EchoTab: Learning Measurement-Grounded Representations for Echocardiographic Diagnosis through Tabular Reconstruction
Authors: Michelle Espranita Liman, Ioana Ionac, Alexander Müller, Eimo Martens, Daniel Rueckert, Philip Müller
Organizations: Chair for AI in Healthcare and Medicine, Technical University of Munich (TUM) and TUM University Hospital, Munich, Germany · Department of Cardiology, TUM University Hospital, Munich, Germany · Department of Computing, Imperial College London, London, UK · Munich Center for Machine Learning (MCML), Munich, Germany
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://github.com/michelleespranita/EchoTab.
Figures & tables
Figure 1: Overview of EchoTab , our measurement-grounded representation learning framework for multi-view echo studies. (1) Given an echo study, each individual view is first encoded by the echo encoder. Next, a tabular decoder lets each measurement query selectively pull information from the echo views that are relevant for estimating that measurement. This results in a set of measurement-factorized tokens for the echo study, which the prediction head uses to predict measurement values. EchoTab is trained with a tabular reconstruction objective, in which the MSE loss between available ground-truth and predicted measurements is calculated. (2) EchoTab ’s measurement-factorized tokens enables accurate diagnosis by using only the measurement tokens that are relevant for the downstream diagnosis.
Figure 2: Variants of the prediction head used in EchoTab . The prediction head is used during the tabular reconstruction of continuous measurements from their corresponding measurement-specific representations. We compare different degrees of parameter sharing to assess the trade-off between shared and measurement-specific components of the prediction function. (1) Independent heads : each measurement token has its own MLP head, enabling fully measurement-specific parameters. (2) Shared head : all measurement tokens share the same MLP head, enforcing fully shared parameters. (3) Partially shared head : all measurement tokens share the base weights of the MLP head, while a hypernetwork generates measurement-specific parameter updates, providing a compromise between shared and measurement-specific parameters. Biases are omitted for clarity.
Table 1: EchoTab ’s measurement-specific representations enable accurate diagnosis for multi-view echocardiography. By enabling diagnosis-specific selection of relevant measurement tokens, EchoTab consistently outperforms echo foundation models, as well as study- and view-level contrastive and MLP baselines, across downstream diagnosis classification tasks. Unlike contrastive approaches, EchoTab does not require large batch sizes, enabling efficient training. Bold and underlined values denote the best and second-best results.
Figure 3: Comparison of the echo views a cardiologist (left) and EchoTab (right) rely on when deriving measurements on MIMIC. EchoTab matches the cardiologist’s view in 16 of 20 (80%) measurement-view pairs, indicating it learns clinically meaningful measurement-view associations.
Method
Echo encoder
Tabular decoder
Tabular decoder queries
Prediction head
MIMIC
Internal
Diag. clf.
Meas. pred.
Diag. clf.
Meas. pred.
MLP
✓
✗
–
Independent
.697
.392
.761
.446
Shared
.683
-16.944
.648
-14.799
✓
✓
Single
Independent
.720
.380
.766
.422
Shared
.628
-16.987
.624
-14.902
EchoTab
✓
✓
Measurement
Independent
.709
.401
.764
.415
Table 2: Ablation of tabular decoder, queries, and prediction head in EchoTab . Bold and underlined values denote the best and second-best results. Diagnosis classification is measured by AUC averaged across diagnoses and measurement prediction by R2 averaged across measurements. In EchoTab , a partially shared head makes the resulting measurement tokens most effective, balancing cross-measurement representation alignment with measurement-specific representational flexibility. As a result, it achieves the best diagnosis performance on both datasets.
Measurement tokens
MIMIC
Internal
Diagnosis-relevant
.748
.778
All
.693
.751
Incorrect
.687
.757
Table 3: Ablation of the measurement tokens used for diagnosis classification (avg. AUC). Choosing diagnosis-relevant measurement tokens is essential for accurate diagnosis.
Remove diagnostic labels (to be used for downstream tasks)
–
0 99
Remove all-missing measurements (second pass)
–
0 71
Appendix
Table 4: Number of measurements remaining after each filtering step. Dashes indicate steps that do not apply to the given dataset.
Internal (patients / studies)
MIMIC (patients / studies)
Train
6,272 / 7,690
3,612 / 5,610
Validation
565 / 0 879
450 / 0 723
Test
566 / 0 861
452 / 0 730
Total
7,403 / 9,430
4,514 / 7,063
Appendix
Table 5: Dataset splits, reported as number of patients and number of echo studies.
Table 6: EchoTab still outperforms baselines when both are given diagnosis-specific prior. Baselines and reference foundation models use only diagnosis-relevant views, whereas EchoTab uses all views with diagnosis-relevant measurement tokens. Bold and underlined values denote the best and second-best results.
Views
Measurement tokens
MIMIC
Internal
All
Diagnosis-relevant
.748
.778
Diagnosis-relevant
Diagnosis-relevant
.725
.753
Appendix
Table 7: Ablation of the echo views used by EchoTab for diagnosis classification (average AUC). Restricting EchoTab to diagnosis-relevant views lowers performance on both datasets, indicating that the tabular decoder benefits from access to all views rather than a defined set of relevant views.
Task
N
T
Mi
Mi-Mo
Mo
Mo-S
S
Train
Aortic stenosis (AS)
115
-
112
-
114
-
112
Aortic regurgitation (AR)
-
9
8
9
9
9
9
Mitral regurgitation (MR)
73
75
74
75
74
73
74
Left atrium size (LA)
417
-
413
-
417
-
-
Right atrium size (RA)
152
-
153
-
152
-
149
Appendix
Table 8: Class distribution of the downstream diagnosis classification tasks of the MIMIC dataset ( N ormal, T race, Mi ld, Mi-Mo mild-to-moderate, Mo derate, Mo-S moderate-to-severe, S evere).
Train
Val
Test
Task
N
Mi
Mo
S
N
Mi
Mo
S
N
Mi
Mo
S
Aortic stenosis (AS)
96
96
96
96
724
28
21
10
736
21
22
20
Mitral regurgitation (MR)
37
37
37
37
557
176
40
10
552
202
35
10
Left atrium size (LA)
113
113
113
113
626
47
79
26
658
46
58
24
Right atrium size (RA)
73
73
73
73
704
29
27
19
738
15
21
17
Left ventricle size (LV)
22
22
22
22
714
12
19
9
711
14
14
6
Appendix
Table 9: Class distribution of the downstream diagnosis classification tasks of our internal dataset ( N ormal, Mi ld, Mo derate, S evere).
Table 11: Diagnosis → Measurement → View mapping for our internal dataset, as defined by a cardiologist.
Figure 4: Distribution of echo studies by view and data split (train/val/test) for the MIMIC and internal cohort datasets.
Token type
Remove background tokens
AUROC
Accuracy
Global (CLS)
-
0.752
0.456
Local
×
0.792
0.554
Local
✓
0.805
0.583
Appendix
Table 12: k -NN ( k=5 ) performance on echo view classification using different configurations of echo tokens from EchoPrime. We use the configuration in the last row for all models that use EchoPrime: local tokens, excluding those that correspond to the background.
Figure 5: List of measurements used from the MIMIC dataset (45 total).
Figure 6: List of measurements used from our internal dataset (67 total).
Hypernetwork variant
# passes
h1w_head output dim.
L
1
2rd=12288
M
2
rd=6144
S
2r=16
d=768
Appendix
Figure 8: Hypernetwork variants used in the partially shared prediction head of EchoTab . Assume r=8 and d=768 . Every other module of the hypernetwork is identical across variants ( Figure 7 ).
Hypernetwork variant
Diag. clf. (avg. AUC)
Meas. pred. (avg. R2 )
L
.735
.405
M
.748
.418
S
.743
.395
Appendix
Table 13: Ablation of the hypernetwork variants used in the partially shared prediction head of EchoTab on the MIMIC dataset. By default, our model uses the M variant.
Figure 9: Per-measurement R2 by EchoTab on the MIMIC dataset.
Figure 10: Per-section R2 by EchoTab on the MIMIC dataset.
Echocardiography is the most widely used non-invasive cardiac imaging modality, providing essential information for cardiovascular diagnosis. Interpreting an echocardiogram requires synthesizing complementary evidence across multiple heart views to identify abnormalities and produce structured clinical reports. While recent efforts focus on improving classification performance, most models lack explicit diagnostic reasoning and spatially grounded anatomical evidence, limiting clinician trust. We present EchoSonar-R, a multi-view reasoning-enabled vision-language model that jointly performs multi-label disease classification and report generation from echocardiography studies. EchoSonar-R combines a spatiotemporal video encoder with a structure-aware cardiac detector that provides spatially grounded anatomical cues to improve interpretability and clinician trust during cross-view reasoning. EchoSonar-R is trained in two stages: supervised fine-tuning (SFT) on reasoning-annotated targets, followed by Group Relative Policy Optimization (GRPO) with task-specific rewards that jointly align classification and report generation within a unified reinforcement-learning framework. Across a private multi-view dataset and two public benchmarks, EchoSonar-R improves macro balanced accuracy by 17.1% on the private set and 6.1% on MIMICEchoQA over the strongest baseline, achieves a GREEN clinical faithfulness score of 0.800, and produces interpretable reasoning traces grounded in multi-view visual evidence.
Darya Taratynova, Ahmed Aly, Numan Saeed +1
Division of Computing and Mathematical Sciences, Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI), Abu Dhabi, UAE
Echocardiography is a widely used modality for cardiac assessment due to its non-invasive and cost-effective nature, but the sparse and heterogeneous spatiotemporal views of the heart pose distinct challenges. Existing masked autoencoder (MAE) approaches typically process images or short clips independently, failing to capture the inherent multi-view structure required for coherent cardiac representation. We introduce Latent Attention Masked Autoencoder (LAMAE), a foundation model architecture tailored to the multi-view nature of medical imaging. LAMAE augments the standard MAE with a latent attention module that enables information exchange across frames and views directly in latent space. This allows the model to aggregate variable-length sequences and distinct views, reconstructing a holistic representation of cardiac function from partial observations. We pretrain LAMAE on MIMIC-IV-ECHO, a large-scale, uncurated dataset reflecting real-world clinical variability. To the best of our knowledge, we present the first results for predicting ICD-10 codes from MIMIC-IV-ECHO videos. Furthermore, we empirically demonstrate that representations learned from adult data transfer effectively to pediatric cohorts despite substantial anatomical differences. These results provide evidence that incorporating structural priors, such as multi-view attention, yields significantly more robust and transferable representations.
Simon Böhi, Irene Cannistraci, Sergio Muñoz Gonzalez +8
Department of Biomedical Engineering, University of Basel, Switzerland · Department of Computer Science, ETH Zurich, Switzerland
Electrocardiography (ECG) is a low-cost, widely used modality for diagnosing electrical abnormalities like atrial fibrillation by capturing the heart's electrical activity. However, it cannot directly measure cardiac morphological phenotypes, such as left ventricular ejection fraction (LVEF), which typically require echocardiography (Echo). Predicting these phenotypes from ECG would enable early, accessible health screening. Existing self-supervised methods suffer from a representational mismatch by aligning ECGs to single-view Echos, which only capture local, spatially restricted anatomical snapshots. To address this, we propose Echo2ECG, a multimodal self-supervised learning framework that enriches ECG representations with the heart's morphological structure captured in multi-view Echos. We evaluate Echo2ECG as an ECG feature extractor on two clinically relevant tasks that fundamentally require morphological information: (1) classification of structural cardiac phenotypes across three datasets, and (2) retrieval of Echo studies with similar morphological characteristics using ECG queries. Our extracted ECG representations consistently outperform those of state-of-the-art unimodal and multimodal baselines across both tasks, despite being 18x smaller than the largest baseline. These results demonstrate that Echo2ECG is a robust, powerful ECG feature extractor. Our code is accessible at https://github.com/michelleespranita/Echo2ECG.
Michelle Espranita Liman, Özgün Turgut, Alexander Müller +3
Chair for AI in Healthcare and Medicine, Technical University of Munich (TUM) and TUM University Hospital, Germany · Department of Cardiology, TUM University Hospital, Germany · Department of Computing, Imperial College London, UK +1