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; Francoise Marvel; Salar Abbaspourazad; Anshuman Mishra; Guillermo Sapiro

cs.LGOct 6, 2026

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

Authors: Robert A. Lewis, I-Min Chiu, Kyle Verrier, Karthik Jayaraman Raghuram, Francoise Marvel, Salar Abbaspourazad, Anshuman Mishra, Guillermo Sapiro, +2 more

Organizations: Massachusetts Institute of Technology (work done while at Apple). · Apple, Inc. · Division of Cardiology, Johns Hopkins Medicine. · Princeton University.

Abstract

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.

Figures & tables

Figure 1: Conceptual overview of MS-ECG-FM. MS-ECG-FM is an ECG foundation model that can support a broader range of ECG detection tasks. It is pre-trained using concurrent alignment to ECG, ECHO, chest X-ray and discharge reports to improve the universality of its representations.

Figure 2: Comparison of MS-ECG-FM to other contemporary ECG foundation models on 12-lead ECG diagnostic tasks using linear probe evaluations. (a): The performance profile shows that MS-ECG-FM universally outperforms the other methods on the 98 downstream labels. (b): MS-ECG-FM is strongest for all diagnostic groups, with large gains on several important diagnosis (dx) groups. Random refers to a randomly initialized ECG encoder. Significance testing is performed using a paired patient-cluster bootstrap procedure (see Methods). 95% bootstrap confidence intervals are shown. AUROC: area under the receiver operating characteristic curve, ARR: arrhythmia, CD: conduction disturbances, ECG-HYP: hypertrophies detected from ECG, MI: myocardial infarction, ISC/STTC: ischemia and ST/T changes, SHD: structural heart disease, VC: diagnoses that rely on voltage criteria. Conditions by diagnosis group are in Table 1 .

		Single Lead			Lead Subsets				12-lead
		I (LCX)	II (RCA)	V2 (LAD)	I,II	All limb	I,II,V2	I,II,III,V2	12-lead
All	Random	76.4	77.7	73.7	81.1	81.7	81.0	82.3	83.6
	MERL	63.5	66.6	68.5	69.7	63.5	75.8	79.0	88.8
	D-BETA	84.4	85.6	82.2	88.2	88.3	89.3	89.2	90.3
	MELP	84.9	86.0	81.9	89.1	89.2	90.6	90.7	91.2
	ECGFounder	86.2	86.4	83.1	89.0	88.9	90.0	90.8	91.5

Table 2: Performance of MS-ECG-FM and contemporary ECG-FMs on detection tasks with common reduced-lead configurations. With the exception of arrhythmias (ARR) and ECG-HYP detected from single lead V2, MS-ECG-FM is consistently the strongest method. AUROC: area under the receiver operating characteristic curve. 2 2 footnotetext: Linear probe evaluations. Random refers to a randomly initialized ECG encoder. The best model in each column is bold ;

\ast

marks models that are not significantly different from the best per a paired patient-cluster bootstrap significance test (see Methods). 2 2 footnotetext: LCX: left circumflex, RCA: right coronary artery, LAD: left anterior descending, ARR: arrhythmia, CD: conduction disturbances, ECG-HYP: hypertrophies detected from ECG, MI: myocardial infarction, ISC/STTC: ischemia and ST/T changes, SHD: structural heart disease, VC: diagnoses that rely on voltage criteria.

Figure 3: Contribution of different clinical note types to downstream performance. The single source oracle represents the performance when selecting the strongest single source method on each label. (a): MS-ECG-FM matches the performance of the oracle by learning the union of information across clinical report types. (b): Alignment to different clinical note types results in stronger representations for different diagnostic tasks. Random refers to a randomly initialized ECG encoder. Significance testing is performed using a paired patient-cluster bootstrap procedure (see Methods). 95% bootstrap confidence intervals are shown. AUROC: area under the receiver operating characteristic curve, ARR: arrhythmia, CD: conduction disturbances, ECG-HYP: hypertrophies detected from ECG, MI: myocardial infarction, ISC/STTC: ischemia and ST/T changes, SHD: structural heart disease, VC: diagnoses that rely on voltage criteria, MIMIC-EST: ICD-based task on conditions with established ECG evidence, MIMIC-EXP: ICD-based labels on conditions with exploratory ECG evidence. Conditions by diagnosis group are in Table 1 .

			AUROC
Group	Condition	Prevalence % (N)	Random	ECG-MR	ECG-CR	ECHO	Chest X-ray	Discharge	MS-ECG-FM
ARR	AFIB	7.2% (152)	93.2	∗ 98.4	∗ 98.5	∗ 98.0	97.9	97.9	98.7
	STACH	3.9% (82)	97.7	∗ 99.4	99.5	99.2	98.8	98.9	∗ 99.5
	SARRH	3.7% (77)	58.2	89.8	∗ 94.1	74.0	75.0	78.7	94.4
	SBRAD	3.1% (64)	93.5	∗ 96.5	∗ 96.6	95.8	∗ 95.4	∗ 96.4	96.7
	SVARR	0.7% (14)	84.8	∗ 95.0	96.4	∗ 94.5	92.6	∗ 94.1	∗ 95.4

Table 3: Contribution of different clinical report types to downstream performance for different condition detection tasks. A glossary of condition codes is in Table 2 . AUROC: area under the receiver operating characteristic curve. 3 3 footnotetext: Linear probe evaluations. Random refers to a randomly initialized ECG encoder. Prevalence is the proportion of positives in the test set the metrics were computed on. The best model is bold ;

\ast

marks models that are not significantly different from the best per a paired patient-cluster bootstrap significance test (see Methods). By default, only conditions with at least 10 positive and negative cases are shown. † Cardiac amyloidosis is excluded from the MIMIC-EXP aggregate score due to low prevalence but shown here given the emerging interest in its detection.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

			MIMIC-IV data usage			Available
Model	Pre-training objective(s)	# Pretrain samples	ECG waveforms	Machine reports	Clinical notes	Code	Weights
ST-MEM Na et al. (2024)	Tokenizes 12-lead ECG recordings at spatio-temporal level (i.e., patches within individual leads). Then uses a modified masked-autoencoder (MAE) reconstruction objective.	345,779	✗	✗	✗	✓	✓
MERL $\dagger$ Liu et al. (2024)	Multimodal contrastive ECG waveform & machine report objective ( InfoNCE ). Auxiliary unimodal contrastive ECG w/latent augmentation ( InfoNCE ).	800,035	✓	✓	✗	✓	✓
KED Tian et al. (2024)	Multimodal contrastive ECG waveform & LLM-augmented machine report objective, with concurrent use of labels from machine report.	800,035	✓	✓	✗	✓	✓
ECG-JEPA Kim (2024)	Uses Joint-Embedding Predictive Architecture ( JEPA ) self-distillation objective where masked latent representations of the ECG from the teacher network are predicted by the student network.	174,140	✗	✗	✗	✓	✓
ECG Founder $\dagger$ Li et al. (2025)	Supervised multi-label classification with labels sourced from an ECG-machine analysis program.	10,771,552	✗	✗	✗	✓	✓

Appendix

Table 1: Review of prominent electrocardiogram foundation models from 2024-2025.

\dagger

indicates models chosen for our experiments due to their strong benchmark performance.

\ddagger

indicates MIMIC-IV-ECG was used in pre-training pooled with other datasets.

Methods	PTBXL-Super			PTBXL-Sub			PTBXL-Form			PTBXL-Rhythm			CPSC2018			CSN
Methods	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%
ST-MEM Na et al. (2024)	61.12	66.87	71.36	54.12	57.86	63.59	55.71	59.99	66.07	51.12	65.44	74.85	56.69	63.32	70.39	59.77	66.87	71.36
HeartLang Jin et al. (2025)	78.94	85.59	87.52	64.68	79.34	88.91	58.70	63.99	80.23	62.08	76.22	90.34	60.44	66.26	77.87	57.94	68.93	82.49
MERL Liu et al. (2024)	82.39	86.27	88.67	64.90	80.56	84.72	58.26	72.43	79.65	53.33	82.88	88.34	70.33	85.32	90.57	66.60	82.74	87.95
D-BETA Hung et al. (2025)	83.15	88.36	90.11	77.74	82.92	85.15	70.10	78.91	83.98	86.61	92.83	96.71	85.46	91.35	94.92	80.04	87.36	90.71
MELP Wang et al. (2025)	85.82	87.61	87.87	79.22	84.40	87.46	63.41	76.71	83.30	88.83	94.65	96.91	88.54	91.75	94.32	78.25	84.83	90.17

Appendix

Table 2: Linear probe performance of contemporary ECG foundation models as reported in their original papers. Results are shown for different percentages of available training labels (1%, 10%, 100%). 9 9 footnotetext: Evaluation metric is AUROC: area under the receiver operating characteristic curve.

ECG-FM	Number of ECG encoder parameters
Baselines
MERL Liu et al. (2024)	$5.47$ M • tokenizer: $0.12$ M • transformer: $5.35$ M
D-BETA Hung et al. (2025)	$63.20$ M • tokenizer: $0.60$ M • transformer: $61.42$ M • class embedding, projection, and pooler: $1.18$ M
MELP Wang et al. (2025)	$62.55$ M • tokenizer: $0.60$ M • transformer: $61.42$ M • attentional pooler: $0.53$ M
ECGFounder Li et al. (2025)	$30.67$ M
MS-ECG-FM ECGViT

Appendix

Table 3: Number of ECG encoder parameters used to produce frozen linear-probe embeddings. Text encoders and unused pre-training or classification heads are excluded. For transformer-based models, tokenizer and transformer counts are listed below the total; MELP and D-BETA additionally use a pooling or projection module at probe time. MS-ECG-FM’s default backbone is underlined .

Category	Diagnosis in clinical practice	ECG detection evidence	Evidence strength
Arrhythmias
Atrial fibrillation (AFib)	Diagnosed from standard ECG (12-lead, 10s) or ambulatory ECG (e.g., with Holter monitors)	Hannun et al. (2019) ; Ribeiro et al. (2020)	★
Silent or future atrial fibrillation	AFib is paroxysmal / episodic. No prototypical signatures in the ECG when AFib is not active. However, machine learning can be used to screen for AFib (future or latent) from normal sinus rhythm ECGs.	Attia et al. (2019c) ; Noseworthy et al. (2022)	★
Other supraventricular and ventricular arrhythmias	Diagnosed from standard ECG (12-lead, 10s) or ambulatory ECG (e.g., with Holter monitors)	Hannun et al. (2019) ; Ribeiro et al. (2020)	★
Conduction Disturbances	ECG-based diagnosis of atrioventricular and intraventricular conduction delays / blocks	Hannun et al. (2019) ; Ribeiro et al. (2020)	★
Ischemic Heart Disease

Appendix

Table 1: Summary of diagnostic categories and clinical outcomes and corresponding evidence for their detection or prediction using ECG. Evidence strength is indicated using symbols: ★ = strong evidence, ▲ = promising evidence, ❍ = exploratory evidence.

#	Diagnosis group	Dataset source and labels
1	All : all diagnostic labels	External model comparison: all labels from datasets (1) PTB-XL, (2) CPSC2018, (3) CSN, (4) EchoNext-Mini Internal model comparison: all labels from (1) PTB-XL, (2) CPSC2018, (3) CSN, (4) EchoNext-Mini, (5) MIMIC-IV-ECG-Ext-ICD
2	ARR : arrhythmias	PTB-XL Rhythm: atrial fibrillation (AFIB), sinus tachycardia (STACH), sinus arrhythmia (SARRH), sinus bradycardia (SBRAD), supraventricular arrhythmia (SVARR)
3	CD : conduction disturbances	PTB-XL Sub: left anterior/left posterior fascicular block (LAFB/LPFB), incomplete right bundle branch block (IRBBB), complete left bundle branch block (CLBBB), complete right bundle branch block (CRBBB), AV block (AVB), non-specific intraventricular conduction disturbance (block) (IVCD)
4	ECG-HYP : hypertrophy detected from ECG	PTB-XL Sub: left ventricular hypertrophy (LVH), right ventricular hypertrophy (RVH), left atrial overload/enlargement (LAO/LAE), right atrial overload/enlargement (RAO/RAE)
5	MI : myocardial infarction	PTB-XL Sub: anterior myocardial infarction (AMI), inferior myocardial infarction (IMI), lateral myocardial infarction (LMI)
6	ISC/STTC : ischemia and ST/T changes	PTB-XL Sub: ischemic in anterior leads (ISCA), ischemic in inferior leads (ISCI), non-specific ischemic (ISC_), ST-T changes (STTC), non-specific ST changes (NST_)

Appendix

Table 1: Diagnosis groups used for aggregated evaluation. Each group pools related binary labels from the indicated dataset sources. Only labels with adequate prevalence are included in the aggregate groups (at least 10 positive and negative class instances). * Only for comparison between internal models and ECGFounder as other external baselines pre-train on all of the MIMIC-IV-ECG data.

Code	Description
AFIB	Atrial fibrillation
STACH	Sinus tachycardia
SARRH	Sinus arrhythmia
SBRAD	Sinus bradycardia
SVARR	Supraventricular arrhythmia
LAF/LPF Blocks	Left anterior/left posterior fascicular block

Appendix

Table 2: Glossary of condition codes.

Category	PTB-XL (21,837)	CPSC2018 (6,877)	CSN (23,026)	EchoNext-Mini (100,000)	MIMIC-IV-ECG-Ext-ICD (468,005)
Arrhythmias	✓	✓	✓	✗	✓
Conduction Disturbances	✓	✓	✓	✗	✓
Ischemic Heart Disease
STE-ACS / STEMI	✓	$\bm{\sim}$	$\bm{\sim}$	✗	✓
NSTE-ACS / NSTEMI	✗	✗	✗	✗	✓
Chronic coronary syndrome (e.g., atherosclerosis, etc.)	$\bm{\sim}$	✗	✗	✗	✓

Appendix

Table 3: Disease coverage by downstream evaluation datasets. This expanded evaluation suite has more universal coverage of diseases that are detectable from ECGs by expanding it to include EchoNext-Mini and MIMIC-IV-ECG-Ext-ICD. Number of total instances in the dataset is in parentheses. ✓ indicates the dataset supports the condition, ✗ indicates it does not, and

\bm{\sim}

denotes partial or limited support of the condition as currently used. Labels for MIMIC-IV-ECG-Ext-ICD are at the visit-level.

			Single Lead			Lead Subsets				12-lead
Group	Condition	Prevalence % (N)	I (LCX)	II (RCA)	V2 (LAD)	I,II	All limb	I,II,V2	I,II,III,V2	12-lead
ARR	AFIB	7.2% (152)	97.7	98.5	97.3	98.5	98.7	98.5	98.5	98.7
	STACH	3.9% (82)	99.2	99.3	99.1	99.5	99.5	99.5	99.5	99.5
	SARRH	3.7% (77)	90.0	92.7	89.3	94.1	93.4	93.5	93.4	94.4
	SBRAD	3.1% (64)	96.1	96.1	95.9	96.3	96.1	96.1	96.0	96.7
	SVARR	0.7% (14)	84.9	96.1	71.9	93.7	93.7	91.8	92.4	95.4

Appendix

Table 1: Label-level linear probe AUROC of MS-ECG-FM by lead configuration. A glossary of condition codes is in Table 2 . 9 9 footnotetext: AUROC: area under the receiver operating characteristic curve. Prevalence is the proportion of positives in the test set the metrics were computed on. † Cardiac amyloidosis is excluded from the MIMIC-EXP aggregate score due to low prevalence.

			Single Lead			Lead Subsets				12-lead
Group	Condition	Prevalence % (N)	I (LCX)	II (RCA)	V2 (LAD)	I,II	All limb	I,II,V2	I,II,III,V2	12-lead
ARR	AFIB	7.2% (152)	87.4	92.3	89.3	93.5	94.5	94.0	94.4	95.6
	STACH	3.9% (82)	86.5	84.8	86.2	85.7	86.7	89.0	89.8	89.5
	SARRH	3.7% (77)	37.8	42.4	34.1	44.9	41.2	46.1	45.0	47.7
	SBRAD	3.1% (64)	58.0	62.5	63.7	63.4	63.5	62.0	62.7	63.3
	SVARR	0.7% (14)	5.9	30.8	2.6	28.5	25.6	34.7	32.2	48.7

Appendix

Table 2: Label-level linear probe AUPRC of MS-ECG-FM by lead configuration. A glossary of condition codes is in Table 2 . 9 9 footnotetext: AUPRC: area under the precision-recall curve. Prevalence is the proportion of positives in the test set the metrics were computed on. † Cardiac amyloidosis is excluded from the MIMIC-EXP aggregate score due to low prevalence.

			AUPRC
Group	Condition	Prevalence % (N)	Random	ECG-MR	ECG-CR	ECHO	Chest X-ray	Discharge	MS-ECG-FM
ARR	AFIB	7.2% (152)	49.6	94.8	∗ 95.5	91.6	91.3	93.0	95.6
	STACH	3.9% (82)	61.8	90.0	∗ 88.0	83.0	80.7	83.6	∗ 89.5
	SARRH	3.7% (77)	5.3	30.8	∗ 45.0	10.6	11.4	14.3	47.7
	SBRAD	3.1% (64)	42.0	65.3	∗ 61.5	58.4	∗ 59.7	57.2	∗ 63.3
	SVARR	0.7% (14)	2.4	∗ 35.4	∗ 32.5	28.8	∗ 36.7	32.6	48.7

Appendix

Table 3: Contribution of different clinical note types to MS-ECG-FM AUPRC for different condition labels. A glossary of condition codes is in Table 2 . 9 9 footnotetext: AUPRC: area under the precision-recall curve. Prevalence is the proportion of positives in the test set the metrics were computed on. The best model is bold ;

\ast

marks models that are not significantly different from the best per a paired patient-cluster bootstrap significance test (see Methods). † Cardiac amyloidosis is excluded from the MIMIC-EXP aggregate score due to low prevalence.

12-lead

Reduced-lead

All

ARR

ECG-HYP

ISC/STTC

SHD

MIMIC-EST

MIMIC-EXP

MS-ECG-FM

90.6

96.9

95.8

93.9

94.4

92.9

83.4

91.7

88.2

78.8

84.5

85.3

82.6

Z-score waveform-level

\downarrow

0.5

\uparrow

0.1

\downarrow

0.2

\downarrow

0.5

0.0

\downarrow

1.0

\downarrow

0.6

\downarrow

1.1

\downarrow

0.2

\downarrow

0.4

\downarrow

0.6

\downarrow

0.7

\downarrow

0.5

Z-score lead-level

\downarrow

0.7

\downarrow

0.2

\downarrow

0.2

\downarrow

2.4

\downarrow

0.8

\downarrow

1.1

\downarrow

1.3

\downarrow

2.8

\downarrow

0.3

\downarrow

0.4

\downarrow

0.6

\downarrow

0.8

\downarrow

0.2

Single layer tokenizer

\downarrow

0.3

\uparrow

0.5

\downarrow

0.3

\uparrow

1.0

\downarrow

1.0

\downarrow

0.3

\downarrow

1.0

\uparrow

0.5

\downarrow

0.6

\downarrow

1.2

\downarrow

0.5

\downarrow

0.3

\downarrow

0.3

No RLM

\downarrow

0.3

\uparrow

0.1

\downarrow

0.6

\uparrow

0.2

\uparrow

0.4

\uparrow

0.2

\downarrow

0.8

\downarrow

0.1

\downarrow

0.9

\downarrow

0.7

\downarrow

2.3

\downarrow

2.4

\downarrow

2.0

Appendix

Table 1: Ablation of other MS-ECG-FM components. Results for individual leads are for the All dx group. The MS-ECG-FM row is AUROC per diagnostic group. Other rows are

\Delta

AUROC versus MS-ECG-FM.

\uparrow

and

\downarrow

mark the sign of the difference. RLM: random lead masking augmentation Oh et al. (2022) .

	All	ARR	CD	ECG-HYP	MI	ISC/STTC	SHD	VC	MIMIC-EST	MIMIC-EXP
Stochastic pairing	90.6	96.9	95.8	93.9	94.4	92.9	83.4	91.7	88.2	78.8
Average embeddings	90.4	97.2	95.5	94.6	94.4	92.6	83.3	91.5	88.0	78.7
Report-specific batch	90.3	96.7	95.5	94.5	93.7	92.5	83.9	91.8	88.0	78.5
Report-specific projectors	90.3	97.3	95.3	93.8	94.0	92.2	83.7	91.4	88.2	78.2

Appendix

Table 2: Ablation of multi-source alignment objectives. Linear probe AUROC by diagnostic group. Default setting for MS-ECG-FM is underlined and the best per column is highlighted in bold .

Figure 1: Comparison of a model pre-trained with MS-ECG-FM’s recipe and architecture except that it is only aligned to the ECG machine reports (ECG-MR) to other contemporary ECG foundation models pre-trained with the same MIMIC-IV-ECG machine reports on 12-lead ECG diagnostic tasks using linear probe evaluations. Significance testing is performed using a paired patient-cluster bootstrap procedure (see Methods). 95% bootstrap confidence intervals are shown. AUROC: area under the receiver operating characteristic curve, ARR: arrhythmia, CD: conduction disturbances, ECG-HYP: hypertrophies detected from ECG, MI: myocardial infarction, ISC/STTC: ischemia and ST/T changes, SHD: structural heart disease, VC: diagnoses that rely on voltage criteria. Conditions by diagnosis group are in Table 1 .

Report	Text encoder	Checkpoint	All	ARR	CD	ECG-HYP	MI	ISC/STTC	SHD	VC	MIMIC-EST	MIMIC-EXP
ECG-MR	Frozen	ER	89.5	95.8	95.2	94.2	94.0	92.3	81.5	91.7	86.4	76.5
ECG-MR	Frozen	ZS	$\downarrow$ 0.1	$\uparrow$ 0.1	$\downarrow$ 0.1	$\uparrow$ 0.2	$\downarrow$ 0.6	$\uparrow$ 0.1	$\downarrow$ 0.4	$\uparrow$ 0.2	0.0	$\downarrow$ 0.1
ECG-MR	Online	ER	$\downarrow$ 3.9	$\downarrow$ 5.8	$\downarrow$ 1.4	$\downarrow$ 2.6	$\downarrow$ 4.3	$\downarrow$ 3.3	$\downarrow$ 2.5	$\downarrow$ 3.3	$\downarrow$ 1.9	$\downarrow$ 2.2
ECG-MR	Online	ZS	$\downarrow$ 24.7	$\downarrow$ 26.1	$\downarrow$ 24.5	$\downarrow$ 20.6	$\downarrow$ 26.1	$\downarrow$ 24.3	$\downarrow$ 21.8	$\downarrow$ 24.1	$\downarrow$ 20.2	$\downarrow$ 15.5
ECG-MR	Partially online	ER	$\downarrow$ 0.1	$\uparrow$ 0.7	$\downarrow$ 0.1	$\uparrow$ 0.4	$\downarrow$ 0.1	$\downarrow$ 0.3	$\downarrow$ 0.2	$\uparrow$ 0.2	$\downarrow$ 0.4	$\downarrow$ 0.1
ECG-MR	Partially online	ZS	$\downarrow$ 1.5	$\downarrow$ 1.4	$\downarrow$ 0.4	$\downarrow$ 1.6	$\downarrow$ 2.3	$\downarrow$ 1.0	$\downarrow$ 1.3	$\downarrow$ 1.0	$\downarrow$ 1.1	$\downarrow$ 1.6

Appendix

Table 3: Ablation of pre-training methods when aligning to ECG-MR reports only. Linear probe AUROC by diagnostic group. The frozen-text, ER-checkpoint row is linear-probe AUROC. Other rows are

\Delta

AUROC versus that setting.

\uparrow

and

\downarrow

mark the sign of the difference. 10 10 footnotetext: AUROC: area under the receiver operating characteristic curve. Underlined text encoder and checkpoint are the default setting. Frozen uses offline MedGemma-27B embeddings; online jointly updates a MedCPT text encoder (partially online freezes the first 6 layers) Jin et al. (2023) . ER is the RankMe / effective-rank pre-training checkpoint; ZS is the checkpoint selected by out-of-distribution zero-shot AUROC as used by MERL Liu et al. (2024) and MELP Wang et al. (2025) .

				# Total Notes (Patients)			Length (tokens)
Report	Description	Pre-processing / Cleaning	How to join to ECG	Raw	Clean	# Train ECGs w/Notes (Patients)	Median	Max
ECG Machine Report (ECG-MR)	Reports and summary measures generated by the ECG machine	Concatenate all statements together; remove one empty report	By study_id	800,035 (161,352)	800,034 (161,352)	720,101 (144,545)	17	125
ECG Cardiologist Report (ECG-CR)	The ECG is read by a cardiologist who then writes a report. They may reference the machine report or ignore it.	(1) Remove clinical indications, (2) Exclude reports with no informative text (e.g., “ECG interpreted by ordering physician, please see corresponding note”)	Joined to waveform_note_links by study_id then to mimic_iv_note_ecg by note_id	623,566 (105,365)	490,780 (102,886)	441,639 (92,377)	33	256
ECHO Cardiologist Report (ECHO-CR)	Ultrasound images of the heart. A cardiologist reviews ECHO images and writes a report.	None	ECG waveform is joined to the closest ECHO report in a 7-day window	81,821 (34,147)	81,821 (34,147)	129,934 (26,944)	442	1,128
Chest X-ray Radiologist Report (ChXR-RR)	X-ray of the chest. Can show cardiomegaly, pulmonary edema, vascular congestion, and other secondary signs of cardiac dysfunction. Radiologist reviews and writes a report.	None	ECG waveform is joined to the closest ChXR report in a 7-day window	715,465 (119,933)	715,465 (119,933)	415,589 (97,686)	108	761
Discharge Summary Physician Report (DS)	Report summarizing a patient’s hospitalization including the reason for admission, past medical history, their hospital course (including summaries of various scans and blood/microbiology tests), and any relevant discharge instructions	None	ECG joined to discharge summary for the corresponding hospital stay by hosp_hadm_id or ed_hadm_id	331,793 (111,647)	331,793 (111,647)	326,779 (78,754)	2,464	15,039

Appendix

Table 1: Summary of clinical reports in the MIMIC-IV-ECG and MIMIC-IV-Note datasets Gow et al. (2023) ; Johnson et al. (2023) . We join these reports to the 800,035 available ECG recordings (720,102 in the training set). Multiple ECGs from the same visit may be joined to the same ECHO, Chest X-ray, Discharge Summary, Chest CT or Chest CTA report. The

\bullet

symbol denotes that preliminary experiments looking at performance on the validation set (i.e., not test set) suggested that these reports did not provide useful supervision so we excluded them from further analysis (cf. Section 10.2 ).

Methods	PTBXL-Super $\dagger$			PTBXL-Sub $\dagger$			PTBXL-Form $\dagger$			PTBXL-Rhythm $\dagger$			CPSC2018 $\dagger$			CSN $\dagger$
Methods	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%	1%	10%	100%
ECG-MR	88.7	91.5	92.3	77.0	87.4	88.9	64.9	79.1	86.7	81.5	96.3	97.3	83.9	90.8	92.9	81.5	89.8	94.5
ECG-CR	88.9	92.0	93.1	80.3	87.8	90.3	67.5	80.4	88.3	91.2	96.4	98.3	86.9	92.3	94.0	79.9	91.8	97.0
ECHO-CR	87.3	90.0	91.7	79.2	85.6	88.6	59.0	71.6	79.1	78.9	86.2	93.1	78.3	88.6	91.7	74.0	86.7	93.2
ChXR-RR	87.1	89.7	91.2	77.8	84.2	87.9	62.7	74.0	82.2	77.0	87.8	94.1	75.7	88.4	92.3	74.2	86.3	92.0
DS	86.5	89.9	91.6	78.3	85.7	89.1	62.9	75.9	81.8	79.8	91.5	95.7	76.7	88.8	93.3	74.3	86.4	93.5

Appendix

Table 2: Linear probe AUROC on the validation set of existing out-of-distribution (OOD) benchmarks. The results are shown for different percentages of available training labels (1%, 10%, 100%).

\bullet

denotes exclusion from further analysis. Note: these experiments were preliminary with an earlier version of the model, so deviations with the other results in this manuscript are expected.

Config	Value
Optimizer	AdamW
Batch size	512
Learning rate	2e-4
Weight decay	0.2
Optimizer momentum	$\beta_{1},\beta_{2}{=}0.9,0.999$
Learning rate schedule	Cosine annealing w/linear warm-up

Appendix

Table 1: Pre-training settings for MS-ECG-FM and single-source models.

Config	Value
Optimizer	AdamW
Batch size EchoNext-Mini	1024
Batch size MIMIC-IV-ECG-Ext-ICD	1024
Batch size PTBXL / CPSC2018 / CSN 1–10%	16
Batch size PTBXL / CPSC2018 / CSN 100%	128
Learning rate	0.001

Appendix

Table 1: Linear probe settings for all ECG foundation models.

Explore similar work

Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals
May 26, 2026Phu X. Nguyen, Konstantinos Kontras, Wei Dai +5Electrocardiogram Foundation Models Multimodal Representations
1KU Leuven · 3MIT · 4DFKI +1
A Signal-Language Foundation Model for Broad-Spectrum Cardiovascular Assessment from Routine Electrocardiography
May 25, 2026Ziqing Yu, Yuhui Tao, Jiayu Huo +19Electrocardiogram Foundation Models Adult-To-Pediatric Electrocardiogram Transfer
Department of Cardiology, Zhongshan Hospital of Fudan University, Shanghai Institute of Cardiovascular Diseases, National Clinical Research Centre for Interventional Medicine, Shanghai, China · Digital Medical Research Center, School of Basic Medical Sciences, Fudan University, Shanghai, 200032, China. · Shanghai Key Laboratory of Medical Imaging Computing and Computer Assisted Intervention, Shanghai, 200032, China. +8
Pretraining Strategies and Scaling for ECG Foundation Models: A Systematic Study
May 12, 2026M A Al-Masud, Nils StrodthoffElectrocardiogram Foundation Models Electrocardiography
AI4Health Division Carl von Ossietzky Universität Oldenburg Oldenburg, Lower Saxony, Germany

May 26, 2026, cs.LGJ/K move · Enter open · S save

Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals

Phu X. Nguyen, Konstantinos Kontras, Wei Dai, Huy Phan+4

1KU Leuven · 3MIT · 4DFKI+1

Electrocardiogram Foundation Models Multimodal Representations Representation Learning

Open paper AskPDF arXiv

Electrocardiograms (ECGs) are widely used non-invasive measurements of cardiac activity and play a central role in clinical diagnosis. Recent multimodal approaches align ECG signals with clinical reports to incorporate diagnostic semantics, but clinical reports often fail to preserve the rich physiological structure of ECG waveforms, particularly across multiple levels of abstraction ranging from coarse diagnostic categories to fine-grained morphology. To address this limitation, we formulate ECG representation learning from an information-theoretic perspective and derive a tractable objective that jointly preserves signal structure and integrates clinical semantics. Based on this principle, we propose \textbf{MERIT} (Multimodal ECG Representation via Information Theory), a dual-branch pretraining framework combining masked ECG modeling with ECG--text contrastive alignment. Extensive experiments on PTB-XL and additional benchmarks demonstrate consistent improvements over prior methods, including gains exceeding

3%

F1 on PTB-XL All and

5%

F1 on SubClass classification. In zero-shot evaluation, MERIT further improves performance by up to

+2.66\%

AUC and

+2.11\%

F1 on PTB-XL SubClass, while also demonstrating robustness under multiple distribution-shift settings. Moreover, leveraging the learned ECG representations for ECG-conditioned clinical text generation with large language models improves text quality across several metrics, including ROUGE and METEOR. Together, these results demonstrate that MERIT learns more informative and clinically meaningful ECG representations, particularly for fine-grained clinical applications.

Abstract

Explore similar work

Information-theoretic Multimodal Representation Learning for Electrocardiogram Signals

A Signal-Language Foundation Model for Broad-Spectrum Cardiovascular Assessment from Routine Electrocardiography

Pretraining Strategies and Scaling for ECG Foundation Models: A Systematic Study