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

arXiv:2610.13599 · cs.LG · Submitted 2026-10-09 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction".

Jane: The gist The ECHOTAB framework introduces a measurement-grounded representation learning approach that uses measurement-specific tokens as an intermediate representation between echo views and diagnosis,

Tom: First, who's behind it and why it matters.

Title and authors: Tom: Let's talk about the paper itself. It’s called EchoTab, and the authors are Michelle Espranita Liman, Ioana Ionac, Alexander Muller, Eimo Martens, Daniel Rueckert. They’re coming from TUM and Imperial College London > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg1

Jane: It sounds like they are tackling a problem where diagnostic information is spread across different views in an echocardiography study, and the existing methods don't explicitly model those measurements as targets yet. This paper aims to fix that gap > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg1

Lu: They are proposing measurement-specific tokens that act as an intermediate representation, which is then used for prediction, and they train this through a tabular reconstruction objective > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Meng: So they are trying to create these tokens so the model can predict the actual measurement values based on what it sees in the echo views > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Lalam: The paper is interested in how we can make echoes more measurement-aware without just encoding those measurements as discrete text representations, which is a limitation of some existing contrastive learning methods > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg2

The paper's summary: Tom: So, summarizing what they did: they use an echo encoder to get view tokens, then a tabular decoder lets each measurement query selectively pull information from the echo views relevant to it > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Jane: And the key part is that this results in a set of measurement-factorized tokens for the study, and then these tokens are used by a prediction head to predict those measurement values using a tabular reconstruction objective > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Lu: They retain local patch tokens, specifically Lv (c),j, where each zv (c),j represents a spatio-temporal patch of the video clip, instead of using just a global CLS token > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Meng: The paper shows that these measurement-factorized tokens allow for accurate diagnosis by letting us use only the measurement tokens relevant to that specific diagnosis > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Lalam: This whole setup is motivated by the clinical workflow, where different measurements come from different subsets of views and different diagnoses are informed by specific subsets of measurements > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg1

The paper's improvements: Tom: Now for the improvements. They show that instead of relying on generic study-level representations, ECHOTAB learns these measurement-specific tokens through tabular reconstruction > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg2

Jane: And this is important because it enables diagnosis-specific selection of relevant measurements, which improves performance on two datasets over pre-training methods that learn generic representations > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg2

Lu: They also show that the learned attention over echo views is consistent with established measurement-view relationships that cardiologists actually use in their work > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Meng: The paper mentions they used a partially shared head variant, which they found provided a good balance between having flexibility for each measurement and keeping things aligned across different measurements > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg5

Lalam: And it’s worth noting that the authors found that choosing diagnosis-relevant measurement tokens is essential; if you average all of them, you mix in information irrelevant to the actual diagnosis and your classification performance goes down > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg5

Conclusion: Tom: So, to wrap up EchoTab, it proposes a measurement-grounded representation learning framework that uses measurement tokens as an intermediate step between echo views and diagnosis > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg1

Jane: It learns these tokens via tabular reconstruction without needing negative samples or large batch sizes, which is a big efficiency win for training complex multi-view tasks > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg2

Lu: The alignment with clinical relationships, where the learned attention matches what cardiologists do, is a strong finding that shows the representation is clinically meaningful > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg3

Meng: From a deployment standpoint, I’d look at how much overhead this measurement token generation adds to inference time when we run this on real traffic > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg5

Lalam: Ultimately, the paper’s success lies in achieving the highest average AUC among echo foundation models and contrastive baselines on MIMIC, outperforming the strongest baseline by four point four points > EchoTab: Learning Measurement-Grounding Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg5 <ref:2610.13599#pg1,Representations for Echocardiographic Diagnosis through Tabular Reconstruction>

Tom: So, we've seen how they use tabular reconstruction to avoid negative samples while still getting strong results on diagnosis classification with EchoTab > EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction#pg1

Michelle Espranita Liman, Ioana Ionac, Alexander Muller, Eimo Martens, Daniel Rueckert, Philip Muller

Technical University of Munich and TUM University Hospital and Department of Cardiology at Technical University of Munich and TUM University Hospital Department of Computing at Imperial College London and Munich Center for Machine Learning

cs.LG

Submitted: 2026-10-09

Updated: 2026-10-09

Code: https://github.com/michelleespranita/EchoTab

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 90/100

The gist: The gist The ECHOTAB framework introduces a measurement-grounded representation learning approach that uses measurement-specific tokens as an intermediate representation between echo views and

Key concepts

Measurement-Specific Tokens: These are intermediate representations that link echo views and diagnoses. Each token is trained specifically to predict a corresponding measurement value, allowing the model to capture clinically meaningful structural and functional properties derived from specific measurements.
Tabular Reconstruction: This is the training objective used to learn the measurement-specific tokens. The model learns these tokens by reconstructing tabular data (measurements) from factorized echo representations without needing negative samples or large batch sizes, enabling effective learning.
Measurement-Grounding: This approach ensures that the learned representation is tied directly to specific measurements. It leverages the clinical workflow where different measurements come from different views, allowing the model to use only the most relevant information for a particular diagnosis.

Summary

The gist The ECHOTAB framework introduces a measurement-grounded representation learning approach that uses measurement-specific tokens as an intermediate representation between echo views and diagnosis, which improves diagnosis classification performance by selectively using relevant measurements

How it works

ECHOTAB represents multiview 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 This approach is motivated by the hierarchical clinical workflow where different measurements are derived from different subsets of views, and different diagnoses are informed by specific subsets of measurements

The framework consists of several components that work together to achieve this goal

** Echo Encoder:**

The echo encoder is initialized with EchoPrime weights and uses the MViTv2 architecture, where each selected view instance is encoded independently Local patch tokens from all selected views are concatenated into the study-level echo token sequence Z

** Tabular Decoder:**

The tabular decoder consists of B cross-attention blocks, each followed by a feed-forward network (FFN), with residual connections, layer normalization, and learnable layer-scale parameters It initializes learnable measurement queries qm for measurement m ∈ M to associate each measurement with the echo views most relevant to it

** Prediction Head:**

The prediction head acts as a lightweight read-out on top of the tabular decoder, reconstructing the corresponding tabular measurement values from the measurement-factorized tokens The framework is trained end-to-end with a tabular reconstruction objective, which does not require negative samples and large batch sizes

Key contributions and findings

The main contributions of ECHOTAB include:

** Proposing ECHOTAB:**

We propose ECHOTAB, a measurement-grounded representation learning framework that represents multi-view echocardiography studies using measurement-specific tokens, which serve as an intermediate representation between echo views and diagnosis

** Learning through Tabular Reconstruction:**

We show that ECHOTAB learns these measurement-specific tokens through tabular reconstruction, enabling diagnosis-specific selection of relevant measurements, improving diagnosis performance on two datasets over pre-training methods that learn generic studylevel representations, while avoiding the need for negative samples and large batch sizes

** Alignment with Clinical Relationships:**

We systematically study ECHOTAB’s architecture and behavior, showing that its learned attention over echo views is consistent with established measurement-view relationships used by cardiologists

Experimental results

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

ECHOTAB consistently improves diagnosis classification across both datasets and achieves the highest average AUC among echo foundation models, contrastive baselines, and the MLP baseline On MIMIC, ECHOTAB outperforms the strongest baseline (view-level SupCon) by 4.4 average AUC points (0.748 vs. 0.704), while requiring only 1 GPU compared to 8 GPUs for SupCon

Ablation and analysis

Ablation studies show that the partially shared head provides a compromise between measurement-specific representational flexibility and cross-measurement alignment, yielding the best diagnosis classification in both cohorts Furthermore, choosing diagnosis-relevant measurement tokens is essential for accurate diagnosis, as averaging all measurement tokens mixes in diagnosis-irrelevant information and reduces classification performance ECHOTAB selects the same echo views as the cardiologist in 80% of measurement-view pairs, agreeing on 81% of measurements

Future directions

ECHOTAB requires specifying the mapping between diagnoses and the measurements relevant to them when constructing diagnosis-specific embeddings, which introduces domain knowledge into the model Future work could investigate its application to other modalities, such as cardiac MRI, CT, or retinal imaging and explore more sophisticated aggregation mechanisms than simple averaging for measurement-specific tokens

AI Use Statement

In this work, we used generative AI tools to design and provide feedback on our research methodology and experiments, implement methods, and interpret results We have not used generative AI tools to generate synthetic datasets, help develop theoretical models or conceptual frameworks, propose or refine hypotheses, or clean and reformat datasets Additionally, we used generative AI tools to create and edit software code, draft parts of a research paper

Reproducibility Statement

We provide our code at https://github.com/michelleespranita/EchoTab, which includes the implementation of our model architecture and the downstream task Our experiments use the MIMIC-IV-Echo dataset v1.0 (Gow et al., 2026; Pollard et al., 2026), which is distributed through PhysioNet under credentialed access

Acknowledgments

This study was approved by the Ethics Committee of TUM Klinikum Rechts der Isar (reference number 2025-395-S-CB, application dated July 13, 2025)

References

Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers Rujikorn Charakorn, Edoardo Cetin, Yujin Tang, and Robert Tjarko Lange. Text-to-LoRA: Instant transformer adaption Mateo Espinosa Zarlenga, Pietro Barbiero, Gabriele Ciravegna, Giuseppe Marra, Francesco Giannini, Michelangelo Diligenti, Zohreh Shams, Frederic Precioso, Stefano Melacci Adrian Weller, Pietro Lio and Mateja Jamnik. Concept embedding models: Beyond the accuracy-explainability trade-off Brian Gow, Tom Pollard, Nathaniel Greenbaum, Benjamin Moody, Ahram Han, Jonathan W Waks, Alistair Johnson, Elizabeth Herbst Marta Hasny, Maxime Di Folco, Keno Bressem and Julia Schnabel. Tables guide vision: Learning to see the heart through tabular data Gregory Holste, Evangelos K. Oikonomou, Marton Tokodi, Attila Kovacs Zhangyang Wang and Rohan Khera. Complete ai-enabled echocardiography interpretation with multitask deep learning Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang and Weizhu Chen. LoRA: Low-rank adaptation of large language models Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock and Evan Shelhamer. Perceiver IO: A general architecture for structured inputs & outputs Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu and Dilip Krishnan. Supervised contrastive learning Myung Jun Kim, Felix Lefebvre Ga etan Brison Alexandre Perez-Lebel and Ga el Varoquaux. Table foundation models:

Improvements for AI systems

Bold measurement-grounded representation learning for echocardiography studies that explicitly models measurements as an intermediate representation between echo views and diagnosis. This allows downstream diagnostic tasks to selectively use relevant measurements, improving accuracy by using measurement-specific tokens that form a factorized echo representation.
Develop a system capable of diagnosis-specific measurement selection by using the learned tokens for accurate prediction, as shown by the result: ECHOTAB consistently improves diagnosis classification across both datasets and achieves the highest average AUC among echo foundation models, contrastive baselines, and the MLP baseline.
Improve efficiency by training without requiring negative samples or large batch sizes due to the use of a tabular reconstruction objective. This enables efficient training for complex multi-view tasks compared to contrastive approaches that require negative samples or large batch sizes.
Enhance interpretability by aligning learned attention with clinical knowledge, as demonstrated by the finding: ECHOTAB selects the same echo views as the cardiologist in 16 of 20 (80%) measurement-view pairs, indicating it learns clinically meaningful measurement-view associations.
Implement a flexible prediction mechanism through various head variants (independent, shared, and partially shared) to balance representational flexibility with cross-measurement alignment, achieving the best diagnosis classification on both datasets with the partially shared head variant.

Sources

EchoJEPA: A Latent Predictive Foundation Model for Echocardiography