IECG: Weakly-Supervised Interpretability-Enhancing Foundation Model for Electrocardiogram

Huynh Dang Nguyen; Trong-Thang Pham; Anh-Thuan Do; Phuong-Nhi Dinh; Brijesh Patel; Ngan Le; Van Nguyen

cs.AIOct 20, 2026

IECG: Weakly-Supervised Interpretability-Enhancing Foundation Model for Electrocardiogram

Authors: Huynh Dang Nguyen, Trong-Thang Pham, Anh-Thuan Do, Phuong-Nhi Dinh, Brijesh Patel, Ngan Le, Van Nguyen

Organizations: Quantum AI & Cyber Security Institute (QACI) - FPT Corporation, Vietnam · University of Arkansas, USA · Hospital Haiphong Medical University, Vietnam · Pham Ngoc Thach University of Medicine, Vietnam ·

Abstract

Electrocardiography (ECG) is a cornerstone of cardiac diagnosis, yet most machine learning methods for ECG analysis lack interpretability, limiting their clinical adoption. We present IECG, a weakly-supervised ECG foundation model that jointly performs diagnostic classification and abnormality localization without requiring manual abnormal region annotations, overcoming a long-standing bottleneck in interpretable ECG learning. Central to IECG is a weakly-supervised attention strategy that explicitly incorporates clinical priors by deriving abnormality heatmap pseudo-labels from established ECG clinical knowledge. These pseudo-labels guide the model's CLS-token attention through a KL-divergence loss, encouraging it to focus on diagnostically relevant waveform regions. During pretraining, IECG is jointly optimized with clinically guided attention supervision, ECG-text contrastive alignment, and report captioning objectives, enabling it to learn rich multimodal representations while producing localized waveform evidence. By explicitly aggregating features from clinically guided attention regions, IECG grounds its diagnostic predictions in interpretable waveform evidence without sacrificing classification performance. Extensive experiments demonstrate that IECG achieves state-of-the-art abnormal region localization while maintaining strong classification performance, advancing trustworthy foundation models for ECG analysis.

Figures & tables

Figure 1: IECG overview. The ECG Encoder learns representations through ECG-Text Alignment and a self-supervised clinical abnormality localization mechanism with CGAL pipeline.

Figure 2: Each ECG token is enriched with Spatial and Temporal embeddings in addition to Positional embeddings, providing explicit lead identity and time frame information.

Figure 3: The Criteria Feature Retrieval (CFR) module retrieves diagnostic criteria by comparing the input diagnosis embedding with a pre-constructed vector database.

Model	Interpretable	Abnormality localization		Classification		Faithfulness
		P@50% ( $\uparrow$ )	AUROC ${}_{\text{loc}}$ ( $\uparrow$ )	FMax ( $\uparrow$ )	AUROC ${}_{\text{cls}}$ ( $\uparrow$ )	Insertion ( $\uparrow$ )	Deletion ( $\downarrow$ )
ST-MEM [ 23 ]	✗	16.56 $\pm$ 1.29	50.07 $\pm$ 0.65	16.18 $\pm$ 0.12	59.07 $\pm$ 0.25	23.25 $\pm$ 0.38	21.57 $\pm$ 0.37
METS [ 13 ]	✗	18.89 $\pm$ 1.26	59.25 $\pm$ 1.24	40.19 $\pm$ 0.33	88.05 $\pm$ 0.30	38.42 $\pm$ 0.58	21.24 $\pm$ 0.48
MERL [ 17 ]	✗	19.73 $\pm$ 1.28	61.13 $\pm$ 1.23	41.82 $\pm$ 0.57	88.80 $\pm$ 0.27	35.50 $\pm$ 0.28	20.07 $\pm$ 0.54
ECG-FM [ 21 ]	✗	16.27 $\pm$ 1.21	50.29 $\pm$ 0.43	34.37 $\pm$ 0.43	83.41 $\pm$ 0.52	25.23 $\pm$ 0.36	12.09 $\pm$ 0.14
HeartLang [ 10 ]	✗	17.08 $\pm$ 1.28	50.32 $\pm$ 1.12	27.58 $\pm$ 0.47	77.07 $\pm$ 0.64	23.50 $\pm$ 0.31	20.76 $\pm$ 0.34

Table 1: Abnormality localization and Classification results on

\text{ECGAD}_{\text{data}}

. Bold indicates the best result, and underline denotes the second-best result. For IECG , attention supervision is applied to either the attention scores from the final Transformer layer ( IECG -Last) or the layer-wise aggregated attention obtained via attention rollout ( IECG -Rollout).

Figure 4: Qualitative comparison of abnormality localization across different ECG conditions. For each example, the top panel shows the ECG waveform with ground-truth abnormalities highlighted in blue. The middle and bottom rows present the localization heatmaps produced by ECGAD ${}_{\text{model}}$ and IECG , respectively. Compared to ECGAD ${}_{\text{model}}$ , IECG produces more accurate and temporally consistent activations that better align with clinically relevant abnormalities across diverse conditions.

Model	FMax	AUROC ${}_{\text{cls}}$
ST-MEM [ 23 ]	11.05 $\pm$ 0.03	64.39 $\pm$ 0.04
METS [ 13 ]	39.88 $\pm$ 0.23	90.82 $\pm$ 0.17
MERL [ 17 ]	40.00 $\pm$ 0.27	91.66 $\pm$ 0.17
ECG-FM [ 21 ]	31.00 $\pm$ 0.33	85.10 $\pm$ 0.32
HeartLang [ 9 ]	26.03 $\pm$ 0.47	80.10 $\pm$ 0.34
TolerantECG [ 24 ]	40.76 $\pm$ 0.38	92.02 $\pm$ 0.12

Table 2: Classification result evaluated on PTB-XL dataset. Bold indicates the best result, and underline indicates the second-best result.

Configuration	AUROC ${}_{\text{loc}}$	AUROC ${}_{\text{cls}}^{\text{ECGAD}}$	AUROC ${}_{\text{cls}}^{\text{PTB-XL}}$
$\beta=1$ ( Default )	75.59 $\pm$ 1.00	89.09 $\pm$ 0.49	92.04 $\pm$ 0.29
$\beta=2$	75.28 $\pm$ 0.95	89.30 $\pm$ 0.26	91.36 $\pm$ 0.13

Table 3: Effect of captioning loss on localization and classification.

Configuration	AUROC ${}_{\text{loc}}$	AUROC ${}_{\text{cls}}^{\text{ECGAD}}$	AUROC ${}_{\text{cls}}^{\text{PTB-XL}}$
$\gamma=0$	61.33 $\pm$ 1.27	89.41 $\pm$ 0.23	92.14 $\pm$ 0.09
$\gamma=0.5$ ( Default )	75.55 $\pm$ 1.00	89.09 $\pm$ 0.49	92.03 $\pm$ 0.23
$\gamma=1.0$	75.31 $\pm$ 0.97	89.38 $\pm$ 0.40	92.61 $\pm$ 0.12

Table 4: Effect of attention supervision on localization and classification.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Method	P onset	P offset	QRS onset	QRS offset	T onset	T offset
ProMac	84.72	84.91	77.05	78.08	81.70	67.62
NeuroKit	84.78	84.60	79.62	80.92	79.52	63.97
UNet3+	96.68	96.75	99.88	99.89	97.82	96.66

Appendix

Table S1 : Performance comparison of segmentation methods on ECG component boundary localization, evaluated using F1‑score on the Semi‑Seg‑ECG test set.

Dataset	P onset	P offset	QRS onset	QRS offset	T onset	T offset
ISP	97.09	97.16	100.00	100.00	99.50	99.48
Zhejiang	94.56	94.60	99.84	99.85	96.92	95.73
LUDB	96.49	96.57	99.86	99.86	99.60	99.27
QTDB	97.74	97.74	99.43	99.43	95.52	89.14

Appendix

Table S2 : F1 score performance of the UNet3+ across multiple datasets in the SemiSegECG test set.

Component	Abnormal Conditions
P wave	Missing (absent P wave)
	Amplitude $>0.25$ mV in limb leads
	Amplitude $>0.15$ mV in precordial leads
	Duration $>120$ ms
	Inverted in leads I, II, III; Upright in lead aVR
QRS complex	Duration $>120$ ms

Appendix

Table S3 : Abnormal criteria for each ECG waveform component. These thresholds are used in CGAL Stage 2 to flag candidate abnormal regions.

Query	Retrieved diagnosis	Retrieve score
Nonspecific ST changes	Septal hypertrophy	-14.26
Marked ST depression in V lead	Pericarditis	-10.04
Short PR interval	Extreme Axis Deviation	-9.29
ST depression V1-V3	Poor R Wave Progression	-8.83
Prolonged PR interval	R Wave Peak Time	-8.04
ST-T changes	Digitalis Effect	-7.53

Appendix

Table S4 : Examples of lowest‑scoring queries in CFR retrieval.

Configuration	AUROC ${}_{\text{loc}}$	AUROC ${}_{\text{cls}}^{\text{ECGAD}}$	AUROC ${}_{\text{cls}}^{\text{PTB-XL}}$
w/o ST embedding	75.01 $\pm$ 0.84	90.31 $\pm$ 0.24	92.08 $\pm$ 0.30
w/ ST embedding ( Default )	75.55 $\pm$ 1.00	89.09 $\pm$ 0.49	92.03 $\pm$ 0.23

Appendix

Table S5 : Effect of Spatial-Temporal embedding on localization and classification.

Figure S1 : First-degree atrioventricular block with sinus rhythm. IECG consistently highlights prolonged PR intervals across multiple leads, showing stronger alignment with ground truth compared to the baseline, which exhibits more dispersed and less structured localization.

Figure S2 : Anterolateral myocardial infarction with first-degree AV block and intraventricular conduction disturbance. As expected clinically, IECG assigns substantially higher attention weight to the anterolateral leads (I, aVL, and V4–V6) to capture the myocardial infarction, while simultaneously maintaining strong focus on the prolonged PR intervals across the waveform to identify the AV block. This demonstrates the model’s ability to accurately localize both spatially-restricted morphological abnormalities and global temporal conduction delays within the same sample, whereas the baseline exhibits fragmented and less anatomically specific activations.

Figure S3 : Inferior myocardial infarction with first-degree AV block. To accurately diagnose the inferior infarction, IECG distinctly concentrates its attention on the inferior leads (II), precisely matching the expected anatomical region of the ischemic injury. Concurrently, it correctly highlights the prolonged PR intervals to account for the atrioventricular conduction delay. In contrast, the baseline model fails to prioritize the inferior leads and produces activations that lack clear anatomical or temporal justification.

Figure S4 : Complete right bundle branch block (RBBB) with left anterior fascicular block (LAFB). IECG accurately targets the broadened QRS complexes characteristic of RBBB, placing prominent attention on the right precordial leads (V1–V3). Furthermore, to identify LAFB, the model correctly distributes its focus toward the inferior leads (II, III, aVF) and high lateral leads (I, aVL), which are critical for recognizing the associated left axis deviation. This precise, anatomically-grounded localization contrasts with the baseline, which generates uncoordinated activations that fail to clearly reflect the underlying multi-fascicular conduction pathologies.

Figure S5 : Example illustrating annotation incompleteness in $\text{ECGAD}_{\text{data}}$ using Lead II. The ECG exhibits complete left bundle branch block (LBBB) and first-degree atrioventricular (AV) block, as indicated by prolonged QRS duration (e.g., 172–176 ms) and prolonged PR interval (e.g., 204–224 ms), respectively. The ground-truth (GT) heatmap highlights regions corresponding to widened QRS complexes associated with LBBB but fails to capture the prolonged PR intervals that define first-degree AV block. In contrast, IECG accurately highlights both QRS and PR regions, reflecting clinically relevant abnormalities across both morphological (ventricular depolarization) and temporal (atrioventricular conduction delay) features. This example demonstrates that missing annotations of temporal abnormalities can lead to underestimation of localization performance.

Method	AUROC ${}_{\text{loc}}$
IECG w/o Attention Supervision ( $\gamma=0$ )	61.33%
Just CGAL (Rule-based Pseudo-labels)	70.13%
IECG ( $\gamma=1.0$ )	75.31%
IECG ( Default , $\gamma=0.5$ )	75.55%

Appendix

Table S6 : Comparison of localization performance between the heuristic-based CGAL pseudo-labels and the trained models.

Configuration	BLEU-4	ROUGE-L	METEOR	BERT-score
w/o CGAL ( $\gamma=0$ )	12.57	34.49	33.61	89.25
IECG ( Default )	12.14	34.28	32.56	89.20

Appendix

Table S7 : Report generation performance on the PTB-XL dataset.

Model	CPSC		Chapman
Model	FMax	AUROC	FMax	AUROC
ST-MEM [ 23 ]	26.95 $\pm$ 0.09	63.03 $\pm$ 0.11	12.47 $\pm$ 0.32	62.22 $\pm$ 0.80
HeartLang [ 9 ]	36.44 $\pm$ 0.61	70.13 $\pm$ 0.45	29.52 $\pm$ 0.65	84.15 $\pm$ 0.64
XAI4ECG [ 32 ]	57.46 $\pm$ 0.65	85.74 $\pm$ 0.40	29.21 $\pm$ 0.78	82.10 $\pm$ 0.58
ECG-FM [ 21 ]	64.94 $\pm$ 0.38	90.24 $\pm$ 0.12	44.03 $\pm$ 0.71	90.76 $\pm$ 0.12
ECGAD [ 6 ]	68.89 $\pm$ 0.90	91.59 $\pm$ 0.82	34.41 $\pm$ 0.73	87.88 $\pm$ 0.38

Appendix

Table S8 : Classification results evaluated on the CPSC and Chapman datasets. Bold indicates the best result, and underline indicates the second-best result.

Electrocardiogram Foundation Models

Electrocardiogram foundation models learn reusable representations from large ECG datasets for diagnosis, risk prediction, and transfer across clinical settings.

See all 29 papers

MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning
Oct 6, 2026Robert A. Lewis, I-Min Chiu, Kyle Verrier +7Electrocardiogram Foundation Models Electrocardiogram Classification
Massachusetts Institute of Technology (work done while at Apple). · Apple, Inc. · Division of Cardiology, Johns Hopkins Medicine. +1
Diagnosing as Cardiologists Do: ECG Agents with Doctor-Grounded Priors for Clinical Reasoning Across Diseases and Populations
Aug 10, 2026Hongxiang Gao, He-yang Xu, Yuwen Li +6Electrocardiography Electrocardiogram Classification
State Key Laboratory of Digital Medical Engineering, Southeast University · School of Instrument Science and Engineering, Southeast University
ECG-InterpBench: Benchmarking the Interpretability of ECG Foundation Models with Matched-Scale Sparse Autoencoders
Jul 29, 2026Yixuan Duan, Wei QiuElectrocardiogram Foundation Models Neural Network Interpretability
Department of Electrical and Computer Engineering Rice University Houston, Texas, USA

Oct 6, 2026J/K move · Enter open · S save

MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning

Robert A. Lewis, I-Min Chiu, Kyle Verrier, Karthik Jayaraman Raghuram+6

Massachusetts Institute of Technology (work done while at Apple). · Apple, Inc. · Division of Cardiology, Johns Hopkins Medicine.+1

Electrocardiogram Foundation Models Electrocardiogram Classification Multimodal Contrastive Learning Electrocardiogram Representation Learning Electrocardiogram Interpretation

Open paper AskPDF arXiv

Electrocardiography (ECG) records the electrical activity of the heart, aiding diagnosis by detecting abnormalities in cardiac function. ECG foundation models have demonstrated promising results, but are limited by a reliance on ECG interpretation reports as their sole supervision. Because interpretation reports only capture the subset of waveform information routinely recognized by clinicians, this constrains representation learning to overlook the broader diagnostic signals present in ECG. We introduce a new ECG foundation model --- MS-ECG-FM --- that is trained through contrastive alignment to multiple distinct clinical note types, including ECG, echocardiography, radiology, and discharge reports. We evaluate MS-ECG-FM on an extended set of ECG detection benchmarks, showing that it comprehensively outperforms existing methods on the full span of conditions that ECG can detect, including in reduced-lead configurations. Different reports improve representations for different diagnostic domains, while multi-source alignment captures their complementary information and produces consistently strong representations across clinically diverse tasks.

Abstract

Electrocardiogram Foundation Models

Explore similar work

MS-ECG-FM: Towards a More Universal Electrocardiogram Foundation Model for Health Monitoring using Multi-source Contrastive Learning

Diagnosing as Cardiologists Do: ECG Agents with Doctor-Grounded Priors for Clinical Reasoning Across Diseases and Populations

ECG-InterpBench: Benchmarking the Interpretability of ECG Foundation Models with Matched-Scale Sparse Autoencoders