acceptodds
Under review as a conference paper at ICLR 2027

EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction

Abstract

Diagnostic information in multi-view medical imaging studies, like echocardiography, is distributed across views, each capturing different aspects of cardiac structure and function. In clinical practice, these views are first interpreted through quantitative measurements, which then support specific diagnoses. However, existing echo representation learning methods do not explicitly model measurements as an intermediate representation between views and diagnosis. As a result, individual measurements are not explicitly represented in a form that downstream diagnostic tasks can selectively use, hindering accurate diagnosis. Inspired by this hierarchical clinical workflow, we introduce EchoTab, a measurement-grounded representation learning framework that represents multi-view echocardiography studies using measurement-specific tokens as an intermediate representation between echo views and diagnosis. Each measurement token selectively attends to the views relevant to its estimation and is trained through tabular reconstruction to predict the corresponding measurement value. The resulting tokens form a factorized echo representation, which encode clinically meaningful structural and functional properties and can be selectively used for downstream diagnosis. Unlike the commonly used contrastive approaches, EchoTab does not require negative samples or large batch sizes. Experiments on two datasets demonstrate that EchoTab learns effective measurement-grounded representations that improve diagnosis classification in multi-view echocardiography and its learned attention aligns with established measurement-view relationships used by cardiologists. Our code is accessible at https://anonymous.4open.science/r/EchoTab.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multimodal Learning with Tabular Data, Representation Learning, Echocardiography, Medical AI

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_R6rqOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces EchoTab, a representation learning method for Echocardiography that leverages key clinical measurements. EchoTab consists of several components: an echo encoder that transforms echocardiography clips into echo tokens, learnable measurement queries, a tabular decoder that connects the measurement queries with their corresponding relevant views through a cross-attention mechanism, and a prediction head. Experiments on MIMIC-IV-Echo v1.0 and an internal dataset show that EchoTab outperforms the baselines.

Critical strengths
  • The idea of using Echo-related measurements as intermediate representations is very well motivated, as it reflects the clinical workflow where diagnosis is conducted by considering multiple information, including measurements.
  • EchoTab enables diagnosis-specific selection of relevant measurements, which could be helpful in clinical settings where Echocardiography view scans are incomplete.
  • Experiments on both public and private data demonstrate EchoTab’s superiority over the baselines, and the ablation study shows the contribution of each of EchoTab's components and design considerations.
  • The cross-attention map analysis shows that EchoTab is able to mimic cardiologists on the measurement-view relevancies to some extent, which demonstrates that measurement-grounded representations align with well-established clinical rules.
  • The paper is well written and organized, with some caveats in atuhors’ claims (see weaknesses).
Critical weaknesses
  • Abstract > "Each measurement token selectively attends to the views relevant ..." > made me wonder the first time reading the paper. Consider adding something like "through cross-attention".
  • In the introduction, the authors say "A natural alternative is to use measurements directly as the target modality for contrastive learning" right after mentioning EchoPrime. However, I believe this claim is not well supported, as the authors utilize EchoPrime as EchoTab's view/clip encoder. Also there was no experiment comparison between the contrastive learning as in EchoPrime and EchoTab's measurement-grounded representation learning. I strongly suggest revising and lowering this claim, as I believe the experiments are inadequate to support such a claim.
  • The authors say "EchoTAB consistently outperforms echo foundation models". However this could be that because they were not pre-trained on the MIMIC-IV-Echo v1.0 data distribution, while EchoTab is trained on that distribution. Thus, I believe the authors should revise and lower down the claim. Also, it is worth comparing the Echo foundation models with EchoTab trained on MIMIC and tested on the private set, or vice versa.
  • No statistical analysis of the experimental results is provided in the main text.
Questions for authors
  • What is the reasoning behind sampling randomly clip during training while deterministically selecting the first clip during inference, as mentioned in Sec. 2.1?
  • Did the authors try to compare using Eq. 5 with simple mean-std normalization of measurements?
  • Table 2: Did the authors run any experiment where the Echo encoder is randomly initialized and trained end-to-end using the proposed measurement-grounded representation learning?
  • How does EchoTab’s performance on measurement prediction compare to PanEcho?
10 comments
AuthorsOpenReview ↗

(W1) Clarification of "cross-attention" in the abstract

We have now added "cross-attention" to this sentence in the Abstract to "Each measurement token selectively attends to the views relevant to its estimation through cross-attention, and is trained through tabular reconstruction to predict the corresponding measurement value."

AuthorsOpenReview ↗

(W2+W3) Comparison of EchoTab and echo foundation models

We agree that the phrase "A natural alternative..." could be misinterpreted as an unsupported claim that tabular-based learning is strictly superior to EchoPrime's text-based contrastive learning. Before introducing our own approach, we simply intended to explain how EchoPrime could be adapted to use measurements instead of text reports as supervision during contrastive learning, following established image-tabular contrastive methods.

We also agree that the phrasing "EchoTab consistently outperforms echo foundation models" could be misinterpreted as claiming our pre-training method is strictly superior to the ones used to train the foundation models without accounting for domain shift. As indicated in Table 1, the foundation models were included only as reference, not baselines, as they were not re-trained on our datasets. Because we initialized EchoTab and the other baselines with one such foundation model (EchoPrime), we included them to demonstrate that through echo-tabular pre-training, EchoTab and the other baselines improve upon the original EchoPrime weights.

To prevent any misunderstanding, we have softened these claims and have made the following changes in the manuscript:

  • Introduction (Section 1): We revised "A natural alternative is to use measurements directly as the target modality for contrastive learning" to: "Following image-tabular contrastive methods (Hager et al., 2023), EchoPrime's echo-text alignment could be adapted to echo-tabular alignment by directly using the tabular measurements instead of text reports as the target modality."

  • Table 1 Caption: We revised "EchoTab consistently outperforms echo foundation models, as well as study- and view-level contrastive and MLP baselines…" to "EchoTab consistently outperforms study- and view-level contrastive and MLP baselines…"

  • Diagnosis Classification (Section 3.3): We revised "EchoTab consistently improves diagnosis classification across both datasets and achieves the highest average AUC among echo foundation models, contrastive baselines, and the MLP baseline." to "EchoTab achieves the highest average AUC across diagnoses among the contrastive and MLP baselines on both datasets."

To rule out the possibility that EchoTab’s performance gains over echo foundation models are only due to dataset-specific training on the target distributions instead of the actual measurement-grounded learning itself, we further trained (FT) the echo foundation models EchoPrime and PanEcho on our datasets according to their original pre-training objectives. Since EchoPrime’s echo-text contrastive objective requires text reports for training, we generated synthetic text reports from the measurement data without any diagnostic labels. This guarantees that EchoPrime (FT) receives the same measurement-only supervision as EchoTab. We then evaluated these models in both in-dataset and cross-dataset settings on diagnosis classification. We have added this experiment to App B.3 (now App B.2).

The reported metric is average AUC across diagnoses in each dataset. A→B denotes that the model was trained on measurement data from dataset A and evaluated on dataset B.

ModelPre-training ObjectiveIn-Dataset: MIMICIn-Dataset: InternalCross-Dataset: Internal→MIMICCross-Dataset: MIMIC→Internal
EchoPrime–0.6590.730––
PanEcho–0.7020.702––
EchoPrime (FT)Echo-text contrastive0.6890.7070.6520.729
PanEcho (FT)Tabular reconstruction0.7130.7340.6630.729
EchoTab (Ours)Tabular reconstruction0.7480.7780.7030.766

As shown above, further training EchoPrime with its original echo-text contrastive objective yields mixed results. EchoPrime (FT) improves on MIMIC (0.659→0.689) but degrades on the internal dataset (0.730→0.707). Meanwhile, PanEcho (FT) improves on both MIMIC (0.702→0.713) and on the internal dataset (0.702→0.734), consistently benefiting from training on the target data distribution.

Meanwhile, in both in-dataset and cross-dataset evaluations, EchoTab outperforms EchoPrime (FT) and PanEcho (FT). This confirms that EchoTab's performance gains come from our measurement-grounded representation learning framework, rather than mere exposure to the target data distribution.

AuthorsOpenReview ↗

(W4) Statistical analysis of the experimental results in Table 1

We added a paired bootstrap analysis to compare average AUCs between EchoTab and all methods in Table 1. Specifically, we resample the test set with replacement 1,000 times, resulting in 1,000 sets of test studies that are shared across methods and diagnoses. For each method, we compute the improvement in average AUC of EchoTab over that method and report its 95% percentile CI and two-sided bootstrap p-value, with p-values corrected using the Holm-Bonferroni procedure. Baselines significantly outperformed by EchoTab are marked with * in Table 1, with CIs and p-values reported in the Appendix B.4 (previously Appendix B.4, now Appendix B.3).

On MIMIC, the improvement of EchoTab in average AUC is statistically significant against all compared methods. On the internal dataset, it is statistically significant against all methods except the MLP baseline (for which the 95% confidence interval narrowly includes zero).

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

The paper proposes EchoTab, a measurement-grounded representation learning framework that represents multi-view echocardiography studies using measurement-specific tokens as an intermediate representation between echo views and downstream diagnoses. During training, EchoTab extracts view embeddings from the available views in an echo examination. These embeddings are then aggregated through QKV cross-attention in a tabular decoder, producing measurement-factorized representations that are passed through measurement-specific prediction heads and supervised using the available clinical measurement labels (using normalized MSE). For downstream diagnosis tasks, the measurement prediction heads are discarded, and clinically relevant measurement-factorized representations are selected based on predefined expert knowledge and mean-pooled to produce a study-level representation for diagnosis prediction.

EchoTab achieves average diagnostic AUCs of 0.778 on an internal dataset and 0.748 on MIMIC-echo, showing better results than the compared echo foundation models and contrastive baselines.

Critical strengths

The paper is generally clear and easy to read, and the experimental evaluation is thorough, providing comprehensive comparisons with existing methods and ablation studies to assess the effectiveness of the proposed components.

Critical weaknesses
  1. Terminology and definition. First, the fact that the model is named EchoTab, along with the mention of tabular reconstruction/embedding/decoder, can be misleading, as tabular data usually refers to a type of input data. Here, tabular refers to the regression-type outputs, so the model is essentially a multi-output regression model. This makes it feel as though the paper is intentionally reframing existing concepts without any de facto changes, adding a layer of confusion for readers. The same goes for the naming of the queries. While they serve as Q in QKV attention, they should be better named or explicitly recognised as conditioning tokens that steer the representations towards measurement-specific counterparts.
  2. Paper organization. Scattered information for the "Diagnosis classification" stage. Fig. 1- (2) is not explained in the methodology section but in the Experiment section (Sec. 3.2). How the relevant measurement tokens are selected is in Sec 6 Limitations - by the experts, presumably.
  3. Contradiction in the conclusion from Tables 1/6/7. Table 6 serves as a result demonstrating that other compared methods do not perform as well as EchoTab even when the relevant view information is used, although the 2nd best MLP reaches 0.717 from 0.697 on MIMIC. This is quite close to the result in Table 7 where EchoTab drops from 0.748 to 0.725, making it marginally better than the MLP. The authors wrote: "This suggests that the view measurement relationships learned by the tabular decoder are more effective than a defined set of relevant views, either because the decoder weights the relevant views more effectively or because it extracts complementary information from views not considered relevant to a given diagnosis. In practice, ECHOTAB can therefore be applied directly to all views of a study, without any per-diagnosis view curation." First, please verify which one is the reason. Then, does this suggest the contradiction that there should not be a relevant measurement token selection in the process depicted in Fig.2?
  4. For the other baselines, my understanding from Tables 1/6/7 is that, since they were not trained with the measurement labels, selecting the relevant views for a downstream task does not necessarily help them much. Therefore, another confounding factor here is that incorporating the measurement labels during training may be the reason for the performance gain. The authors should therefore carefully examine whether the contribution comes from the design of intermediate measurement tokens or from incorporating the auxiliary measurement information during training. If it is the latter, then it reduces the potential significance and perceived novelty.
Questions for authors

Please see critical weaknesses 3 and 4.

8 comments
AuthorsOpenReview ↗

(W1) Terminology

We thank the reviewer for the constructive feedback regarding our terminology. Below, we clarify each terminology:

  • Tabular. We adopted the term "tabular" because measurements are a form of tabular data. This usage is consistent with prior vision-tabular pre-training frameworks, such as MMCL [1] and TGV [2], which we use as our baselines. Given that EchoTab extends this line of vision-tabular representation learning to echocardiograms, we maintained this terminology to clearly situate our work within the broader literature.

  • Tabular reconstruction. We used the term "tabular reconstruction" to clearly differentiate how tabular data is used as supervision in EchoTab compared to other vision-tabular pre-training methods, such as MMCL/Contrastive (where tabular data is used as an alignment target) and TGV/SupCon (where tabular data is used as guidance to define positive and negative pairs for contrastive learning) [see "Role of tabular data" in Table 1]. We agree with the reviewer that "tabular reconstruction" is a multi-output regression task. We have now clearly stated the connection between tabular reconstruction and multi-output regression in Section 2.4: "To ground each measurement token xmx_m in measurement-specific information, we train the prediction head to reconstruct the corresponding measurement values. We use a masked, variance-normalized mean squared error for this multi-output regression task to account for the sparsity and varying scales of echo measurements."

  • Tabular embedding. "Tabular embedding" refers to an embedding of all tabular measurements belonging to an echo study, produced by a tabular encoder (e.g., TARTE [3]). This term appears only when we describe the Contrastive baseline. EchoTab itself does not compute any tabular embedding.

  • Tabular decoder and measurement queries. This terminology is inspired by DETR [4], where learnable object queries attend to the encoder output in a transformer decoder to "decode" individual objects from image features. Analogously, in EchoTab, learnable measurement queries attend to the echo encoder output in the tabular decoder to "decode" individual measurements from echo views.


References

[1] Paul Hager, Martin J. Menten, and Daniel Rueckert. Best of both worlds: Multimodal contrastive learning with tabular and imaging data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 23924-23935, June 2023.

[2] Marta Hasny, Maxime Di Folco, Keno Bressem, and Julia Schnabel. Tables guide vision: Learning to see the heart through tabular data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1716-1725, 2026.

[3] Myung Jun Kim, Félix Lefebvre, Gaëtan Brison, Alexandre Perez-Lebel, and Gaël Varoquaux. Table foundation models: on knowledge pre-training for tabular learning. Transactions on Machine Learning Research (TMLR), 2025. ISSN 2835-8856.

[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pp. 213-229, 2020.

AuthorsOpenReview ↗

(W2) Paper organization of the details on diagnosis classification and measurement token selection

To improve clarity, we have made the following updates to the manuscript:

  • We moved the explanation about how diagnosis-specific study embeddings are generated for diagnosis classification from Section 3.2 ("Baselines and Evaluation Strategy") to a new Section 2.5 ("Diagnosis-Specific Study Embedding for Diagnosis Classification").

  • We clarified that the relevant measurement tokens per diagnosis are defined by a cardiologist. This is now stated in the caption of Fig. 1 (2) ("EchoTab’s measurement-factorized tokens enable the construction of different diagnosis-specific study embeddings from the same echo study by aggregating cardiologist-defined measurement tokens relevant to each downstream diagnosis.") and in the new Section 2.5 ("Specifically, for diagnosis kk, let MkM_k denote the set of measurements considered relevant to its prediction, as defined by a cardiologist.").

AuthorsOpenReview ↗

(W3) View selection vs. measurement token selection

We thank the reviewer for this insightful observation. To address both questions, we first clarify that EchoTab performs diagnosis through two steps:

  1. View selection (input): which echo views are given to the model
  2. Measurement token selection (output): which of the learned measurement tokens are averaged for diagnosis classification

Table 7 varies only the input views while keeping the measurement tokens fixed to those relevant to the target diagnosis.

To answer the first question, Table 7 already supports the fact that EchoTab’s tabular decoder learns effective view-measurement relationships by extracting information from both diagnosis-relevant views and complementary views not conventionally considered relevant to a given diagnosis. When the input to EchoTab is restricted to only a pre-defined set of relevant views, its performance drops (0.748 → 0.725 on MIMIC, 0.778 → 0.753 on internal). We agree that the original statement lacks precision and have revised it accordingly in Appendix B.2 to "This indicates that EchoTab's tabular decoder extracts information from both diagnosis-relevant views and complementary views not conventionally considered relevant to a given diagnosis."

As the reviewer correctly pointed out, EchoTab’s performance drops to almost the same as MLP’s when both models are restricted to relevant views (0.725 vs. 0.717 on MIMIC). However, EchoTab's best setting (which uses all views), outperforms MLP's best setting (which uses only relevant views) (0.748 vs. 0.717 on MIMIC). Moreover, MLP's benefit from restricting its input to relevant views on MIMIC does not generalize to the internal dataset, where its performance decreases from 0.761 (when using all views) to 0.746 (when using only relevant views). In contrast, EchoTab consistently performs better when using all views than when restricted to relevant views across both datasets. Therefore, EchoTab does not require diagnosis-specific view selection to achieve its best performance.

To answer the second question, the results above do not suggest a contradiction regarding the need for measurement selection. The fact that EchoTab does not need diagnosis-specific view selection does not negate the need for diagnosis-specific measurement token selection. As mentioned above, view selection and measurement selection represent two different stages in EchoTab:

  1. Views (input): As shown in Table 7, EchoTab benefits from seeing all views, extracting information from diagnosis-relevant as well diagnosis-irrelevant views. In clinical practice, echo acquisitions are often imperfect and incomplete so EchoTab might use information from diagnosis-irrelevant views as context or fallback option when some of the diagnosis-relevant ones are not available.
  2. Measurements (output): As shown in Table 3, EchoTab benefits from the selection of diagnosis-relevant measurement tokens. In clinical practice, diagnoses are defined by specific measurements according to clinical guidelines (e.g., LVEF for systolic dysfunction) so including irrelevant measurements in the average adds noise instead.
Reviewer_PUP9OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a method to estimate multiple echocardiography measurements from multiple cardiac views (videos). It first runs an MViT encoder on each video and concatenates the output tokens. Then, it assigns learnable queries for the measurements and cross-attends them with the encoder output to create what the authors call “measurement-factorized tokens”. These factorized tokens are then fed to regression heads; of specific note is a regression head based on a hypernetwork—namely, a shared MLP tuned differently by the hypernetwork for each measurement query. A loss is then applied by matching the predicted measurements to the ground truth. During inference, the paper proposes pathology classification based on the factorized tokens corresponding to the measurements relevant to the pathology. They performed an experimental evaluation on two datasets: one public (MIMIC-Echo) and another internal.

Critical strengths

– The paper proposes an interesting tabular approach based on queries to predict echocardiography measurements, comparing it, e.g., to standard MLP regression or image-text contrastive learning such as EchoPrime. It experimentally demonstrates the advantage of this query-based approach over others, albeit on a specific and relatively small set of known and predefined measurements.

– The hypernetwork is an interesting addition. While I understand the benefit of dynamic weight generation by the hypernetwork model, I wonder if this is truly a critical architectural component.

Critical weaknesses

– I am struggling to assess the technical novelty. The authors use known components of query-based learning and apply them to echocardiography tasks. The same applies to the multi-measurement head based on the hypernetwork. Thus, it feels more like an application paper than a fundamental ML contribution.

[more Minor] – In general, the relatively small number of echocardiographic measurements (a few dozen), many of which are interrelated, makes it difficult to fully evaluate the impact of the proposed tabular model. I am well aware that this reflects the available echo data, but it remains difficult to test this approach at scale.

Questions for authors

Have you evaluated a shared head with the same number of parameters as the hypernetwork head (i.e., the partially shared head)?

4 comments
AuthorsOpenReview ↗

(W1) Technical novelty and contribution

While EchoTab builds on existing components (query-based decoders, hypernetwork-generated LoRA updates), our contribution lies not in these individual components, but in a new formulation of measurement-grounded representation learning: we factorize a multi-view echo study into measurement tokens, which can be selectively combined to construct measurement-grounded, diagnosis-specific representations of the same echo study. This design mirrors clinical practice, where the measurements relevant to a diagnosis vary across diagnostic tasks.

As shown in Table 1, EchoTab's diagnosis-specific study representations improve diagnostic accuracy over the generic study-level representations produced by prior vision-tabular methods, such as MMCL [1] and TGV [2]. Furthermore, in Table 2, we investigate how to best generate measurement tokens for constructing diagnosis-specific study representations by comparing different prediction head variants.

We included the multi-measurement head based on the hypernetwork (partially shared head) not to propose a technical novelty. Since EchoTab already has a separate query for each measurement, we wanted to see whether there is any benefit to sharing only part of the head’s parameters across measurements through the generation of per-measurement weights by a hypernetwork, providing a middle ground between the shared head and independent heads. As such, the partially shared head is one of the possible architectural variants of the prediction head, and our contribution lies in the analysis of these variants.

We have revised the manuscript to make our contributions more explicit, emphasizing that the measurement-token factorization in EchoTab enables the construction of different diagnosis-specific study representations from the same echo study. Specifically, we made the following updates:

  • Abstract: Added the sentence "By selectively combining the measurement tokens relevant to each diagnostic task, EchoTab constructs diagnosis-specific study embeddings, allowing the same echo study to be represented differently for different diagnoses."

  • Introduction (Section 1): Added the sentence "This factorization enables the construction of diagnosis-specific study embeddings by selectively combining the measurement tokens relevant to each diagnosis. Therefore, the same echo study can be represented differently depending on the diagnostic task, rather than being reduced to a single, fixed study-level representation." and revised the second contribution list to include "This enables the construction of different diagnosis-specific study embeddings of the same echo study by combining clinically relevant measurement tokens."

  • Method (Section 2): Added a new Section 2.5 titled "Diagnosis-Specific Study Embedding for Diagnosis Classification".

  • Table 1 Caption: Revised the caption from "EchoTab’s measurement-specific representations enable accurate diagnosis for multi-view echocardiography." to "EchoTab's measurement-grounded, diagnosis-specific study representations enable accurate diagnosis for multi-view echocardiography studies."

  • Ablation (Section 4.2, Table 3): Explicitly framed the measurement-token-selection as an evaluation of the importance of selecting diagnosis-relevant measurement tokens for constructing these diagnosis-specific study embeddings.

  • Discussion (Section 6): Highlighted the broader implications of this formulation ("Rather than representing a multi-view echo study with a single, fixed embedding, this formulation factorizes an echo study into measurement-specific tokens, enabling the construction of different diagnosis-specific study representations by combining relevant measurement tokens depending on the diagnostic task.").


References

[1] Paul Hager, Martin J. Menten, and Daniel Rueckert. Best of both worlds: Multimodal contrastive learning with tabular and imaging data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 23924-23935, June 2023.

[2] Marta Hasny, Maxime Di Folco, Keno Bressem, and Julia Schnabel. Tables guide vision: Learning to see the heart through tabular data. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 1716-1725, 2026.

AuthorsOpenReview ↗

(W2) Impact of EchoTab to larger datasets

We agree that the number of measurements available in echocardiography datasets is limited and that many of them are clinically correlated. Nevertheless, our results indicate that the learned measurement tokens encode distinct information rather than collapsing into a few shared factors.

First, the cross-attention map of the tabular decoder (Fig. 3) shows that correlated measurements still attend to different echo views. For example, the measurements la_4ch_length, la_dimen, and la_vol describe left atrium dimensions and are therefore correlated, yet EchoTab attends to different views for each of them:

  • la_4ch_length → A4C
  • la_dimen → PLAX
  • la_vol → A2C & A4C

This suggests that each token captures the view-specific information relevant to its own measurement.

Second, the choice of measurement tokens in constructing diagnostic-specific study representations matters for diagnosis (Table 3). Averaging all measurement tokens instead of only the diagnosis-relevant ones degrades the average AUC in diagnosis classification (0.748 to 0.693 on MIMIC, 0.778 to 0.751 on internal). Moreover, choosing incorrect measurement tokens for diagnosis also degrades performance (0.687 on MIMIC, 0.757 on internal). If the tokens mostly encode redundant information, the choice of tokens would have little effect.

Finally, we evaluated EchoTab on two datasets with different numbers of measurements (M=45 for MIMIC, M=67 for internal) and observed consistent improvements over all baselines in both datasets. We acknowledge that testing the framework with a larger number of measurements is not possible with the currently available echo data and have now added this to the Limitations in Section 6 ("Furthermore, the relatively small number of measurements in our two datasets, many of which are clinically correlated, limits our ability to fully assess how our framework scales to larger and more diverse measurement sets.").

AuthorsOpenReview ↗

(Q1) Comparison of shared vs. partially shared heads with matched number of parameters

To match the number of parameters of the 2-layer partially shared head, which additionally contains a hypernetwork, we train a 4-layer shared head. Furthermore, we also train a 4-layer partially shared head to see whether the benefit of the partially shared head over the shared head at 2 layers (as shown in Table 2) holds for 4 layers as well.

Head# MLP layers# paramsMIMIC Diag. clf.MIMIC Meas. pred.Internal Diag. clf.Internal Meas. pred.
(1)Partially shared head21.75M0.7480.4180.7780.445
(2)Shared head41.77M0.7650.4180.7710.448
(3)Partially shared head42.93M0.7500.3900.7750.438

Diagnosis classification is reported in average AUC across diagnoses. Measurement prediction is reported in average R2{R}^2 across measurements.

With a matched number of parameters (rows 1 & 2), the shared head performs better on diagnosis classification on MIMIC (0.765 vs. 0.748), while the partially shared head performs slightly better on the internal dataset (0.778 vs. 0.771). These mixed results suggest that part of the performance gain of the partially shared head over the shared head in Table 2 is attributable to capacity, rather than from partial sharing of the parameters.

With a matched number of MLP layers (rows 2 & 3), the shared head performs better on MIMIC (0.765 vs. 0.750), while the partially shared head performs slightly better on the internal dataset (0.775 vs. 0.771). Thus, the results are mixed across datasets, suggesting that given sufficient capacity, the shared head can match or even outperform the partially shared head.

Due to the mixed results between shared head and partially shared head, we soften our original claim that "a partially shared head makes the resulting measurement tokens most effective" in Section 4.1 & Table 2 to "...sharing parameters in the prediction head makes the resulting measurement tokens effective for the construction of diagnosis-specific study embeddings by aligning them in a common representation space." We have also added the experiment above to Appendix G.2.

The choice of the prediction head still matters. As shown in Table 2, with measurement queries, independent heads perform worse than both shared and partially shared heads despite having much more parameters (~26.6M). However, we do not think this is due to overfitting because measurement queries with independent heads also underperform the single query with independent heads (0.709 vs. 0.720 on MIMIC, 0.764 vs. 0.766 on the internal dataset), even though both models have essentially the same number of parameters. A more likely explanation is that in the former setting, the measurement tokens are not mapped into a common representation space. In the latter, this issue does not arise because there is only a single token from which all the heads read. In contrast, the shared and partially shared heads enforce the measurement tokens to be mapped to a common representation space.

Dear Reviewer PUP9,

Thank you for your thoughtful feedback on our submission.

We have posted a detailed response to your comments and updated the manuscript:

  • W1: We revised the manuscript to state our contributions more clearly. All the changes are listed in our response.
  • W2: We added a statement regarding EchoTab's scalability to the Limitations (Section 6).
  • Q1: We added an experiment in Appendix G.2 comparing a shared head with a partially shared head at a matched parameter count, and softened our claim about the partially shared head in Section 4.1 & Table 2.

These changes and additions are highlighted in blue in the revised manuscript.

As the discussion period ends on Wednesday, we wanted to kindly check whether our response has addressed your concerns. If anything remains unclear, we would be glad to provide further clarification or additional results.

Thank you for your consideration!

Best regards,

The Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Revised Manuscript

We thank the reviewers for the constructive feedback, which has substantially improved the paper. We especially thank Reviewer ozeT for their engagement and for raising their score, and Reviewer R6rq for confirming that our responses addressed their concerns and for maintaining their positive rating. We welcome further feedback from Reviewer PUP9 on whether our responses and the revised manuscript address their concerns. We are happy to answer any remaining questions before the discussion period ends.

We summarize our main contributions and revisions below.

Main contributions


  • Measurement-grounded representation learning for echocardiography. Reviewers found measurement tokens as intermediate representations between echo views and diagnosis well-motivated (R6rq). They also found the query-based approach interesting and noted its advantage over MLP and image-text contrastive learning such as EchoPrime (PUP9).
  • Diagnosis-specific, clinically aligned study representations. Reviewers highlighted that EchoTab enables diagnosis-specific selection of relevant measurements (R6rq). The cross-attention map shows that EchoTab mimics to some extent cardiologists' judgments of which echo views are relevant to each measurement, which shows that the measurement-grounded representations align with established clinical rules (R6rq).
  • Thorough evaluation with consistent gains on two datasets. Reviewers considered the experimental evaluation thorough, with comprehensive comparisons to existing methods and ablation studies (ozeT, R6rq). Furthermore, EchoTab outperforms the baselines (PUP9, R6rq) on both public and private datasets (R6rq).

Key revisions


  1. Terminology and contributions (PUP9 W1; ozeT W1+W2). We clarified that our key contribution lies in the new formulation of measurement-grounded representation learning: the factorization of a multi-view echo study into measurement tokens, which enables the construction of diagnosis-specific study representations (Abstract, Introduction, new Sec. 2.5 in Method, Fig. 1 (2) caption, Table 1 caption, Sec. 4.2/Table 3, Discussion in Sec. 6). We also made explicit the connection between tabular reconstruction and multi-output regression in Sec. 2.4.
  2. Source of EchoTab’s performance gain.
  • Not merely due to exposure to the target data (R6rq W2+W3). We further trained EchoPrime and PanEcho on our echo-measurement datasets using their original pre-training objectives. EchoTab outperforms both models in both in-dataset and cross-dataset evaluations (App. B.2).
  • Not merely due to measurement supervision (ozeT W4). All vision-tabular baselines were trained on exactly the same measurement data as EchoTab (Sec. 3.2). EchoTab also outperforms echo foundation models further trained on this data (App. B.2).
  • Not merely due to domain-knowledge prior (ozeT follow-up comment 2). In EchoTab, domain knowledge is used only downstream to select relevant measurement tokens for each diagnosis, not during training. When each method is given one domain-knowledge prior (diagnosis-relevant measurements for EchoTab, diagnosis-relevant views for baselines), EchoTab still performs best (Table 6, App. B.1). When domain-knowledge prior is unavailable, a data-driven prior could serve as an alternative (Sec. 6).
  1. Statistical analysis of the main results in Table 1 (R6rq W4). We added a paired bootstrap analysis to compare avg. AUCs between EchoTab and all methods. EchoTab's improvement is significant against all methods on MIMIC, and against all except MLP on the internal dataset (Table 1, App. B.3).

Additional experiments


  1. Comparison of shared vs. partially shared head in EchoTab at matched parameter count in App. G.2 (PUP9 Q1). As results are mixed across datasets, we softened our original claims regarding the partially shared head in Sec. 4.1 & Table 2. Both shared head variants outperform independent heads, showing that sharing parameters in the head aligns the measurement tokens in a common representation space, making them effective for the construction of diagnosis-specific study embeddings.
  2. Ablation of the target formulation in App. G.3 (R6rq Q2). Using unnormalized measurements with variance normalization in the loss (Eq. 5) performs better on MIMIC and comparably on the internal dataset than using mean-std-normalized measurements without variance normalization in the loss.
  3. Comparison with PanEcho on measurement prediction in App. I.1 (R6rq Q4). EchoTab achieves higher R2R^2 both when PanEcho is used off-the-shelf and when it is further trained on our datasets.

Additional clarifications


  1. Impact of EchoTab at scale (PUP9 W2). We acknowledge in the Limitations in Sec. 6 that our analysis of EchoTab is limited to the relatively small number of measurements in our two datasets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.