EchoingECG: Probabilistic ECG–ECHO Distillation

• EchoingECG is a cross-modal model that represents ECG and ECHO embeddings as probability distributions to capture uncertainty.
• It integrates PCME++ with ECHO-CLIP supervision, achieving improved balanced accuracy (e.g., from 57.5% to 63.2% in zero-shot settings) for cardiac predictions.
• The model’s uncertainty estimates guide segment selection and triage, helping clinicians decide when additional echocardiographic evaluation is needed.

EchoingECG is an electrocardiogram cross-modal model for echocardiogram tasks that introduces a probabilistic student–teacher model leveraging uncertainty-aware ECG embeddings and ECHO supervision to improve ECG-based cardiac function prediction (Gao et al., 30 Sep 2025). It is motivated by the asymmetry between echocardiography, which provides gold-standard measurements of cardiac structure and function but requires considerable hospital resources, and electrocardiography, which is low-cost, ubiquitous, and routinely collected. The model combines Probabilistic Cross-Modal Embeddings (PCME++) with ECHO-CLIP, a vision-language pre-trained model trained on ECHO-text pairs, so that echocardiographic knowledge can be distilled into ECG representations while preserving ECG–text alignment.

1. Clinical motivation and problem setting

Echocardiograms provide gold-standard measurements of cardiac structure and function, including examples such as left ventricular ejection fraction and chamber volumes, but they are resource-intensive and require specialized equipment and personnel (Gao et al., 30 Sep 2025). Electrocardiograms, by contrast, are low-cost, accessible, and routinely collected. Recent work had already shown that ECGs can be used to predict some ECHO-derived outcomes, creating the possibility of rapid and widely accessible cardiac screening.

EchoingECG is framed around two limitations of prior ECG-based methods. First, most prior ECG-to-ECHO foundation models, including ECG-CLIP, MEDBind, and ECG-FM, employ deterministic contrastive losses based on InfoNCE that learn a one-to-one mapping between each ECG segment and its paired modality. Second, ECG signals exhibit natural variability and noise, including arrhythmic beats, motion artifacts, and electrode misplacement, and the same underlying cardiac condition may manifest in many subtly different ECG waveforms. The paper characterizes this as a “many-to-many” uncertainty that deterministic embeddings struggle to capture. A further constraint is data availability: large-scale, expert-annotated ECG–ECHO pairs are scarce, making it difficult to train high-capacity models without external supervision.

Within this formulation, EchoingECG addresses both representation mismatch and supervision scarcity by explicitly modeling uncertainty in ECG embeddings via probabilistic contrastive learning and by using a vision-language ECHO teacher to transfer echocardiographic information into the ECG embedding space.

2. Probabilistic student–teacher architecture

The central design choice in EchoingECG is to represent each modality-specific embedding as a distribution rather than as a single deterministic vector (Gao et al., 30 Sep 2025). For each modality $m$—an ECG window, text report, or ECHO video—the encoder outputs a mean $\mu_m \in \mathbb{R}^D$ and a log-variance $\log \sigma_m^2 \in \mathbb{R}^D$, defining

$$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$

Similarity between two probabilistic embeddings is computed with the Closed-Form Sampled Distance,

$$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$

The first term enforces similarity of means for aligned pairs, while the second encourages small total variance for confident associations and allows larger variance for ambiguous pairs. The probabilistic contrastive objective is a sampled-distance variant of InfoNCE,

$$L_{\text{contrast}} = - \mathbb{E}_{(x,y)} \left[ \log \frac{\exp(\operatorname{sim}(\mu_x,\mu_y)/\tau)} {\sum_{y'} \exp(\operatorname{sim}(\mu_x,\mu_{y'})/\tau)} \right] + \lambda_{\text{uncertainty}} \cdot \operatorname{Tr}(\Sigma_x+\Sigma_y),$$

where $\operatorname{sim}(a,b)=a \cdot b / \|a\|\|b\|$, $\tau$ is a temperature hyperparameter, $\Sigma_x=\operatorname{diag}(\sigma_x^2)$, and the trace term penalizes large uncertainty for matched pairs.

The teacher side is supplied by ECHO-CLIP. For an ECHO video with frame embeddings $e_i \in \mathbb{R}^d$, EchoingECG treats $\mu_m \in \mathbb{R}^D$0 as independent samples and computes

$\mu_m \in \mathbb{R}^D$1

This yields

$\mu_m \in \mathbb{R}^D$2

The student ECG encoder is a 1D-ResNet that outputs $\mu_m \in \mathbb{R}^D$3 and $\mu_m \in \mathbb{R}^D$4, while the frozen ECHO-CLIP produces $\mu_m \in \mathbb{R}^D$5. Training uses two probabilistic contrastive terms,

$\mu_m \in \mathbb{R}^D$6

combined as

$\mu_m \in \mathbb{R}^D$7

with $\mu_m \in \mathbb{R}^D$8. In this construction, $\mu_m \in \mathbb{R}^D$9 functions as a distillation loss, transferring echocardiographic knowledge into ECG representations while retaining ECG–text alignment.

3. Data, preprocessing, and evaluation protocol

EchoingECG is trained and evaluated on three datasets with distinct roles (Gao et al., 30 Sep 2025). MIMIC-ECG contains 10-second, 12-lead ECGs downsampled to 100 Hz, paired with text reports, with splits of 88,291 train, 12,065 validation, and 24,644 test. A linked MIMIC-ECHO subset contributes ECHO videos with text reports, comprising 814 train, 67 validation, and 194 test samples, matched to MIMIC-ECG by patient ID and admission time. MUSIC serves as an external dataset and contains 20-minute vectorcardiograms transformed to 12-lead ECGs, with ECHO labels, split into 512 train and 125 test.

Preprocessing follows modality-specific pipelines. ECGs receive normalization and standard augmentations as in Raghu et al. 2022. Text uses full clinical notes tokenized and padded to 244 tokens, then encoded by BioBERT. ECHO videos are processed frame by frame by the frozen ECHO-CLIP teacher.

Three evaluation settings structure the downstream analysis. In zero-shot evaluation, EchoingECG uses direct cosine-distance scoring between ECG $\log \sigma_m^2 \in \mathbb{R}^D$0 and textual prompt embeddings $\log \sigma_m^2 \in \mathbb{R}^D$1, with no finetuning. In few-shot evaluation, using 10% of the labeled data, the ECG encoder is frozen and a linear classifier is trained on $\log \sigma_m^2 \in \mathbb{R}^D$2. In fine-tuning with 100% of the labeled data, the ECG encoder and classifier are jointly fine-tuned on the task. Optimization uses AdamW with weight decay $\log \sigma_m^2 \in \mathbb{R}^D$3, learning rate $\log \sigma_m^2 \in \mathbb{R}^D$4, batch size $\log \sigma_m^2 \in \mathbb{R}^D$5, 150 epochs, and embedding dimension $\log \sigma_m^2 \in \mathbb{R}^D$6.

This protocol is designed to test not only end-task accuracy but also whether the representation learned by cross-modal distillation remains useful under limited supervision. The paper’s emphasis on zero-shot and few-shot regimes is significant because it targets the practical scenario in which large labeled ECG–ECHO corpora are unavailable.

4. Quantitative performance

On MIMIC for LVEF $\log \sigma_m^2 \in \mathbb{R}^D$7, EchoingECG outperforms the listed foundation ECG baselines in zero-shot, few-shot, and fine-tune settings (Gao et al., 30 Sep 2025).

Setting	Strongest listed baseline	EchoingECG
ZS balanced accuracy (%)	MEDBind 57.5	63.2
10% balanced accuracy (%)	MEDBind 66.9	69.3
100% balanced accuracy (%)	ECG-CLIP 77.1	78.9

The comparison against individual baselines is also explicit: ECG-CLIP achieves 56.3, 64.5, and 77.1; MEDBind 57.5, 66.9, and 76.2; ECG-FM_M has 55.0 and 66.8 in 10% and 100% settings; and ECG-FM_P has 61.5 and 70.7 in those same settings. EchoingECG is reported as best across settings.

External validation on MUSIC extends this pattern. For balanced accuracy across the tasks LVEF$\log \sigma_m^2 \in \mathbb{R}^D$8, SLVH, DLV, and combined, EchoingECG with the start window gives zero-shot $\log \sigma_m^2 \in \mathbb{R}^D$9, fine-tune $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$0, and overall rank $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$1. After selecting lowest-variance segments, EchoingECG reaches rank 1 with zero-shot $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$2 and fine-tune $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$3. The listed baseline models—ECG-CLIP, MEDBind, and ECG-FM*—range around zero-shot $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$4–$$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$5 and fine-tune $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$6–$$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$7.

The model also improves text-to-ECG retrieval on the MIMIC-ECG test set. Recall@1 / Recall@10 are 50.2 / 93.9 for ECG-CLIP, 53.6 / 94.5 for MEDBind, and 67.0 / 94.5 for EchoingECG. The gain in Recall@1 occurs without sacrificing Recall@10.

Ablation results isolate the contributions of probabilistic contrastive learning and the ECHO teacher. On LVEF classification in MIMIC-ECHO and Recall@1 in MIMIC-ECG, the reported values are 56.8% / 53.2% for InfoNCE only, 61.2% / 52.9% for InfoNCE + ECHO Teacher, 62.0% / 67.9% for PCME++ only, and 63.2% / 67.0% for PCME++ + ECHO Teacher. The stated interpretation is twofold: PCME++ outperforms deterministic InfoNCE, and ECHO-CLIP supervision further boosts ECHO-related tasks without harming ECG–text retrieval.

5. Uncertainty estimation and segment-level behavior

A distinctive feature of EchoingECG is that the learned diagonal covariance $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$8 is treated as a per-embedding uncertainty score (Gao et al., 30 Sep 2025). This is not an auxiliary diagnostic computed after the fact; it is part of the embedding representation itself, arising directly from the PCME++ formulation.

Several empirical observations support the claim that this uncertainty is meaningful. On MIMIC, uncertainty splits based on median $$Z_m \sim \mathcal{N}(\mu_m, \operatorname{diag}(\sigma_m^2)).$$9 show that lower $$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$0 correlates with higher balanced accuracy. On text-to-ECG retrieval, low-variance ECGs yield Recall@1 $$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$1, whereas high-variance ECGs yield Recall@1 $$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$2. In MUSIC, which contains 20-minute recordings, a 10-second sliding window of ECG Lead I shows that noisy or artifact-laden segments produce higher $$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$3, while clean and stable intervals yield lower $$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$4.

These results matter because they operationalize uncertainty at the level of both whole recordings and local windows. The paper therefore proposes two uses. First, $$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$5 can guide segment selection in long recordings by choosing low-variance windows. Second, it can help clinicians identify low-confidence predictions. A plausible implication is that the model’s uncertainty estimates can serve as a triage signal, separating cases in which ECG-only inference is likely to be reliable from cases in which echocardiography remains necessary.

6. Position within ECG–ECHO research

EchoingECG sits within a broader line of work that links ECG and echocardiography, but it addresses a specific problem: distilling ECHO supervision into ECG embeddings for prediction and retrieval tasks (Gao et al., 30 Sep 2025). Related systems differ in modality target, supervision, or aggregation strategy. ECHOPulse is an ECG-conditioned ECHO video generation model based on VQ-VAE tokenization and masked visual token modeling, emphasizing controllable synthetic ECHO generation rather than ECG-based ECHO prediction (Li et al., 2024). Echo2ECG enriches ECG representations with cardiac morphology from multi-view Echos by aggregating up to approximately 128 views through attention pooling and aligning ECG and study-level echo embeddings in a shared 512-dimensional space; it reports 12.5 M trainable parameters and explicitly contrasts this with 126.6 M in EchoingECG (Liman et al., 9 Mar 2026). Echo-SyncNet, in turn, addresses self-supervised synchronization of multiple echo cines without any external input, including ECG, and learns a phase-sensitive embedding space for temporal alignment (Dezaki et al., 2021). EchoXFlow provides a beamspace echocardiography dataset with multiple ultrasound modalities paired with a synchronized single frontal-lead ECG sampled at 600 Hz, making cross-modal and phase-aware learning problems directly formulable from acquisition-time data (Stenhede et al., 6 May 2026). AnyECG-Echo extends echo-report supervision to wearable-compatible single-lead ECGs for 13 fine-grained structural heart disease subtypes and reports external validation in a geographically independent cohort of $$d(Z_{m1}, Z_{m2}) = \|\mu_{m1} - \mu_{m2}\|_2^2 + \|\sigma_{m1}^2 + \sigma_{m2}^2\|_1.$$6 (He et al., 8 Jun 2026).

Against this background, EchoingECG is notable for combining two technical commitments in a single framework: probabilistic cross-modal embeddings and student–teacher distillation from a vision-language ECHO model. The paper concludes that its gains in zero-shot and few-shot regimes suggest deployability in low-resource settings where large labeled ECG–ECHO pairs are unavailable, and that uncertainty quantification is critical for clinical safety and triage, especially in ambulatory or remote monitoring scenarios. This suggests a broader methodological shift in ECG–ECHO research from deterministic alignment alone toward uncertainty-aware cross-modal representation learning.