across all datasets, or at least on Chapman and MIMIC-III, where the performance margins are largest? Could you also evaluate the strongest baselines using your classification backbone? These comparisons would have the greatest impact on my assessment.

Q2. Stochasticity. Is any noise added to x0x_0 during training or inference? If not, how is the distributional formulation in Equation 1 justified? Could you also report results using a single integration step (NFE = 1)?

Q3. Prior controls. How is the “No Prior” variant in Figure 3 defined? Could you include Gaussian-source controls using both a single draw and the mean of KK draws, alongside a zero-source control? Please report reconstruction MSE for each variant, separately for within-modality and modality-level missingness. For Table 4, how many samples were drawn for CSDI and SSSD, and could you also report CRPS?

Q4. Benchmark design. In PTB-XL, the augmented leads are linear combinations of the limb leads; in Chapman, leads III, aVR, aVL, and aVF can be derived from leads I and II. These missing modalities are therefore fully recoverable when the required leads remain observed. Could you report modality-level results separately for each dropped group? Also, because the modality-level prior is a batch statistic, does a test sample’s prediction depend on the composition of its batch?

Q5. Qualitative results. Could you provide example reconstructions, particularly for entirely missing modalities, to illustrate whether the imputed signals are realistic or reflect averaging over plausible signals?

6 comments
AuthorsOpenReview ↗

Response to KWDV (1/2)

Dear Reviewer KWDV,

We thank you for your careful evaluation and for recognizing the breadth of our experiments and the simplicity of our design. We address your concerns below.

W1/Q1: Contribution of task-oriented imputation and the classification backbone

With the backbone fixed, task-oriented imputation improves Macro-F1 on PTB-XL and Chapman. On Chapman, the gain over Prior Only exceeds the largest standard deviation across seeds.

Classification backbone. Under 20% mixed missingness on PTB-XL, our backbone without imputation (No Imputation) obtains Macro-F1 of 0.605 ± 0.022, compared with 0.608 ± 0.009 for MAESTRO, the strongest baseline. With complete inputs (Table 11), our backbone has the highest F1 on Sleep-EDF and PTB-XL, while MAESTRO performs best on PPG-DaLiA. The backbone is competitive, but does not alone explain the gains under missingness.

Value of imputation. With the same backbone, Prior Only raises F1 from 0.605 ± 0.022 to 0.629 ± 0.009; CSDI paired with our classifier obtains 0.626 ± 0.016 (Table 4). These results support imputation for classification under missing inputs.

Value of task-oriented imputation. PAMF transfers task-informed representations to the imputation encoder and adapts the classifier to completed inputs. On PTB-XL, PAMF reaches 0.641 ± 0.011 F1, improving by 1.2 points over Prior Only and 1.5 points over CSDI, and outperforming No Transfer (Figure 4). As these gains are comparable to seed variation, we also evaluated the fixed-backbone comparison on Chapman under 20% mixed missingness.

Results on Chapman. PAMF has the highest F1 for each seed, improving the mean over Prior Only by 2.0 points, and the highest mean AUROC. Results are mean ± sample standard deviation over seeds 42, 52, and 62.

Chapman test set, mixed 20%Macro-F1 ↑AUROC ↑
No Imputation0.487 ± 0.0050.818 ± 0.010
Prior Only0.499 ± 0.0080.819 ± 0.006
PAMF0.519 ± 0.0140.826 ± 0.004

W2/Q2: Stochasticity of the source, the learned flow, and NFE = 1

The source shapes the learned flow; Equation 1 does not require a Gaussian source. In our evaluation, one integration step does not match the 20-step result.

No noise is added to the default source during training or inference. For a fixed observed input, mask, batch, and inference realization, it is a point initialization. Its push-forward through the flow is a point distribution corresponding to the resulting completion. Equation 1 describes this model-induced distribution and does not require a non-degenerate Gaussian source. Generalized conditional flow matching permits non-Gaussian sources [2]. We have clarified the source and flow in Section 3 of the revised manuscript.

Integration steps. We evaluated the same trained model with NFE = 1 and NFE = 20 on PTB-XL under 20% mixed missingness:

PTB-XL, mixed 20%Reconstruction MSE ↓Macro-F1 ↑Latency (ms/batch) ↓
NFE = 10.619 ± 0.0030.584 ± 0.00531.01 ± 1.92
NFE = 200.477 ± 0.0040.641 ± 0.01175.07 ± 7.72

Training uses the linear bridge xt=(1−t)x0+tx1x_t=(1-t)x_0+tx_1 and target velocity ut=x1−x0u_t=x_1-x_0. The network regresses these targets across examples toward v∗(x,t,c)=E[x1−x0∣xt=x,c]v^*(x,t,c)=\mathbb{E}[x_1-x_0\mid x_t=x,c] (Lipman et al. [1]). A straight path to the conditional mean describes the exact optimum under the deterministic-source setting; the trained field can depart from it. In our evaluation, 20 steps improve both reconstruction and classification over one step.

Role of the prior. Changing x0x_0 changes the intermediate states xtx_t and velocity targets x1−x0x_1-x_0 used in training. Replacing the missingness-specific prior with a zero source lowers F1 from 0.641 ± 0.011 to 0.619 ± 0.007 (Figure 4). For entirely missing modalities, the batch mean of z-scored data carries little sample-specific information: batch-matched initialization gives similar F1 (0.650 ± 0.002 vs. 0.648 ± 0.002 for the default; Table 10). Sample-specific information comes from the observed modalities that condition the velocity field.

Reference

[1] Lipman et al., Flow Matching for Generative Modeling, ICLR 2023. https://arxiv.org/abs/2210.02747

[2] Tong et al., Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport, TMLR 2024. https://arxiv.org/abs/2302.00482

AuthorsOpenReview ↗

Response to KWDV (2/2)

Q3: Source controls, reconstruction MSE, and sampling

PAMF achieves the lowest reconstruction MSE with one completion, even when competing imputers average five. PAMF also has the highest mean Macro-F1 in the five-completion comparison, while CSDI, SSSD, and CFMI obtain lower five-sample CRPS.

Figure 4's No Prior replaces the missingness-specific initialization with a zero source. The Table 4 results use one completion per test record (K = 1). We evaluated retrained PAMF, source controls, CSDI, SSSD, and CFMI on PTB-XL under 20% mixed missingness. For reconstruction MSE, PAMF uses its one draw (K = 1); source controls and other imputers are evaluated using both their first completion and the pointwise mean of five completions. Macro-F1 is evaluated on the mean of five completed waveforms for each method.

Reconstruction MSE and downstream classification. Macro-F1 is reported as mean ± sample standard deviation across inference seeds 42, 52, and 62.

MethodWithin MSE K=1 ↓Within MSE K=5 ↓Modality MSE K=1 ↓Modality MSE K=5 ↓Mixed MSE K=1 ↓Mixed MSE K=5 ↓Macro-F1 K=5 ↑
PAMF0.567—0.308—0.478—0.645 ± 0.001
Zero source0.9720.9721.0091.0090.9750.9750.615 ± 0.000
Gaussian source1.1151.0031.1381.0271.1201.0060.610 ± 0.002
CSDI1.4290.8260.6250.3821.1610.6780.628 ± 0.010
SSSD0.9100.7200.5430.4350.7830.6230.630 ± 0.005
CFMI0.9420.5810.5190.3300.7840.4890.640 ± 0.015

The zero-source Macro-F1 standard deviation is nonzero but rounds to 0.000 at three decimal places.

CRPS ↓. We compute five-sample CRPS from the five individual completions, before averaging. CSDI, SSSD, and CFMI obtain lower CRPS than PAMF across the three missingness groups.

MethodWithinModalityMixed
PAMF0.3900.2690.349
Zero source0.5970.6310.600
Gaussian source0.5210.5440.523
CSDI0.3540.2120.307
SSSD0.3400.2500.309
CFMI0.2800.1990.252

PAMF is designed for task-oriented completion. Its single-completion reconstruction MSE and five-completion mean classification performance remain strong when competing methods also average five completions, while CSDI, SSSD, and CFMI achieve lower five-sample CRPS.

Q4: ECG lead structure and batch size

Under modality-level missingness on PTB-XL, PAMF stays within 0.9 F1 points of its complete-input performance without being given any lead-derivation formula. A prediction depends on its batch through the modality-level prior, but performance is stable across batch sizes 32 to 128 (Macro-F1 0.636 to 0.644).

ECG lead structure. ECG leads contain physiological and algebraic relationships that provide useful structure for multimodal completion. PAMF is not given explicit lead-derivation formulas; it learns to use the remaining observed modalities to produce a completion.

On PTB-XL, PAMF obtains 0.657 ± 0.005 Macro-F1 with complete inputs and 0.648 ± 0.002 under modality-level missingness. Table 2 reports the aggregate modality-level results; the submitted results do not provide a separate breakdown for each dropped lead group.

Batch-size sensitivity. The modality-level prior uses batch statistics, so its initialization can vary with batch size. Across batch sizes 32, 64, and 128, reconstruction and classification remain stable. Batch size 128 achieves slightly lower mean MSE and higher mean Macro-F1; the reported configuration uses batch size 64.

Batch sizeMSE ↓Macro-F1 ↑AUROC ↑
320.487 ± 0.0010.636 ± 0.0030.866 ± 0.001
640.477 ± 0.0040.641 ± 0.0110.867 ± 0.003
1280.475 ± 0.0010.644 ± 0.0020.867 ± 0.002

Q5: Qualitative reconstructions

In both selected examples, PAMF more closely follows the ground-truth waveforms than CSDI and SSSD. We have added these examples as Figure 2 in Section 4 of the revised manuscript. In both examples, the entire precordial modality (V1–V6) is masked. The figure compares the ground truth with reconstructions from PAMF, CSDI, and SSSD using the same masked input. Each row shows one of the six leads, with amplitudes expressed as per-record, per-lead z-scores. PAMF follows several sharp peaks in V2–V6, illustrating how it preserves waveform morphology under modality-level missingness. Both examples were selected for their lower PAMF reconstruction error relative to CSDI and SSSD.

Thank you for clarifying that the default source contains no added noise and that deterministic inference produces one completion for a fixed observed input and batch. This clarifies the implementation, but my concern about the generative interpretation remains. Generalized CFM permits non-Gaussian sources; however, a deterministic flow maps a point initialization to another point and therefore cannot represent multiple plausible completions under fixed conditioning information. Equation 1 can describe this degenerate model distribution, but it does not establish that the model captures the conditional distribution of missing signals. I acknowledge that NFE = 20 improves substantially over NFE = 1. This supports the practical value of iterative refinement, while leaving the uncertainty-modelling concern unresolved. I would also qualify my earlier statement that one Euler step must suffice: this is not generally guaranteed by the state-dependent CFM regression objective. Please clarify definitively whether PAMF models the conditional variability of missing signals or produces a single task-oriented completion. With the observed input, mask, and batch fixed, how can deterministic initialization and inference represent multiple plausible completions?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.